Home

Ideas and outcomes

This Wiki is the concise, living index of ideas raised in Unjuno/agent-interface Issues and what happened to them. Issues remain the intake and discussion space; this page records each idea's current state and linked outcomes. Closure is workflow state, not proof of success. Preserve every prior STOP, HOLD, FAIL, and scoped result; a successor never rewrites its predecessor.

The detailed repository ledger is docs/IDEAS_AND_OUTCOMES.md; the selected Issue inventory is docs/RESEARCH_ISSUE_INDEX.md. Issues, PRs, and retained reports remain authoritative for exact evidence. Reconciled 2026-10-02 from GitHub: keep the complete non-PR Issue collection distinct from the selected research inventory; their scopes and snapshots differ, so this Home page does not present stale totals as current. The selected inventory links to its separately maintained ledger. Recent ideas through #6723 and outcome/status updates reviewed during the 2026-10-02 14:13 UTC reconciliation through Issue #6723 and PR #6732, including additional closed-Issue lineage review are included below; source-review limitations are distinguished from executed results.

Durable idea clusters

Idea	Current understanding
Interface efficiency	Reduce redundant observations, model boundaries, serialization, waiting, recovery, and relearning without weakening correctness. Issue #57 was reopened after a completion-checklist audit; its five completion checklist items and full integration/comparison remain unresolved. Matched efficiency remains HOLD; no general end-to-end or human-tempo benefit is established.
Rich planner + bounded local refinement	Keep semantic intent with the rich model; use bounded deterministic local loops for frequent refinement, verification, invalidation, and recovery. Components have scoped evidence; end-to-end usefulness remains unproven.
Universal fallback	Specialized routes are optimizers, not prerequisites. Broad cross-application promotion remains unproven.
Compact control + observation gating	Complementary optimization directions, evaluated at equal correctness. No universal efficiency claim follows from design alone.
Local deterministic control + layered invalidation	Keep frequent input, verification, release, and fine correction local where reasoning is unnecessary; invalidate only the narrowest stale layer. Broad live-app validity remains unresolved.
Pending effects + bounded authority	Preserve unresolved effects as dependencies and bound physical authority. Finite escrow evidence is scoped; counterexamples and live distributed-system limits remain.
Evidence integrity	Preserve first outcomes and counterexamples; distinguish proxies from measured tokens, bytes, performance, and task results.

Latest reconciliation — 2026-10-02 17:03 UTC follow-up

Selected new outcome; other changes from the latest bounded Issue/PR search remain to be reconciled. No experiment was performed by this indexing task.

• #6613, OPEN — Separate EDF deadline successor also retained FAIL_HYPOTHESIS. Fresh40 seeds20000–20039 ×3 strata ×3 policies, distinct from the earlier service-debt A01 failure. One native local WSL2 CPU candidate and one independent raw-only auditor, both exit0,360/360 schedule rows replayed,5/5 corruptions rejected,retries0. Asymmetric on-time optional completions per trace: EDF1.80 versus FIFO2.65 and shortest-service4.95; burst/recovery EDF1.025 versus FIFO1.70 also missed the frozen noninferiority floor. Mandatory events and hard-eligibility/accounting gates passed: audit agreement is not hypothesis success. Wrong-working-directory launch wrapper failed before starting the auditor; corrected command invoked the auditor once, not a scientific retry. The auditor independently derives metrics/disposition but does not compare the candidate's redundant summary envelope. PR #6763 is OPEN/unmerged; local sparse-index checking limitation remains, hosted CI not inferred. No container/GPU/model/GUI, human fairness, real authorization/safety or product benefit; earlier service-debt FAIL and provenance STOP remain unchanged. First outcome.

Selected source-bound audit follow-up — 2026-10-02 17:05 UTC:

• #1839 — Retained no-positive-effect corpus: A01 auditor HOLD, separate A02 mutation-contract PASS. Merged #6761 preserves HOLD_AUDITOR_NOT_COMPLETED: one OrbStack candidate exited0 with six sessions/194 scorer samples/three attack physical joins/zero positive endpoints or authority grants; sole auditor exited1 on the frozen ghost-actuation mutation harness AssertionError, with no audit summary. No retry or scientific PASS. Distinct merged #6764 retains PASS_AUDIT_MUTATION_CONTRACT_SCOPED: fresh A02 candidate/auditor once each exit0, independently matching six sessions/194 samples, attack joins3/3, positive endpoints0 and authority0; six summary mutations rejected6/6 (drop/duplicate session, forged effect, erased physical join, changed sample count, authority grant). This qualifies that auditor against the retained #4193 no-positive-effect corpus only; positive-effect sensitivity and causal attribution are not established. No new game/GUI/model/input or efficacy run. A01 HOLD, historical whitespace-lineage HOLD_CONTRACT_GAP and source-identity HOLD remain unchanged; the documentary endpoint review separately retains HOLD_NO_DECISION_VALUE_ENDPOINT, not a negative feedback-value finding. Merge delivers evidence, not a live plan-bound useful-effect channel or resolution of the broader instrumentation contract.

Selected prefix-obligation outcome follow-up — 2026-10-02 17:10 UTC:

• #6749, OPEN — Preserve pending obligations after decisive FAIL; keep disposition counts separate from derived metrics. The earlier preparation now has two distinct finite-method results. OPEN/unmerged #6760: OrbStack candidate and independent auditor once each in separate offline containers, retries0, PASS_METHOD_SCOPED/errors0/five mutations rejected; authored32-world/4,064-unique-prefix fixture includes CLEAR→CONFLICT→COMPLETE, added before formal execution. Distinct merged #6762: user-authorized WSLc variant, candidate/auditor1/1/retries0, agreement8,640 ordered-prefix rows across24 worlds, five mutations rejected, PASS_METHOD_SCOPED; INVALID2,160/PASS60/PROVISIONAL5,250/STABLE_FAIL1,170. WSLc omits the later-conflict continuation, so different world contracts and unique-versus-ordered denominators must not be pooled or called same-fixture engine replication; its narrower PASS does not supersede #6760. WSLc requested512MiB with cgroup/swap warning: effective memory enforcement unverified. Both test authored prefix/accounting contracts only, not real late corrections, verifier truth, useful latency, GUI/runtime safety or action authority. Preserve #6689/#6706 FAIL_METHOD and #6704 FAIL_AUDIT/STOP_CLASSIFICATION_COUNT_CONTRACT unchanged. Coverage clarification.

Selected safe-stop presentation outcome — 2026-10-02 17:11 UTC:

• #6342 — Directive versus autonomy-supportive safe-stop cards: structural checks passed, overall HOLD. Merged #6758 now retains the earlier preregistered T0 first execution: separate pinned offline OrbStack candidate/auditor once each, exits0, no reruns; four matched pairs/eight synthetic cards and two planted controls. Structural auditor errors0 and unsafe-retry/unsupported-success controls both rejected; construction6/6 is separate. Overall HOLD_INDEPENDENT_REVIEW_REQUIRED because the frozen independent blinded semantic review did not occur: structural parity is not full method PASS. No equivalent comprehension, perceived autonomy/reactance, reading-time, user behavior, safety or product benefit is measured; no participants/GUI/audio/model. Main delivery does not close the idea or authorize T1, and the required review remains missing. Result and merge confirmation.

Selected safe-probe successor follow-up — 2026-10-02 17:13 UTC:

• #6147 — Choose the post-probe effect envelope, not merely an identified initial state. Merged #6756 retains distinct host-CPU analytical A04/A05, not repairs of A03 FAIL_AUDIT_CONTROL_COVERAGE. A04 independently audits p→q initial-state distinction that ends in stale-target envelopes: both deny action/YIELD; one-step p-only convergence, action-equivalent aliases and safe bisimulation are separate controls, with seven mutations rejected. A04 does not cover the exact p→q convergence requirement. Fresh A05 covers that specific case: unique resolved depth2 p→q tree converges C0/D0 to current Z; all13 safe words retain equal initial observation traces, so initial identity is not claimed; current terminal effect envelope is used, informative unsafe u rejected, six mutations rejected. Finite synthetic T0 boundary only: no real GUI/model/container, safety transfer or T1 authorization, broader idea remains unresolved. Preserve A02 STOP_CANDIDATE_RUNTIME_ERROR/NOT_EVALUATED and its raw-byte discrepancy, A03 audit-control FAIL, and original A01 formal0 source preparation separately. The later archive/ref cleanup report changes A01 preservation status only and is not experiment execution; this index performed no branch operation. Successor scope.

Selected evidence-integration follow-up — 2026-10-02 17:14 UTC:

• Delivery status changed, scientific outcomes did not. Merged #6754 preserves #6081 S1/S2 runner STOP packets (scientific scripts unrun/result JSON absent), replacing the older Draft/open status. Merged #6750 preserves #59 model-store mount prelaunch STOP, candidate/auditor0/0, not mount failure. Merged #6757 supplies the replacement delivery for earlier #6752 WSL CUDA smoke: GPU-route PASS/response-contract FAIL and missing full-log qualification remain; no new smoke, benchmark or formal live control. Merged #6755 integrates #5424 T4 severity-ranking method result without changing fallback severe-event increase2→3/completed-offer decrease6→5 or adding calibrated/live benefit. Earlier open-PR descriptions are timestamped history, not current integration status.

• #6121 finite T0 candidate runtime STOP is now preserved, not evaluated. Merged #6149 retains seven originals: sole host-CPython candidate exited1 at initial obligation creation (emit received multiple values for kind), no candidate JSONL, auditor0/corruption controls0/retries0, STOP_CANDIDATE_RUNTIME_ERROR/NOT_EVALUATED. Docker inventory timeout and WSL startup E_FAIL were setup observations, no container/WSL/GPU/model/GUI run. Frozen PLAN Git-blob identifier has39 characters, so archive does not attest that source identity; prior failed index check stays failed, and the displayed traceback was truncated (no full-stderr byte hash claimed). Separate #6160 successor does not upgrade this consumed allocation; no scientific capacity finding or source repair follows from merge.

Selected modal-context lineage follow-up — 2026-10-02 17:16 UTC:

• #6102 — Match nested modal call/return to the exact parent and generation. Canonical merged #6106 retains PASS_METHOD_SCOPED for14 authored depth≤3 finite-ID traces: stack and depth-matched FSM agree with independent oracle14/14, invalid controls10/10 fail closed, flat modal flag false-accepts6/10. OrbStack offline execution, mutation tests6/6, byte-identical frozen-source repeat; prefreeze auditor-schema failure remains excluded and preserved. No stack superiority over a depth-matched FSM is shown; bounded-depth behavior can be encoded finitely. Finite GUI IDs/generations do not prove an arbitrary live-identity VPA theorem; non-LIFO/interleaved contexts yield UNKNOWN. Closed UNMERGED #6107 retains a distinct encoding (5,461 words;51 complete identity matches;306 flat depth-only false accepts;FSM/stack disagreement0;overflow UNKNOWN432) but was adjudicated duplicate, excluded from additional progress/result counts, not a new successor result. Another host candidate emitted58 unaudited event rows, then its allocated auditor launch used the wrong cwd: STOP_AUDITOR_LAUNCH_PATH/NOT_ADJUDICATED, auditor code0/output absent/retries0; do not pool with either corpus. Latest archive/ref cleanup preserves duplicate source/history only, no rerun or change to the canonical result. No real GUI-event eligibility, task effect, action authority, arbitrary nesting/identity, safety or performance claim. Duplicate qualification.

Selected opportunity-denominator lineage review — 2026-10-02 17:18 UTC:

• #5694 — Keep externally scheduled opportunities in the denominator, including those never attempted during busy periods. Canonical merged #5738 retains eight authored synthetic scenarios and auditor PASS_METHOD_SCOPED: dense completed-cycle p95=30ms/useful3/3 versus sparse5ms/useful2/3, missing O2 during busy time; safe-stop, expiry, censoring/unsynchronized UNKNOWN, overlap matching and no-opportunity NOT_APPLICABLE stay distinct. Candidate/auditor once each, no retries; raw exists/auditor exit0, but candidate exit0 was not reliably captured, so the frozen prerequisite before auditing and formal-protocol compliance remain unverified. Preserve prefreeze absolute-timestamp-as-latency construction failure separately; no live benchmark bias, latency correction or MAP01 regrading follows. Closed UNMERGED #5940 and the separate41-opportunity parallel probe are explicitly redundant/noncanonical, not successor or independent replication: preserve their outputs/provenance limits without pooling or counting added progress. Latest archive/ref cleanup changes source retention only, no rerun. Unverified refinements remain: separate event-count from duration-weighted coverage, and partition misses by acquisition/delivery/decision/effect boundary rather than assume planner stalls. Natural gameplay opportunities can depend on prior actions; same seed does not establish common exogenous denominators. Overlap needs frozen effect-to-cue matching; scorer-only cue presence is not model-visible evidence. Synthetic accounting only, no live GUI/model/game safety or human-tempo benefit.

Selected route-selector lineage review — 2026-10-02 17:19 UTC:

• #5911, OPEN — Derive route-change labels from explicit min/max selectors and ties. Predecessor delivery merged #5913 preserves a failed counterexample criterion: original candidate emitted no numeric half-cost elasticity, independent paths80ms versus260ms did not change route; conservative/inconsistent labeling is not a demonstrated runtime defect. #5913 is a corrected evidence copy, not another allocation. Separate merged #5918 retains finite T0 PASS_METHOD_SCOPED: five frozen selector/tie cases agree with independent enumeration, three corruptions rejected, candidate/auditor once each; two startup/binding failures and runtime-adapter limitations preserved. Execution used a Codex V8 function adapter, no GPU/container/GUI/real trace. A separate overlapping CPython3.14.5 supplemental five-case result adds max-cost control, source/raw/process retained on its branch with no duplicate PR; do not conflate it with canonical #5918 or infer independent live replication. Latest cleanup distinguishes allocation01 syntax-only/source preparation (candidate/auditor unrun) from allocation02's already-published result; archive/ref removal adds no scientific outcome. This validates authored selector/argmin/argmax/tie consistency only, not original #5851 live semantics, causal optimization validity, runtime performance or #59 completion.

Selected locale-invariance lineage review — 2026-10-02 17:21 UTC:

• #5919, OPEN — Normalize declared locale presentation, never target/value/effect/authority differences. Merged #5923 retains eight synthetic cases labeled H_PASS_SCOPED_METHOD_PASS: two benign pairs accepted, five seeded faults rejected at declared stages, ambiguous translation UNKNOWN. English-only/visual-proxy controls missed five faults, but the proxy is stipulated metadata, not measured images; no natural translation or rendered-GUI result. Candidate/auditor once each in separate V8 isolates. Preserve the subsequent auditor gap: its corruption gate checked candidate-provided probe metadata, not actual altered-result rejection, and duplicate/extra rows were accepted. Distinct merged audit-only #5928 retains PASS_RAW_REVIEW: independently recomputes original8 cases and rejects three actual altered copies plus duplicate/missing/extra rows, candidate0/audit1/retries0, original source/raw/result/audit unchanged. A separate Issue-reported pinned-container v2 audit accepted original raw/rejected6 controls without candidate rerun; do not identify that local packet with #5928's no-Docker supplement. Parallel nine-row prefreeze sketch and failed/corrected corruption prototype remain construction-only, not added H-pass. Latest original-tip archive/ref cleanup changes preservation only. Finite authored semantics and audit completeness do not establish multilingual agent robustness, external translation validity, real app effects or product benefit; broader live/natural-language validation remains open.

Selected verifier-exposure lineage review — 2026-10-02 17:23 UTC:

• #5941, OPEN — Separate blind first-pass evidence from peer-exposed critique; do not double-credit copied conclusions. Merged #5954 retains allocation01 HOLD_AUDITOR_NEGATIVE_CONTROL_NOOP:66 synthetic rows reconstructed, but mutation rejection4/5 because forge_independence changed no decision; original candidate diagnostics are not validated PASS. Distinct merged #5968 retains allocation02 PASS_METHOD_SCOPED:66-row raw-only reconstruction/errors0, six effective mutations rejected, candidate/auditor once each/no reruns, host-CPU stdlib only. Copied peer-exposed vote witness refuses UNKNOWN_INDEPENDENCE, shared raw observations do not themselves imply peer-verdict exposure, unknown history/invalid commitments receive no credit. Residual false PASS remains1 versus count-only10: two valid unexposed PASS votes against independently assigned FAIL truth. Exposure accounting is not protection from independent/correlated errors; these are authored case counts, not operational rates. Prefreeze circular vote-derived truth correction is preserved separately; metadata/call-bound provenance is assumed, not attested. Another128-case preparation retains STOP_DUPLICATE_ACTIVE_T0_PR/NOT_EVALUATED, formal0/0; later archive/ref cleanup adds no experiment. Unverified T1 distinguishes prior verdict from additional raw evidence with matched exposure and commit-then-critique; no LLM conformity, real quorum safety, GUI effect, runtime or product benefit was measured. Old HOLD remains unchanged and empirical hypothesis stays open.

Selected human-decision queue lineage review — 2026-10-02 17:25 UTC:

• #6150, OPEN — Batch presentation must preserve per-card consent; count all offered versions, not only answered cards. Earlier4/4-tested authority-queue preparation was parked stale-base before formal0/0 and remains unexecuted; latest exact-tip archive/ref cleanup is preservation only. Distinct all-offered T0 result report labels PASS_METHOD_SCOPED: host-CPython candidate/auditor once each,13 independent fixture checks; authored batch timely-correct4/10 with answered-only mean1.0 versus immediate8/10/mean2.0 reverses the speed ranking. Denial, pre-notice cancellation, revised card versions, expiry/right-censoring and wrong-target hard violation remain separate; silence is never consent. Record this as Issue-reported outcome, not independently read-back raw or verified main integration: bounded #6150 PR search returned none, and latest recovery found no matching result PR/main package on the older preparation branch; differing hashes forbid attributing that result to the parked branch or reconstructing missing evidence. Synthetic service-time/accounting is not measured human attention, preference correctness, safe consent or real queue performance. Unverified refinements separate effect expiry from adjustable presentation timeout, offered/noticed/answered denominators, factual comprehension from genuine accept/deny/defer preference, and concurrent-task interference from queue ordering. No participants/GUI/model or implementation by this indexing task; batching is not blanket approval and real benefit remains untested.

Selected screening-claim lineage review — 2026-10-02 17:26 UTC:

• #5739, OPEN — Confirm only supported survivor claims; speed cannot outweigh forbidden effects. Merged #5747 retains PASS_CLAIM_BOUNDARY_SCOPED on32 authored deterministic assignments: shared-baseline multiple survivors HOLD_MULTIPLE_COMPARISON_UNADJUSTED, screened-out sealed best yields SURVIVOR_VS_BASELINE_ONLY_NO_GLOBAL_OPTIMALITY, fastest forbidden-effect arm HARD_SAFETY_STOP. One host-Windows candidate exit0; first auditor-v1 exited1 with NameError during report serialization, STOP_AUDITOR_REPORT_SERIALIZATION/no result file. Separately frozen read-only audit-v2 reconstructed the unchanged32 raw assignments and rejected7/7 mutations, no candidate rerun; old STOP remains. Distinct duplicate allocation preserves STOP_AUDIT_BASE_PIN_MISMATCH/NOT_EVALUATED: candidate once exit0, sole auditor wrong hard-coded base exit1, candidate labels unaudited;10/10 preparation tests missed the file-level gate. Latest exact-tip archive/ref removal delivers retention only, not a method result or evidence against canonical #5747. No calibrated statistical uncertainty, real candidate/global optimum, task effects, container/GPU/model/GUI, runtime savings or product benefit; finite claim-label integrity only, no T1 authority. PR-required checks passed but formal check skipped by design, not a formal re-execution; merge closes neither broader hypothesis nor live #59 acceptance.

Selected image-only looming lineage review — 2026-10-02 17:28 UTC:

• #6030, OPEN — Identical pixels/timestamps cannot distinguish physical contact from harmless sprite growth. Preserve S1 preformal truth-pair-audit qualification STOP and S2 frozen-base/commit-parent mismatch STOP, candidate/auditor0/0 each; latest original-tip archival/ref cleanup changes retention only. Merged #6051 retains S3 STOP_AUDIT_INPUT_FORMAT/NOT_EVALUATED: sole host-Python candidate exit0 emitted12 JSONL records, sole auditor exit1 before row parsing because plain JSONL was opened as gzip, no audit result/no retry. Distinct merged audit-only #6057 retains S4 PASS_AUDIT_ONLY_SCOPED/PASS_VISUAL_EQUIVALENCE_BOUNDARY_SCOPED: one corrected frozen auditor on exact unchanged S3 raw, candidate0, exact timestamped raster pairs6/6/distinct hidden truth6/6/paired decision invariance6/6, four effective corruptions rejected. Pixel-change/radius-growth yield6/6 pairs and radius-tau4/6; each yield also occurs for the visually identical no-contact counterpart. S4 does not retrospectively convert S3's first audit into success. Six synthetic64×64 monochrome counterfactuals establish only an input-identifiability boundary, not natural prevalence or failure of cues under independently certified rigid/contact assumptions. No live camera/GUI/game/model/GPU/container, physical task effect, safety or product result; #5905 predecessor STOP remains unchanged and live transfer remains unverified.

Selected #59 key-evidence lineage review — 2026-10-02 17:30 UTC:

• Per-key occupancy, nullable cleanup, then virtual-server keymap witnesses are distinct rungs. Merged #6094 retains exploratory host-only two-key synthetic reconstruction (SPACE32–50ns/W70–90ns, candidate/auditor once, five corruptions rejected); full source hashes were recorded after first outcome, so this is not prospectively source-frozen physical occupancy evidence. Separate merged audit-only #6122 retains PASS_AUDIT_OWNER_BOUNDARY_SCOPED on immutable #6105 T0 inputs: accept key=null only for exact owner-lease-cleanup close with bound keycode25/owner/intent/reason/order/time/source; reject null explicit-up and forged/non-boolean checks. Candidate/auditor1/1 denotes JSON audit processes, not owner execution; #6105 FAIL_AUDIT remains unchanged and explicit-up T1 is separate.

• T3 real-Xlib construction and corrected publication. Closed UNMERGED #6140 retains PASS_XVFB_KEYMAP_WITNESS_CONSTRUCTION_SCOPED8/8: private WSL namespace/Xvfb, two XTEST W down/up occurrences, distinct interval IDs,32-byte keymap false→true→false, matched releases and empty close; candidate/auditor once/retries0. Python-Xlib0.33 list[int] return differed from fake-Xlib bytes assumption; normalization changed only isolated T3 copy, predecessors unchanged. Shared-/tmp startup failure preserved. #6140's17 Git paths had literal backslashes, a publication defect, not scientific refutation; corrected merged #6158 delivers the same17 blobs under slash paths, no experiment rerun. Latest provenance review reports original head archived by tag and17/17 identities match main; old branch absence is not lost evidence. No physical HID duration, application consumption/effect, safety/latency, model usefulness or live MAP01/#59 completion follows from these server/audit fixtures.

Selected German-XKB lineage review — 2026-10-02 17:35 UTC:

• Correcting the release oracle did not make #3792 a completed test. Closed #3792 formal-02 stopped at receiver_focused: Lookup lacked process DISPLAY, then run_row returned None and the wrapper failed before raw.json. No layout/preflight/formula candidate was reached. Merged #4560 preserves six exact frozen source files, not missing raw evidence; no rerun. Its current-main CONTAINER_RUN.md still imports the later live v2 directory instead of the archived bundle (also flagged by review), so it is not verified relocation-safe reproduction guidance. Source preservation does not resolve that command-path caveat or promote the STOP.

• Distinct #3794 receiver-delivery result. Closed #3794, merged #3800: one formal pinned OrbStack run, three German two-level XKB rows plus US control, exact =B2*A2 delivery4/4, unsupported Euro refusal with zero emissions/events4/4, exact paired trace and empty release4/4; independent second-container audit zero errors and five corruptions rejected. Character payload is counted from KeyPress only; non-empty KeyRelease lookup is not duplicate text. Eight preceding construction attempts and separate #3799 construction publication are not extra formal allocations. PASS_GERMAN_FORMULA_DELIVERY / PASS_AUDIT_CONFIRMED_DELIVERY remains distinct from predecessor STOP and proves private X11 receiver delivery only, not Calc/task effect, hardware, IME/level3, broad layouts or latency.

Selected closed-Issue backfill — 2026-10-02 17:40 UTC:

• Referenced image bytes and cursor frontier should commit together under the tested storage fault. Closed #4093, merged #4128: retrospective publication, not GitHub preregistration or new run;18 private SQLite cases, raw audit18/18 and10 rejecting mutations. ATOMIC_PAGE retains old cursor in6 fault cases and all images9/9 after declared capacity raise/recovery. Metadata/cursor-first comparator leaves6/9 incomplete,9 images missing: FAIL_SAVED_CURSOR_IMAGE_COMPLETENESS remains beside PASS_IMAGE_CAPACITY_TRANSACTION_SCOPED. Both policies returned native capacity errors; not false-success-return or permanent-loss evidence. Single trusted writer, available source images and ordinary saved-cursor recovery only; no device ENOSPC/power-loss, multiwriter, eviction, model consumption or performance claim. Original GitHub-write/tool STOPs and separator-newline transfer failure are retained; later28-part exact patch delivery resolved publication HOLD with zero scientific reruns.

• Reconnect starts a new observation epoch, not justified continuity of old Shift state. Closed #4135, merged #4142: PASS_RECONNECT_KEY_STATE_BOUNDARY_SCOPED on24 fresh private Tk/Xvfb sessions, raw audit233 checks/no errors,12/12 rejecting mutations. Carry-old-state gives exact b8/12, wrong B2/12, unresolved2/12; epoch-bound current keymap bootstrap gives exact b8/12, wrong0, unresolved4/12 (required missing/foreign-bootstrap refusals, not task successes). It repairs the directed disconnected-press/release cases only. First allocation retains STOP_EXTERNAL_TOOL_TIMEOUT / HOLD_FORMAL_INCOMPLETE: seven complete prefix rows, one incomplete start, missing batch receipt/outer exit. Separately frozen v2 changes only outer batching to six4-case batches; no v1 pooling. Cooperative serialized one-Shift fixture does not establish physical HID, silent-loss detection, causal cross-stream provenance, bootstrap-to-input atomicity, model utility or deployment.

Selected closed-Issue backfill — 2026-10-02 17:43 UTC:

• A fixed byte limit does not fix eager-index allocation cost. Closed #4122, merged #4127: retrospective delivery of locally frozen c5d2, not GitHub preregistration/new execution. Exactly1MiB input,128/1024/8192 records,three fresh processes each; median retained tracemalloc bytes29,721/259,521/2,130,297 (high/low71.67649,gate>=8), prefix entries129/1025/8193 despite max_records32. PASS_INDEX_DENSITY_MEMORY_SCOPED,9 worker exits0,raw audit no errors,10/10 corruptions rejected. This measures retained tracked Python allocation, not RSS/peak/native memory or an unbounded-memory claim; no throughput/live append/default recommendation. Separate demand-index lane stopped before formal at STOP_PARALLEL_DUPLICATE_BEFORE_FORMAL when #4068 appeared; construction is not pooled into c5d2.

• Same callback is insufficient if it pumps events between history check and Undo. Closed #4090, merged #4099, superseding closed UNMERGED publication-only #4096 without new science. Retrospectively delivered40 private Tk/Xvfb cases: SPLIT_KEY correct/wrong/refused4/4/2; INLINE_NO_YIELD8/0/2; INLINE_YIELD4/4/2; INLINE_RECHECK4/0/6 (each10). PASS_COMPARE_UNDO_REENTRANCY_BOUNDARY_SCOPED preserves8 wrong comparator outcomes as failures and12 refusals as unresolved, not compensation successes. Native competing ABA edit restores equal text but changes Undo history; Tk update admits reentry, a subsequent recheck refuses it, while no-yield callback performs target Undo then processes pending input rather than dropping it. Raw-only audits20+20,12 corruption designs reject in each batch,no formal rerun. Cooperative single-thread fixture only: not OS/GIL lock, universal GUI transaction, arbitrary-app safety, crash durability or latency benefit; inline-versus-split also changes API, so reentry contrast rests on the inline routes.

Selected closed-Issue backfill — 2026-10-02 17:41 UTC:

• Upper schema/grounding validity is not subprocess completion evidence. Closed #4081, merged #4112: archival delivery complete, original HOLD_FORMAL_INCOMPLETE unchanged (17 complete cases,one partial without entry exit,30 unstarted;two observed outer exits of eight planned batches). Exact upper call/real validators with substituted local subprocess transport expose CURRENT accepting child exit23,completed+failed events and string usage; schema-invalid/colliding-point controls refuse. JOIN_ONLY even refuses normal child0 because unchanged broker exits1; completed JOIN_FIXED normal control returns. Unbalanced prefix counterexamples are not rates/full48 acceptance. Full auditor still lacks batch02 receipt; prefix audit12/12 mutations and10 tests certify retained prefix only. No Docker/model/GUI run or scientific rerun. Strict consumer needs a compatible reviewed producer exit/envelope contract, not schema checks alone.

• Qualify image content without rewriting action outcomes. Closed #4080, merged #4100: retrospective locally frozen39 synthetic subprocess cases over three public presentation functions, PASS_REVIEW_IMAGE_QUALIFICATION_SCOPED. Same returned bytes yield12 decode+dimension matches,18 content HOLDs,9 preserved upstream refusals/no-image outcomes; all39 original receipts/outcome summaries unchanged, including synthetic partial/recovery/release-failure fields. Raw audit no errors,12/12 coherent mutations rejected,10 tests; excluded construction supervisor STOP retained separately. Digest/signature integrity is weaker than decode/geometry; historical image_status contract is not retroactively wrong. Same-dimension wrong-pixel control still passes, so no capture origin/currentness/target/semantic/model-viewing proof. Presentation HOLD does not erase execution results or authorize replay; no real capture/host integration, concurrency, task benefit or runtime promotion.

Selected closed-Issue backfill — 2026-10-02 17:43 UTC:

• Check after the last relevant same-event mutation, not merely at key-handler entry. Closed #4091, merged #4115: retrospectively delivered36 private Tk applications, PASS_DISPATCH_PLACEMENT_BOUNDARY_SCOPED. Correct/refused/incorrect: HOST_ONLY4/0/8,EVENT_EARLY4/4/4,PRE_CLASS4/8/0 (12 each). Widget callback moves caret or selects ab after early check, producing cab or c; post-callback/pre-class recheck refuses insertion and retains ab. Native c is still emitted/received/released in every case: eight PRE_CLASS refusals are unresolved partial states, not OS-input refusal or recovered tasks. Raw audit29,913 checks/no errors,12/12 mutations,15 tests. Known cooperative binding order with no later mutation/reentry only, not universal GUI atomicity. Formal/source244-file capsule plus six exact outer launch files is hosted; full402-member historical bundle is hash-bound, not claimed fully hosted. Publication truncation/digest/receipt fixes changed no science.

• Transfer a temporal monitor to real event transport without claiming application readiness. Closed #1786, merged #4118: unchanged A3 A_THEN_B_WITHIN(80ms) monitor consumes distinct X11 A/B/heartbeat PropertyNotify atoms and server timestamps. One frozen12-session formal: positive SATISFIED4/4,expiry EXPIRED4/4,repeated-A no-restart EXPIRED4/4; exact events/nondecreasing timestamps/oracle agreement12/12,worker exits0,task input0. PASS_TEMPORAL_CONTRACT_X11_PROPERTYNOTIFY_SCOPED; frozen audit five corruptions, separate postformal raw/source audit12 corruptions. Original #1801 source-only PR closed UNMERGED with scientific sessions0, not a failed transfer. Subsequent environment propagation and two process-isolation construction STOPs remain separate from eligible construction04/formal result; post-result summary-print and Base64 trailing-LF fixes were not reruns. No XTEST,semantic completion,loss/reconnect/wrap completeness,cross-clock comparison,model/task utility or production ABI follows.

Selected event-history lineage review — 2026-10-02 17:45 UTC:

• A delayed current property read is not an earlier notification payload. Open #3942 retains HOLD_INFRASTRUCTURE: one invocation exit1,23 complete private XTerm sessions/45 audited WM_NAME events plus one partial A event (46 retained,not48),missing final child acknowledgement/terminal receipt; frozen audit FAIL with four errors. Complete prefix exposes eight delayed AB/ABA naive-history misattributions and zero EARLY ones, but cannot satisfy full24/48 acceptance. XTerm exit0 does not supply the missing protocol receipt. Direct-write partial JSON is a plausible mechanism, not a confirmed cause of the unrecovered transient failure.

• Distinct atomic-command/receipt-complete successor. Closed #3956, merged #3970 retains predecessor HOLD and separate PASS_ATOMIC_TRANSPORT_PROPERTY_BOUNDARY_SCOPED: fresh24 sessions,48 consumer+48 observer events,48 current snapshots with historical UNKNOWN,complete command/ack/quit/terminal journals; independent audit no errors,eight corruptions rejected. Temp-file/fsync/atomic replacement and retained diagnostics change transport/observability,not title traces or policy gates. Separate-reader construction demonstrates partial direct-write possibility,not old failure attribution. Naive historical attribution still fails eight deliberately delayed AB/ABA cases; separated policy avoids inventing history but does not recover it. No XTEST/model/input work,natural race rate,crash durability,authentication,arbitrary-app completion or runtime promotion.

Selected closed-Issue backfill — 2026-10-02 17:47 UTC:

• Historical snapshot lifetime and logical history capacity are different resource bounds. Closed #4086, merged #4102: retrospective six-case SQLite study,768 accepted write transactions with final history exactly two rows. HELD_TX WAL apparent extent65,952/263,712/1,054,752 bytes after8/32/128 writes; COPY_RELEASE and BEGIN_ONLY8,272 throughout. Held-reader TRUNCATE busy2/2,after release0bytes; raw audit2,479 checks/no errors,12/12 mutations,12 tests. PASS_WAL_SNAPSHOT_RETENTION_BOUNDARY_SCOPED: copying the same coherent historical view and ending transaction permits reclamation,not currentness. BEGIN without initial SELECT is exposure control,not preserved history. PASSIVE checkpoint status0 with zero checkpointed frames is not full reclamation; journal_size_limit is not active-WAL hard quota. Apparent file size is not physical disk use/RAM/performance. Single reader/writer,finite128-write horizon only,no durability/multiwriter/runtime promotion.

• Receipt freshness can expire during validation, and even freshly validated pixels can already be outdated. Closed #4010, merged #4025: source-first21 native-XGetImage cases, PASS_VALIDATION_COMPLETION_AGE_BOUNDARY_SCOPED. Six receipt-fresh packets with explicitly injected>=60ms validation delay exceed20ms and completion-sample gate refuses; six prompt positives qualify; three already-stale and six digest/identity controls retain refusals. Three prompt-changed packets qualify at0.411–0.428ms yet differ from current display. Raw audit12/12 mutations,three batch/actor/server exits0,no retries. FRAME_FRESH remains receipt-time contract,not retroactively broken. Delay injection is not ordinary hash-cost/GIL/speed evidence; age at validation sample grants no later-use freshness,semantic currentness or input authority. Earlier13/32 STOP_CONTAINER_TOOL_TIMEOUT / HOLD_EVIDENCE_INCOMPLETE stays separate; full old archive conversation-hosted,not claimed uploaded. No model/XTEST/task/production benefit follows.

Selected closed-Issue backfill — 2026-10-02 17:49 UTC:

• Launcher exit, stopped execution and reaped descendants are different completion states. Closed #3982, merged #4021: PASS_OWNED_GROUP_REAP_BOUNDARY_SCOPED on24 Linux process-tree cases. Each policy has two normal ABSENT controls; six orphan cases are LIVE under LEADER_ONLY,ZOMBIE under GROUP_SIGNAL,ABSENT under GROUP_REAP. Same subreaper setting in all arms; raw wait statuses and independent PID/start-ticks/group snapshots reconcile,24 outside-group sentinels survive candidate observation,common rescue removes all owned residue. Raw audit12/12 corruptions; four actual batch/outer exits0. Pre-Python log-redirection STOP consumed no formal cases and remains separate. Old image-cost supplement still STOP_EXTERNAL_EXECUTION_ENVELOPE_PARTIAL_EVIDENCE / STOP_EVIDENCE_INTEGRITY at155/168 pairs,missing terminal/launcher exit; no filling/pooling. Explicit non-daemonizing tree only,no escaped descendants,supervisor-death/PID-reuse stress,input-release/rollback or production adoption.

• Abort request received is not abort applied before release. Closed #3930, merged #3980: one frozen18-case Tk/Xvfb/XTEST allocation, PASS_LIVE_ABORT_ACK_BOUNDARY_SCOPED. Current app-applied ACK before release suppresses ordinary Button callback3/3; request-only,ACK timeout and foreign-session ACK each still commit3/3 and remain QUERY_EFFECT. Unsupported abort refuses before press3/3. Unsafe request-received reporting retains FAILS_FALSE_ABORT_CLAIM3/3; candidate false-abort claims0,15 input-bearing cases released,18 terminal/cleanup neutral and app/server exits0; raw audit14/14 corruptions. Neutral release is not cancelled application effect. Explicit app disable is fixture-specific,not generic Escape/rollback; barriers are not natural race rates. Construction auth/byte-recording STOPs and publication422 incidents are preserved without scientific rerun; no #706 XI2 retry,#2197/model/task-efficiency or production completion.

Selected closed-Issue backfill — 2026-10-02 17:51 UTC:

• Docker schema preflight must connect to the actual allocation path; compatibility is not task success. Closed #2849, merged #3781: initial engine STOP,additive model seams and no-task mount/import/response replays are distinct from formal GUI allocations. v26 consumed one host call but response-accounting rejection prevented submission; v29 exact history conflicts with frozen zero-submission trace and remains FAIL_TASK1_SCOPED despite passing posthoc classification/offline reconciliation. Fresh v30 passes one task only,not six. Later #3864 stops after three schema preflights; distinct #3873 seed284936 completes three arms6/6 exact tasks with releases,17 unique reconciled host calls (three preflights+14 task calls),raw audit PASS and frozen RETAIN gate. Persistent repair excludes old-target pointer admissions; token/planner-generation break-even at task2. This closes the finite connection/allocation question,not current-main reliability,production,human tempo or population efficiency; every earlier STOP/FAIL remains provenance.

• Caller wait expiry is not source-time event expiry. Closed #4067, merged #4071: retrospective locally frozen18-case pipe study,unchanged event monitor,80ms source allowance vs160ms caller wait. PASS_CALLER_BUDGET_EVENT_TIME_BOUNDARY_SCOPED: SATISFIED3/EXPIRED6/TIMEOUT_UNRESOLVED6/UNKNOWN_INCOMPLETE_PREFIX3; unsafe synthetic-timer comparator six false expirations plus three unsupported silence classifications. Raw audit12/12 corruptions,36 child exits0,no formal rerun. Contiguous trusted source prefix is required before watermark-based expiration; late packets do not rewrite returned outcome or grant replay authority. Publication unexpectedly triggered three legacy GTK workflows (24 incidental rows); incident retained separately,not pooled or independently certified as GTK effect success. Merged engineering change restricts those legacy allocations to explicit manual dispatch; it is not zero-GUI publication nor additional temporal evidence. No real-producer completeness,model utility,hard-real-time or deployment claim.

Selected compensation-lineage review — 2026-10-02 17:52 UTC:

• Primary GUI restoration does not erase prior collateral effects. Closed #4006, merged #4020: PASS_LIVE_RESTORATION_COLLATERAL_SCOPED,24 fresh Checkbutton/Text sessions/96 reporting views using unchanged #2076 reducer. Six notification-enabled restores return initial primary value but retain two configured callback journal entries each; primary-only comparator falsely calls all six complete,full evidence yields PARTIAL_PRIMARY_RESTORED. No-notification restores and no-input controls satisfy declared contract; change-only fails;72 withheld/foreign/generation views UNKNOWN are shadow views,not extra GUI cases. Raw audit10/10 mutations,four malformed-evidence refusals,all exits/release queries reconcile. Declared private notification side effect is fixture configuration,not universal stock-widget behavior; no external-business rollback,concurrency/crash durability or model benefit.

• Compensation must bind edit scope and history, not just equal current text. Closed #4043, merged #4054 retains distinct24-case locally frozen scope study and12-case prospectively frozen history study. Generic Undo removes protected prior p3/3; explicit app separator preserves it3/3 (24 outcomes9 compensated/3 collateral failures/12 partial-no-Undo). In twelve history cases,CLEAN both gates compensate,CHANGED both refuse; content ABA restores base:px but revision2→4,TEXT_MATCH Undoes deletion/reinserts y2/2 while REVISION_BOUND refuses/unresolved2/2. Totals4 compensated/2 FAIL_WRONG_COMPENSATION/6 unresolved,raw audits6+6,nine rejecting mutations,no old rerun/pooling. Revision evidence is cooperative and assumes no writer after final check; #4090 separately tests reentry. Issue also reports separate retained45-case supplement18 correct/9 wrong/6 no effect/12 refusals; exact-path PR search found no delivery,so not asserted main-integrated or pooled. No generic history certificate,automatic Undo,arbitrary-app safety or product completion.

Selected text-lock lineage review — 2026-10-02 17:54 UTC:

• Held-key neutrality does not describe locked modifiers; preflight can expire before dispatch. Closed #4003, merged #4019: PASS_TEXT_LOCK_BOUNDARY_SCOPED on24 private US-map Tk sessions using unchanged backend methods. Exact/wrong/refused: EXACT_BACKEND4/4/0,PREFLIGHT_LOCK_GUARD2/2/4,DISPATCH_LOCK_GUARD4/0/4 (eight each). Preflight misses OFF→ON and needlessly refuses ON→OFF; dispatch guard reads current LockMask and refuses currently locked rows without emissions. Wrong Ab2 remains FAIL_TEXT_EFFECT despite backend completion/release. Raw audit10/10 mutations,four batch/outer exits0,configured Caps state retained; all24 Euro controls refuse. Auth construction STOP preserved. Assumes no mutation after final check; no automatic lock clearing,public CLI admission,German replication or generic modifier/IME guarantee.

• Equal boundary lock samples cannot detect mid-text ABA. Closed #4038, merged #4052: retrospective locally frozen15 sessions, PASS_LOCK_ABA_BOUNDARY_SCOPED. STABLE_OFF and LOCK_AFTER_TEXT yield aB2; LOCK_STAYS_ON and ABA_WHOLE yield Ab2; ABA_MIDDLE yields ab2 (three each). Both ABA schedules sample OFF before/after. Authored completion-only comparator nine false exact claims,boundary-equality comparator six; these are not upstream task-success claims. Reporting-only exact-value check gives6 PASS/9 FAIL and30 missing-or-foreign HOLD shadow reports; it prevents/repairs no input and grants no replay authority. Raw audit2,236 checks/no errors,13 mutations,15 units; old199 files unchanged. Publication helper temp-path string equality failure remains recorded,not byte-identical controls claim. Cooperative post-action evidence detects declared mismatches only,no authenticity,future stability,model utility or production promotion.

Selected closed-Issue backfill — 2026-10-02 17:55 UTC:

• Retention protection must follow each active reader lifetime. Closed #4062, merged #4078: PASS_READER_LOCK_OWNERSHIP_SCOPED,18 private Linux cases. INDEPENDENT_UNLOCK and SHARED_CLOSE allow exact B reads6/6/no early reclaim; SHARED_UNLOCK allows4/6,with both FIRST_RELEASE cases reclaimed before B reads/SOURCE_GONE. Initial probes all defer,final probes remove both files;54 worker exits (six registered73),two batch/outer exits0;raw audit3,896 checks/no errors,10/10 corruptions. Duplicated open file descriptions share flock; explicit unlock differs from closing one reference. Stable inode/cooperative collector only,no noncooperative deletion prevention,lock replacement,crash durability or model consumption proof. Old48-case retention study (partial overlap with #3978) preserved separately,not pooled;784-file capsule resolves its delivery gap without erasing historical HOLD.

• Truthful NOT_FOUND is not cancellation or future non-arrival. Closed #4069, merged #4073: PASS_LATE_REQUEST_IDENTITY_BOUNDARY_SCOPED on32 local SQLite/pipe cases,64 app/relay exits0,two outer exits0,raw audit12/12 mutations,10 tests. Both policies receive identical truthful read-only queries/arrival schedules; ATTEMPT_KEY commits22 effects in16 cases,four late duplicates plus two altered-payload second effects; stable INTENT_KEY commits16,zero duplicates/changed-payload extra effects. Intent key also binds exact parameters: mismatch is conflict,not reused success. Same-attempt redelivery deduplicates both; original-drop retries complete once. Cooperative single receiver with local atomic effect+receipt and retained intent lifetime,not GUI retry permission,external-effect atomicity,distributed exactly-once,expiry/crash safety or production adoption. Full new331-file capsule delivered; older42/45-case complete archives not included,remaining preservation debt disclosed rather than pooled.

Selected focus/lifecycle review — 2026-10-02 17:57 UTC:

• OS focus readback is not toolkit keyboard-recipient proof. Closed #4036, merged #4053: PASS_FOCUS_ACTIVATION_RECIPE_BOUNDARY_SCOPED,30 fresh Tk apps/three10-case batches. Top-level and native-child focus deliver to A in stable3/3,but B in before/after-activation changes3/3 each; child A XID readback succeeds9/9 without repairing Tk recipient. Ordinary click repairs A in stable/before6/6,but later change sends to B3/3. Totals12 correctA/15 wrongB/three blank no-input controls,raw audit12 rejecting mutations,all app/server exits0. Click requires supplied current geometry and no later focus change,not semantic target discovery/generic safe recovery. Old21-case and separate30-case focus archives reported as retained conversation-local evidence; full retrospective delivery path search found no PR,not asserted integrated or pooled. No public CLI/core/model/latency acceptance.

• Existence-only bootstrap can avoid unrelated lifecycle over-invalidation without inventing birth history. Closed #4039: merged #4048 preserves allocation01 STOP_SUPERVISION_TIMEOUT/HOLD_FORMAL_INCOMPLETE,nine complete rows,partial case9,actual outer-9,absent END/full-audit failure. Distinct freshly frozen allocation02,merged #4070,PASS_BOOTSTRAP_LIFECYCLE_SCOPE_SCOPED on18 new cases/36 classifications/three actual exit0 batches. Legacy MATCH9/UNKNOWN9 vs bootstrap MATCH12/UNKNOWN6: three known preexisting-bystander exits no longer invalidate target;nine ordinary positives still match,target-destroy/restart six UNKNOWN. Birth tokens stay null,allowlisted inputs exclude oracle/scenario. Raw audit12 inherited semantic mutants;six envelope controls explicitly postformal,not preformal;no old-row pooling. Candidate has extra existence information,not equal-information efficiency. Quiescent direct-root commissioning only,no coverage under unexplained loss,XID ownership/action authority or production guarantee. Older24-case subscription-handoff remains HOLD_PREREG_INPUT_METADATA_SEPARATION;full old ZIP conversation-hosted.

Selected closed-Issue backfill — 2026-10-02 17:59 UTC:

• Exact final value is not request-specific single-commit evidence. Closed #4042, merged #4051: retrospectively delivered36 SQLite cases, PASS_REQUEST_EVENT_SCOPE_SCOPED. State-only PASS33; own single commit plus current value match9. Naive unsupported completions24 split18 contradicted and6 UNKNOWN (not24 proven execution failures); candidate unsupported0. Event states APPLIED_ONCE12/NOT_APPLIED_IN_WINDOW12/APPLIED_MULTIPLE3/REQUEST_CONFLICT3/UNKNOWN6; causal necessity unresolved throughout. Complete trusted request/content/session/resource/instance/floor-bound event window is extra information,not equal-information efficiency. Same-value operation can commit without change,already-correct state can succeed without operation,own commit can later be overwritten. Raw audit15 rejecting controls,14 tests,original110 files unchanged,no rerun. Cooperative local atomic state+journal only,no authentication,external exactly-once,power-loss,GUI/model benefit or replay authority.

• KeyRelease representation, sampled neutralization and application text are separate. Closed #4041, merged #4050: retrospective15 private Tk sessions/30 simultaneous legacy+detectable observer views, PASS_X11_REPEAT_RELEASE_BOUNDARY_SCOPED. Fifteen authored down/up pairs yield45 characters; legacy60 releases,45 subsequent keymap samples still DOWN; detectable15 releases,zero sampled DOWN. Of45 legacy witnesses,30 later same-hold repeats and15 onset-compatible (including repeat-OFF); onset cause not isolated. Detectable negotiation is per observer connection and does not suppress app repeats. Original path-bound auditor fails intact relocation/confounds initial controls; preserved unchanged,postformal audit_v2 validates3,773 checks/13 mutations only after intact controls pass,no science rerun. Current backend already queries tracked-key/button state rather than counting release events: no production release-count defect alleged. Samples are server-logical,not event-time/continuous/hardware truth; no general backend/model/tempo benefit.

Recent idea/outcome reconciliation — 2026-10-02 18:06 UTC:

• Event coverage and time coverage answer different questions. Open #5694, merged #6771: new synthetic A02 PASS_METHOD_SCOPED, three scenarios/six routes, raw-only audit and four rejecting mutations. Heterogeneous durations: long-only covers 1/11 events but 90/100 time units; shorts-only covers 10/11 events but 10/100 time units. Equal-duration control agrees (2/4 events, 20/40 time); absent onset schedule stays HOLD with null denominators. This extends denominator validation, not live benefit, planner-stall attribution or human-tempo evidence. WSLc cgroup/swap memory-enforcement limitation retained. Earlier #5738 method result and #5940 parallel disposition are unchanged.

• An untested idea is not a statistical test; an abandoned started trial must not disappear. Open #5890, new intake A02 #6772 currently open/unmerged: synthetic PASS_METHOD_SCOPED, 25/25 rows, five rejecting corruption controls. Sixteen not-tested ideas remain in the qualitative funnel outside the statistical family; five started opportunities retain two FAILs (one abandoned), PASS/HOLD/STOP; four eligible claims are distinct from four p-value-free deterministic method/safety checks. Intake A01 STOP_MAIN_ADVANCED_BEFORE_CANDIDATE remains candidate/auditor/raw 0/0/0. Separately merged #6766 archives the older allocation-01 STOP_MAIN_ADVANCED_AND_STORAGE_ZERO preparation/comments, not missing raw evidence or a new result; older #5945 is separate. No empirical FDR guarantee, intake representativeness, GUI/model or safety benefit follows.

• When can a summary determine a safe action across every compatible full state? Open #6774 is a preregistered, still-unverified observation-fiber updateability successor to #5368; closed duplicate #6716 remains preserved. Proposed finite enumeration compares freshness+label against a gate requiring every compatible state to share the intended target/effect without forbidden collateral effects, plus the smallest predeclared distinguishing-field hint. Ambiguous/impossible/collateral/out-of-model acceptance is FAIL_METHOD; false refusal/wrong hint HOLD_UTILITY. No candidate/result is recorded at this check. Even a future T0 PASS would establish only completeness within the enumerated model, not live GUI safety, natural ambiguity rates or runtime utility.

Selected closed-Issue backfill — 2026-10-02 18:08 UTC:

• Same process, server, numeric XID and pixels do not prove the same window incarnation. Closed #3950, merged #3974: PASS_WINDOW_LIFECYCLE_BOUNDARY_SCOPED, four private Xvfb/client pairs, 24 comparisons/72 classifications and 28 captures. Deliberate destroy/recreate retains coarse typed and process/server identity MATCH 4/4; observed lifecycle rejects stale replacement 4/4, accepts unchanged/repaint/remap/fresh positives 16/16, and returns UNKNOWN after observer restart 4/4. Raw audit rejects 12 mutations; 24 offline malformed/gap/authority controls fail closed. Missing-Xauthority construction STOP remains. Observation-only, cooperative controlled reuse: no natural race rate, authenticated ownership, unseen-gap detection, reparenting coverage or atomic action guarantee. A later separate 24-case subscription-handoff study remains HOLD_PREREG_INPUT_METADATA_SEPARATION despite raw audit PASS and posthoc metadata-nonuse checks; opaque renaming does not retrospectively blind its inputs. Its original archive was conversation-local, not asserted fully hosted. Neither this HOLD nor #881's earlier finding is replaced by the scoped window result.

• An unresolved-gap deadline is not an inactivity timeout, or a guarantee that the event loop can run. Closed #4031, merged #4035: retrospective PASS_WAIT_SCHEDULING_BOUNDARY_SCOPED, 30 private pipe/SQLite cases, unchanged #3986 ELAPSED_80MS predicate. Relative waits renewed by later traffic notify at median 320.820 ms versus fixed gap deadline 80.369 ms; burst then quiet 140.757 versus 80.900; quiet both approximately 80.56. Pre-deadline predecessor recovery yields no notices. A synchronous 150 ms handler block still delays both (261.462 versus 180.966): timer eligibility is not execution availability. Raw audit reconstructs 666 worker rows/252 commands, all 30 worker exits and formal exit 0; nine unit tests and 12 rejecting mutations. Outer-tool timeout occurred after retained formal exit and is not a scientific failure. The 265-file capsule preserves original failures/publication history; delivery re-audit is not a formal rerun. No hard deadline, production scheduling defect, model-visible latency, task effect or recommended 80 ms default follows.

Selected closed-Issue backfill — 2026-10-02 18:10 UTC:

• Encoded-byte identity and decoded-pixel equivalence imply different cache contracts. Closed #4029, merged #4034: PASS_CACHE_MAINTENANCE_CONTRACT_SCOPED, 48 cases in two fixed batches, ORIGINAL versus byte-pin and pixel-check research wrappers. Under completed quiescent cache maintenance, original returns eight exact/six wrong-or-invalid images/two occupied-name refusals; each guard returns 14 exact/zero wrong/two refusals. Lossless re-encoding is byte-different but pixel-equal: byte pin regenerates, pixel guard reuses. All six occupied-name sentinels remain unchanged; raw-only stdlib PNG audit 1,434 checks, 12 rejecting mutations per batch. The expanded maintenance envelope does not invalidate the earlier producer-exclusive contract. Bounded RGB PNG without color-management metadata; no concurrent check/use guarantee, freshness/identity/authenticity, model receipt or measured validation-cost benefit. Preformal zlib metadata AttributeError is retained, not a failed formal case. Original 157-file capsule remains unchanged; closure is evidence delivery, not production adoption.

• Retired missing history is UNKNOWN, not proof of nonexecution. Closed #4027, merged #4033: PASS_RETIRED_OUTCOME_COVERAGE_SCOPED, 48 cases/three immutable batches. Lookup-only recovery duplicates retired same-payload effects 3/3 and adds conflicting changed-payload effects 3/3; coverage-aware recovery returns OUTCOME_UNKNOWN_RETIRED and adds none. Retired already-applied and never-accepted histories deliver the same query packet: the candidate also withholds three genuinely unsubmitted old requests, a liveness cost, not completion. Fresh independent current-epoch work succeeds 3/3 in both arms. All queries read only metadata/receipts and preserve DB bytes; raw audit rejects 12 mutations, ten unit methods. Construction Boolean-epoch oracle failure remains in the 173-file archive. The deliberately unfenced receiver isolates client choice and is not a production-defect allegation or weakening proposal; #531's receiver watermark is not rediscovered. No automatic epoch/ID restamping, exactly-once guarantee, GUI/model or speed claim. Historical merge preceded CI completion and external approval was not claimed; publication is separate from scientific disposition.

Selected closed-Issue lineage backfill — 2026-10-02 18:10 UTC:

• Duplicated retained descriptors preserve file identity, not independent read position. Closed #3984 remains STOP_COMMAND_EXECUTABLE_NOT_FOUND: the transcribed interpreter path was absent, shell exit 127, Python/worker/formal cases 0/0/0. Closed command-corrected #3995, merged #4016, is a distinct allocation with unchanged scientific sources/schedule, not a relabeling of that STOP. PASS_SHARED_OFFSET_BOUNDARY_SCOPED: 54 cases/108 real workers, named or unlinked immutable regular file, serial/both-reset/interrupted-chunk schedules. Unchecked dup exposes 12 incomplete acquisitions (empty/end or one-record prefix/end); extent-checked dup refuses those 12 but does not restore liveness; explicit-offset pread delivers all six exact records in 36/36 responses across 18 cases. Serial positives and unlink controls reconcile, all worker/supervisor exits 0, raw audit rejects ten corruptions. UNCHECKED_DUP separately remains FAIL_ACQUISITION_COMPLETENESS despite the boundary hypothesis PASS. The unchanged upstream reader legitimately classifies the bytes supplied and does not attest acquisition completeness. pread does not guarantee immutability/atomic snapshots with concurrent writers; extent equality is not authenticity/currentness. No ACK/model consumption, production adoption, performance, cross-platform or broader #3876 completion claim. Both first outcomes are preserved together; public freeze committed hashes, not already-public full source bytes.

Selected closed-Issue lineage backfill — 2026-10-02 18:12 UTC:

• Receiver validation cannot reconstruct version identity discarded by a lossy bridge. Closed #3994 remains STOP_OUTER_TOOL_TIMEOUT_PARTIAL: 49/132 complete rows, 245 classifications, no outer exit/END or full formal audit; observed child exits are not proof of outer completion. Closed fresh bounded-batch successor #4009, merged #4023, has PASS_VERSION_WIRE_BOUNDARY_SCOPED: 132 new Python-worker cases, 88 real Node conversions, 660 classifications, 11 fixed batches each exit 0; raw-only and batch-retention audits pass with eight rejecting corruptions. No old partial rows pooled. Legacy equality and receiver-only int64 each propose for six originally unequal pairs: proposal 9007199254740995 becomes 9007199254740996 through Node while current 9007199254740996 remains outside the bridge. Safe53 receiver, safe53 ingress+receiver and canonical decimal strings have zero inequality losses in this directed matrix, but receiver-only safe53 still accepts two whole-message source 1.0-to-1 normalizations; zero inequality loss is not original-type validation. Numeric safe-range refusal and canonical-string preservation have different domains, not equal high-numeric liveness. PROPOSE is authority-neutral; no production Node deployment, authenticity, freshness, action safety, model quality or speed claim. Public preformal manifests committed hashes, not complete source bytes. The 134-file capsule preserves v1 STOP/v2 result; earlier chat-local intent-relative learning FAIL remains separate, with full old ZIP not included in this PR.

Selected closed-Issue backfill — 2026-10-02 18:14 UTC:

• Checkpoint order trades missing retained notifications against possible redelivery; matching consumed bytes do not identify a restarted producer lifetime. Closed #3989, merged #4011: retrospective publication of v2 PASS_BOUNDARY_CONFIRMATION_4CELL_SCOPED, four directed cases, 18 actual reader CLI calls, six producer processes and two host SIGKILLs. Replacement inode and truncate/rewrite with identical prefix let legacy epoch/cursor omit the new lifetime's first record; fresh epoch with old cursor refuses, fresh epoch without cursor recovers both. Response-first retains three and redelivers three, missing zero; cursor-first retains zero and continuation returns zero, missing three. Local notification accounting is not repeated task actions or model consumption. v1 remains STOP_EXECUTION_SURFACE_TIMEOUT with only five complete rows from a planned 36 and missing terminal evidence; no pooling. The unchanged 173-file archive re-audits identically with 12 rejecting corruptions, no new formal run. Known HOST_CONTRACT boundary, not an upstream defect, power-loss guarantee, durable epoch allocator, exactly-once, integrated producer/host proof or performance result. Successful CI statuses alone are not scientific PASS: one recorded GTK workflow may exit 0 with an unscored STOP.

• An empty end tail is not producer finality; even zero exit needs the declared completion contract. Closed #3996, merged #4014: PASS_STREAM_FINALITY_BOUNDARY_SCOPED, 12 real subprocess streams. All 12 idle reads are empty/end with the producer still alive: baseline falsely says COMPLETE, candidate WAIT_PRODUCER. Nine complete late second notifications arrive after that idle point. Terminal states are three each COMPLETE (joined zero exit, owned final extent/sequence/digest seal and drained cursor), UNKNOWN (zero exit unsealed), PRODUCER_FAILED (sealed exit 7, both records retained), INCOMPLETE (missing last LF, cursor after record one). Zero-exit unsealed bytes are not declared wrong; failed producer exit does not erase delivered records. Raw audit rejects 12 corruptions; 70 original files retained. Trusted finite producer with no further writer after join, not upstream interactive_v17 support, crash-atomic seal publication, authenticated provenance, model viewing or task completion. Historical premerge CI remained in progress, so closure is not evidence of CI-green or runtime promotion; #3876/#57 remain unresolved.

Selected closed-Issue lineage backfill — 2026-10-02 18:16 UTC:

• A fresh worker's configuration difference does not establish the cause of a historical numerical mismatch. Closed #3939 remains STOP_EXECUTION_ENVELOPE_INCOMPLETE_DENOMINATOR: nine of 15 workers/two references retained, no final manifest or outer exit; unchanged auditor fails on the absent manifest. Nine observed exact matches are diagnostic, not a complete result. Closed fresh-seed batch successor #3958, merged #4002, completes 15 workers/three references and six rejecting corruptions but retains HOLD_THREAD_MECHANISM_NOT_EXPOSED. Every fresh DEFAULT reports five threads/determinism false versus parent one/true, and configured T1_D1 matches reference; nevertheless all five arms in all three seeds match full state, float32 logits and predictions exactly. No multithread contrast was exposed, so component H1/H2 support is not mechanism PASS or repair of #3911's earlier FAIL. PyTorch 2.10 CPU differs from the historical pinned 2.5.1 engine; first-arrival synthetic scope does not establish full 16-arrival equivalence or historical causality. Three timing samples per arm and partial order rotations are descriptive, not p95/60 ms acceptance or speed benefit. Both STOP/HOLD source and original evidence are retained across 238 archive members; publication re-audit performs no training. Source-preserving delivery is distinct from cross-engine reproducibility, model/GUI utility and a production fix.

Selected closed-Issue backfill — 2026-10-02 18:18 UTC:

• Arrival time cannot replace source acquisition age. Closed #3953, merged #3990: PASS_CAPTURE_AGE_BOUNDARY_SCOPED, 24 private Xvfb native-capture/IPC rows. Prompt valid frames admitted 3/3 by both policies; source-age rejects six deliberately delayed frames, whereas arrival-restamped comparator admits all six, including three whose delivered pixels differ from independently sampled current display. Fifteen missing/Boolean/reversed/future-bracket or stale-request controls refuse. Raw audit rejects 12 mutations. Capture age at prompt consumption 0.348–0.386 ms; barrier-held frames 150.477–151.088 ms old against a frozen 100 ms budget: these are directed timing witnesses, not natural latency or recommended defaults. Missing clock-namespace metadata, inherited-Xauthority and construction pixel/cleanup failures remain retained. Software acquisition brackets are not physical scanout; age admission is not semantic stability or action authority. No model consumption, same-model/load/application benefit or production restamping defect alleged; broader #2117 remains open.

• Neutral server input and zero current effects can coexist with a queued future application effect. Closed #3945, merged #3972: PASS_APP_QUEUE_NEUTRAL_BOUNDARY_SCOPED, 15 real private Tk sessions. All nine paused click cases show stopped app, responsive neutral X server and zero effects before resume. Paused-once and UNKNOWN_PENDING_APP/no-replay yield one effect each (3/3); blind retry yields two (3/3); live once one and no-input zero. Raw audit rejects 12 corruptions; all final input neutral and owned processes reaped. Construction Xauthority STOP and metadata incident remain in the 114-file capsule. This is a controlled app pause with ordinary callback effects, not failed delivery/cancellation, exactly-once recovery or model-policy efficacy. Callback-owned journal/file evidence is not independent physical-world measurement. Later #3998 changes effect placement to a deferred worker and is separate, not a revision of the callback-inline result. Historical merge preceded completion of two CI checks; full CI/automated approval was not claimed. Parent recovery/application contracts remain unresolved.

Selected closed-Issue backfill — 2026-10-02 18:19 UTC:

• A byte bound does not bound blocking path acquisition. Closed #3947, merged #3975: PASS_REGULAR_FILE_ADMISSION_BOUNDARY_SCOPED, 42 real reader processes. Exact reader produces no response by the ready-barrier-based one-second deadline in all six FIFO cases while still alive; this is finite no-response, not proof of infinite deadlock. One-open O_NONBLOCK/descriptor-regularity/snapshot adapter refuses nine FIFO/directory cases without exposing receipt/cursor. Across policies 18 regular complete/incomplete/symlink receipts match exact source/cursor bytes; six overflow cases reject. All readers/six live writers are reaped; raw audit passes and eight corruption challenges reject. Construction's outer timeout/unrecorded test-command exit remains separate from later test-only success. Local trusted regular-file admission is not a general execution deadline: slow/remote filesystems, faults, concurrent rewriting, parsing and output blocking remain excluded. Snapshot-copy overhead is unmeasured; no production promotion, producer identity, ACK/model viewing or speed claim.

• No observed cue plus a clear final frame is not historical absence. Closed #3948, merged #3966: PASS_TRANSIENT_NEGATIVE_COVERAGE_SCOPED, 24 private native-Xlib/Xvfb cases. Six covered clear traces qualify ABSENT_MIN_DURATION under a genuine >=6 ms cue-duration contract; six clear gaps UNKNOWN; six independently witnessed gap pulses are missed and baseline falsely says ABSENT, while candidate says UNKNOWN; six observed pulses remain historical SEEN after final clearing. Raw audit rejects 12 mutations; three preformal plumbing STOPs retained. Certification requires matching surface/clock and measured capture-bracket gaps strictly below the duration bound, including interval edges; nominal 2 ms sampling is insufficient. A separate posthoc finite-grid diagnostic (4,165 configurations/91,432 event checks) tests the strict boundary, not extra formal samples or a natural-GUI duration guarantee. Shorter/boundary-straddling cues, physical scanout and production completeness remain unproven. Historical SEEN is not current presence or action authority; related #2789 integration remains HOLD_INTEGRATION_INCOMPLETE, not solved by this component PASS.

Selected closed-Issue backfill — 2026-10-02 18:21 UTC:

• Matching consumed prefixes cannot bind notifications to a producer lifetime. Closed #3933, merged #3971: PASS_PRODUCER_EPOCH_BOUNDARY_SCOPED, 48 fresh producer/reader cases. Legacy caller-ID protocol aliases new-lifetime records 9/9 in identical/shared-prefix/mixed-suffix cases and rejects changed/short prefixes 6/6; validating an in-band producer epoch on every returned record refuses stale epochs 15/15 without exposing records/cursor. Same-lifetime append, reader restart and explicit fresh-epoch adoption succeed in 18 positive cases across both protocols; eight corruption challenges reject. Construction's 14-row timeout/partial fifteenth and absent terminal receipt remain excluded. Caller-owned distinct epochs are fixture assumptions, not authenticated/global production allocation; no automatic cursor reset, hostile-producer defense, crash durability, ACK/model consumption or exactly-once guarantee. The source contract already predicted aliasing; this is controlled boundary validation, not an undocumented upstream defect or integrated producer/host result.

• A no-effect cancellation recipe depends on widget and effect phase. Closed #3936, merged #3969: PASS_WIDGET_PHASE_BOUNDARY_SCOPED, 30 real private ttk.Button/Tk Scale sessions, all final server input states neutral, six rehashed corruption controls reject. Release-only cancellation retains an effect in both widgets (three each). Moving away before release suppresses Button effects 3/3 but not Scale effects 3/3, which already exist at post-press observation. Hand-scoped current capability permits Button move-away without effect, refuses Scale before press; stale receipts refuse both widgets (six total). Refusal is not performed cancellation, and physical cleanup does not reverse an already committed effect. Construction Xauthority/cleanup STOP, wrong Scale trough/no-effect gate miss and metadata failure remain in the 195-file bundle. This does not automatically certify arbitrary widget capabilities, acquire real-app current receipts, establish model-facing acceptance or production safety; #2197 and broader roadmap remain open.

Selected idea/outcome reconciliation — 2026-10-02 18:23 UTC:

• Exact fallback can preserve authored changes without saving comparison time. Closed #2001, merged #2015: six GUI-like cases retain PASS_EXACT_FALLBACK_REQUIRED_SCOPED, one task collision and all four changed cases forwarded by the full policy. Later source-bound interpretation identifies the equal-hash/unequal-pixel collision as panel brightness, not task_micro; old hash thresholds 512 individual pixels, not 64 block means. Old fallback_forward counts only the equal-hash fallback leg, so false on an unequal-hash task row is not full-policy suppression. Separate #2561, merged #3993, measures 300 actual Chromium raster pairs with a different 64-block hash: global misses 145/200 authored task changes; fallback/exact miss zero; fallback invoked 245/300. Charging both hashes gives mean fallback 377.981 microseconds versus exact 6.940, retaining HOLD_NO_RUNTIME_SAVING. Controlled authored templates are not natural prevalence or independently sampled applications; capture/decode/transport excluded, batch timing not call-tail latency. Compact GitHub evidence preserves original RGB/rows but not original PNG encodings (derived re-encodings are labeled); full originals remain in conversation archive. Old counts/source/result stay unchanged and broader #2561 transfer remains open.

• Missing evidence under a duplicate Issue number is not grounds to repeat an already consumed protocol. Closed-not-planned #493 and #3321: early closure audit retains HOLD_CLOSURE_UNSUPPORTED, then later reconciliation identifies exact matching 10-pair/20-case protocol in #455/merged #479, not nuisance-only #448. #493 stops before any new case as STOP_DUPLICATE_EXISTING_FORMAL; old rerun request is administratively retired, not fulfilled by a separately executed #493 run. Scientific outcome remains HOLD_LIVE_GIL_TARGET_DETECTION: 1 ms arm detects 8/10 versus 5 ms arm 1/10, failing the frozen >=9/10 gate. Neither duplicate closure nor recovered lineage promotes this to PASS or authorizes a shared GIL-default change. Distinct 12-pair #462 freeze and residual thread/process work are not pooled or silently cancelled.

Selected execution-history reconciliation — 2026-10-02 18:26 UTC:

• Raw semantic reconstruction is not missing outer-exit evidence. Closed-not-planned #3955 is allocation history under #3931, not a new host-persistence hypothesis. First #3931 allocation remains STOP_EXTERNAL_EXECUTION_TIMEOUT (25/45 rows plus partial, no terminal record). Second allocation retains all 45 rows/90 worker exits and a raw-only PASS_HOST_PERSISTENCE_BOUNDARY_SCOPED with ten rejecting corruptions, but overall HOLD_EXECUTION_EXIT_UNOBSERVED remains: absent shell exit cannot be supplied by worker exits, finished timestamp, later process absence or successful CI. Cursor-first omits 18 retained notifications, response-first duplicates 18, atomic retains without either in the declared process-exit cuts. These are local retained notifications, not model presentation or task effects. Merged #3968 preserves both outcomes; integration does not repair provenance or establish power-loss/production exactly-once behavior. Execution-envelope work is routed back to #3931 rather than another wrapper-only Issue.

• Execution-envelope repair proposals are not separate scientific advances. Closed-not-planned #3957 is #3932 image-content allocation history: earlier 81/100-row outer timeout and later STOP_OUTER_LIMIT_NOT_HONORED with 84 journal rows/missing normal launcher exit remain separate and unpooled. Requested 240-second outer allowance did not establish effective launcher lifetime; no image-content FAIL/PASS follows from the stop. Historical further-wrapper request is retired, with existing #3961 history and engineering retained under #3932. Closed-not-planned #3976 was not executed: proposed supervisor-exit repair routes to #3952. That parent retains 22 cases/88 reader calls and semantic checks, yet HOLD_SUPERVISOR_EXIT_UNOBSERVED and frozen auditor missing-receipt exception remain, not erased by a favourable journal. Administrative closure supplies no terminal evidence, new experiment or permission to rerun.

• A consumed setup STOP can remain closed while a distinct successor answers the scientific question. Closed #1563 remains STOP_FORMAL_ENVIRONMENT_INTERPRETER_MISSING_XLIB/scientific NONE, one invocation, zero GUI/scientific rows. Reopen audit later corrected itself after finding #1584 and #1601; it did not reclassify the original STOP. #1584 also stops before serialized science. #1601 fresh 64-session allocation completes PASS_XTERM_NATIVE_EXIT_TAIL_LOCALIZED_A4_SCOPED, four rejecting controls, no predecessor rows pooled. Effect-to-child-ready p95 21.038 ms versus child-ready-to-native-XTerm-exit p50 164.065/p95 214.750 ms localizes the dominant terminal tail in this fixture. It does not identify an internal XTerm/kernel mechanism, revise old endpoints or establish general speedup, model utility or production performance.

Selected closed-Issue backfill — 2026-10-02 18:27 UTC:

• Protocol correction supports one synthetic multi-skill routing result, not rehabilitation of the deviated predecessor. Closed #3912, merged #3925: fresh seed 3914, exactly 400 base updates/120 per adapter, one Windows RTX 3080 Laptop GPU invocation, PASS_MULTI_SKILL_ROUTING_SCOPED. Independent CPU audit reconstructs 24,576 rows: routed A/B/C accuracies 0.970459/0.953613/0.936035; sequential shared B/C 0.000244/0.912598; mean routed-minus-shared B/C 0.488403 exceeds frozen 0.10. Six invalid routes YIELD, base immutable, full-state snapshot/rollback passes; four audit corruption controls. #3895 remains HOLD_PROTOCOL_DEVIATION for executing 120 rather than declared 400 base updates. Initial length-limited GitHub staging corrected by re-uploading exact frozen local bytes, not a source change or formal rerun. Host-GPU evidence is not container parity, natural skill transfer, role-graph/model-task utility, production inference speed or authority.

• Prediction agreement does not satisfy a bit-exact resume contract or a frozen latency gate. Closed #3911, merged #3921: three local network-isolated Docker seeds, 16 feedback arrivals with fresh resume processes. Adapter/AdamW/cursor/base/chain states and six invalid-snapshot controls agree, but strict logits differ by up to 1.91e-6/3.82e-6/1.91e-6 although row predictions match. Update-only p95 209.22/101.57/68.32 ms exceeds <=60 ms in all seeds; frozen overall disposition FAIL remains. B accuracy 0.8613/0.9429/0.9314 is separate from persistence equivalence and not uniformly >=0.90. Initial auditor metadata-lookup failure and corrected audit retained; no training rerun or threshold relaxation. Later #3939/#3958 thread-setting work does not explain or repair this result. No full-state mismatch is invented, and semantic label agreement does not waive original logit/time criteria; no crash-atomic persistence, general reproducibility, GUI safety or production benefit follows.

Selected closed-Issue backfill — 2026-10-02 18:29 UTC:

• Strong IID accuracy does not rescue isolated near-boundary shift failure. Closed #3869, merged #3882: one three-seed Docker CPU allocation, FAIL_NEAR_BOUNDARY_SHIFT. Only CORRECT covariates move toward the decision boundary; CONTINUE/WATCH keep their baseline generators, correcting #3847's earlier treatment confound without reusing its data. Shift accepted accuracy 0.7412/0.7015/0.7830 misses >=0.95; CORRECT recall 0.187/0.026/0.301 misses >=0.95 despite coverage approximately 0.795–0.821. IID accuracy 0.9969/0.9947/0.9989 is diagnostic, not a passing shifted result. Latency, false-CORRECT, other-class gates, 1,536 boundary YIELDs per seed and five invalid probes pass separately. First auditor produces 6,156 errors; audit-only RNG-shape/float32/boundary reconstruction correction returns zero without retraining or relaxing scientific gates. The already-indexed #3892→#3899→#3906 lineage subsequently rechecks those immutable bytes: audit integrity can PASS while this model remains FAIL. Retained invalid input vectors and raw source binding address auditability, not real-world label validity, Astra intent, pixels/GUI/task effect, general capability or authority.

Selected closed-Issue backfill — 2026-10-02 18:31 UTC:

• Fresh-evidence anchoring can still expire before application; safe refusal is not complete recovery. Closed #1476 retains HOLD_LIVE_ANCHOR_NOT_SUFFICIENT_A2, eight immutable three-pair batches/24 pairs, no #1227 timeout rows pooled. SOURCE_APPLIED expires 24/24; CURRENT_EVIDENCE admits/effects 22/24 with zero wrong/stale effects and negative-control admissions zero. Two exact current observations still expire at capture-end-to-future application 132.919/121.298 ms against unchanged 120 ms validity. Admitted future-to-effect p95 1.210 ms does not cover these misses. Frozen binary auditor reports FAIL_LIVE_ANCHOR_TRANSFER_A2/parent_evidence_anchor; retained reconciliation applies the preregistered HOLD for safety-correct incomplete recovery rather than changing raw/auditor output. Seven mutations reject but are not discriminative evidence because the unchanged base already fails that auditor. Reopen/closure audits corrected their mistaken successor-state readings; #1510 remains open for the separate tail-decomposition question. No lease lengthening, batch rerun, natural-tail prevalence, model/task benefit or promotion.

• Omitted focus dependency can redirect a delayed effect, but a favourable first output cannot waive a no-rerun protocol. Closed #1752 remains STOPPED_EXECUTION_PROTOCOL_UNAUTHORIZED_RERUN/scientific NONE: warning-prefixed stdout failed JSON parsing and formal was mistakenly invoked twice. First output retained descriptively, second excluded; neither pooled into closed fresh #1762. Output-capture-only successor invokes once with raw bytes retained before parsing: PASS_DYNAMIC_FOCUS_FOOTPRINT_X11_A2_SCOPED, four arms six cases each, audit 10/10, four rejecting controls. Focus-bound scheduler overlaps independent Y write while A targets X, serializes focus-changing B; unbound negative redirects A to Y 6/6. Synthetic both-effects median 221.018 versus 120.498 ms (100.520 ms difference) is fixture-scoped, not general app speed. Broader selector/dependency completeness, toolkit/modal behavior and external focus mutation remain unvalidated. Corrected reopen audit recognizes this successor without reclassifying #1752's STOP.

Selected closed-Issue backfill — 2026-10-02 18:37 UTC:

• #3890, closed; merged #3902 — Serialize receipt-gated role adapters into inert JSON and reload them in fresh processes. One formal orchestration, fresh seeds3788–3790, one builder and two independent loaders per seed in offline local CPU Docker. Each loader reproduced all12,288 held-out predictions exactly; A/B/C accuracies were .967529/.903076/.928223, .970215/.905273/.900635, and .961914/.923584/.947021. Both fresh graph generations traversed A→B→C; malformed-package/route/stale-receipt controls yielded without fixture emissions or state mutation. Scoped synthetic PASS; corrected raw-evidence audit PASS_AUDIT_SCOPED/errors0, no model rerun. Preserve first audit STOP on manifest/path assumptions and a freeze-hash transcription typo; construction-only seed retirements and tamper-probe graph mutation remain excluded. Seed3789/C is only narrowly above .90, not robust transfer. SHA integrity is not authenticity; synthetic effect receipts are not real GUI effects, execution authority, general skill reuse, portability or concurrency. Result and correction disclosure.
• #3909, closed — Recover the exact executed source and six-session raw evidence before trusting historical provenance. Recovery-only successor to #3866/merged #3881 ended STOP_UNRECOVERABLE. Current main and merge contain identical six source blobs, but all six disagree with RESULT.md's recorded SHA-256 values, independently checked in the intended cached offline/read-only container. Seven PR files omit the six raw rows, XRecord JSONL, workbooks and audit JSON; original branch is absent and available history/Actions/conversation searches surfaced no exact source/evidence bundle. Preserve the original reported PASS and unresolved review, qualified as not durably source-bound or independently re-auditable: this is not proof that its reported fixture outcome was false. No GUI session/new experiment was run; a replacement allocation cannot validate or rewrite the historical six sessions. Recovery result and search boundary.

Selected closed-Issue backfill — 2026-10-02 18:40 UTC:

• #3903, closed; canonical merged #3916 — Distinguish module-executed launch calls from deferred AST calls without importing the target. One frozen offline/read-only Docker Desktop linux/amd64 formal run: eight controls agree with an independent stdlib verifier, AUDIT_PASS/errors0, source/result bindings match. The pinned session_entry.py has unguarded session_map01_v13.main() at line18, supporting the actual-target import STOP; deferred function/lambda calls illustrate overbreadth of the old whole-tree scan. Target import/execution, game, model and input all0. PASS_STATIC_IMPORT_BOUNDARY_SCOPED is syntax classification, not transitive import safety or gameplay correctness. Preserve predecessor #3896/#3900 engine-mismatch setup STOP. Closed UNMERGED #3920 explicitly records overlapping parallel duplicate work, excluded from another result/replication count; its separate launcher preflight path failures are not attributed to canonical #3916. Canonical result.
• #3744, closed as superseded — Synchronize the receiver's Xlib keymap after layout changes. Diagnostic only: fresh German connection decodes keycode19 levels0/1 as 0/=, old pre-layout connection as 0/) with two pending MappingNotify events; no formula input emitted. This supports the stale receiver-cache explanation for #3741, not formal delivery success. Disk-full/WSL EIO/Docker unavailability stopped before frozen formal rows; later official-image pull violated the no-network constraint and remains a setup deviation/STOP, with no formula/formal row. Closure relies on the separately indexed #3794/merged #3800 fresh-receiver result, not a completed #3744 allocation. Later redundant branch-ref cleanup adds no science; predecessor diagnostics and STOPs stay unchanged. German private-X11 receiver scope only, not Calc/physical keyboard/IME/general-layout benefit. Supersession.
• #3885, closed; merged #3891 — Require every matching/intervention/ledger/scorer/clock/window gate before attributing a declared task effect. Fresh independently frozen successor leaves #1873's manifest-failing source bundle unchanged. One offline/read-only pinned OrbStack linux/arm64 formal invocation enumerates15 Boolean gates × two useful/harmful polarities:65,536 unique rows, useful1/harmful1/unresolved65,534, candidate/oracle disagreement0, failed-gate causal promotions0, authority grants0. Separate auditor imports neither candidate nor oracle, reconstructs all rows, checks30 polarity-specific single-gate corruptions plus3 laundering controls. PASS_MAP01_MATCHED_CAUSAL_TASK_EFFECT_CONTRACT_SCOPED certifies authored deterministic-fixture admissibility only; attested matching facts are inputs, not measured causal identification in stochastic MAP01. No gameplay, recovery efficacy, comprehensive utility, model/input/runtime benefit or live-allocation authority; no retry or repair of predecessor provenance. Formal result.

Selected sampler-stream lineage review — 2026-10-02 18:43 UTC:

• #3851, closed — Test whether changing only the online minibatch RNG stream rescues rank4 adaptation on matched seeds. Preserve two distinct source/branch histories, not one contradictory outcome. Branch issue-3851-minibatch-rng-local-20260921 retained trained pre-rollback curves but frozen audit FAIL_AUDIT_INTEGRITY/STOP: rollback reset live adapters before final scoring, producing10 curve_final_mismatch errors (corrected from the first five-error narrative). Separate needle-lora-rank4-minibatch-seed-3851-20260921, merged #3884, retains STOP_RUNNER_INDEXING_FAILURE: one CUDA invocation completes seed3451 base training then indexes original support-row4 into a one-element seen list; first online update fails before adapter optimizer/evaluation, stdout0 bytes, no metric audit. Neither STOP is sampler-quality evidence; do not attribute the other branch's curves to empty stdout or pool executions into replication. Distinct indexing STOP.
• #3865, closed not_planned — Independently audit retained pre-rollback curves without training again. Sole CPU auditor stops before parsing/decompression at STOP_PREDECESSOR_STDOUT_SHA256: byte hash matches, but uppercase frozen digest is compared case-sensitively with lowercase hexdigest. No row findings, no retry; closure is not audit success. Fresh #3875, closed; merged #3878, normalizes digests and retains PASS_CURVE_ONLY_AUDIT_SCOPED: one CPU-only audit verifies160 B curve points ×4,096 rows, independently regenerated A/B labels, A final rows and declared state/route controls, exactly10 explained rollback mismatches/no others. Arrival16 legacy mean .94189453125/shared .941015625, paired delta −.00087890625: collapse not reproduced/no material rescue in this retained five-seed synthetic setup. Audit-only, no optimizer/CUDA or repaired #3851 final metrics; these curves belong to the nonempty-payload branch, not #3884. Curve-only result.
• #3887, OPEN/reopened; merged #3894 — Distinct corrected-index allocation leaves legacy collapse unreproduced, with independent verification unresolved. One local RTX3080 invocation, no retries: reported HOLD_LEGACY_COLLAPSE_NOT_REPRODUCED, legacy .94189453125/shared .943896484375, gain .002001953125, feedback p95 9.8888ms. Bundled auditor NameError ea before verdict means no complete row-level independent PASS; #3875 audits another payload and cannot fill this gap. Separate merged audit-only #3922 retains STOP_INPUT_PROVENANCE before scoring:326,706-byte embedded raw payload verifies, but pinned runner/auditor hashes disagree with retained blobs; old checksum file has two warning lines/eight mismatches. Later offline Docker read-only recheck confirms raw-envelope/seed identity only and that LF/CRLF does not explain source discrepancies, not model-output reconstruction. Original HOLD and indexing STOP remain unchanged; current open state overrides body text claiming historical closure. Five synthetic seeds/host CUDA only, not container training, real task benefit, production latency or causal support for RNG rescue. Latest limited provenance recheck.

Selected vision-input / epoch-composition lineage — 2026-10-02 18:47 UTC:

• #3827, closed; merged #3842 — Test local vision JSON mode separately from structured-schema transport. One Qwen2.5VL7B request through an offline OrbStack runner/loopback Ollama bridge, HTTP400 Failed to load image or audio file, completed answers0/retries0/actions0. Independent audit retains STOP_OBSTAC_OR_MODEL_TRANSPORT, not semantic failure. Later inspection finds all280 predecessor PNG rows use invalid filter255: earlier sips/CRC/decompression acceptance did not establish format validity. Preserve #3204's earlier schema-request/HTTP-body-retention STOP separately; changing format to json alone did not yield an answer. STOP and successor clarification.
• #3843, closed; merged #3855 — Validate the image fixture before interpreting model compatibility. Fresh one-cell corrected filter0 RGB448×280 PNG request, same pinned Qwen digest, HTTP200/exact READY, retry/fallback0/actions/effects0. Separate isolated audit PASS_QWEN_VISION_BACKEND_PROBE_SCOPED/20 checks, including CRC, row lengths/filter validity and provenance/request-response bindings. Supports malformed-image explanation for the earlier load STOP, not general JSON-schema support, visual reasoning or composition. Issue bodies list different seeds3827/3843 despite the later only-image/same-options narrative, so do not claim every request option was identical or an exhaustive causal isolation. Proposed Gemma3 comparison was deferred/unrun. One-cell result.
• #3858, closed; merged #3872 — Join image/UI-tree/state only under declared epoch and skew bounds; abstain before inference on conflict. Fresh valid-PNG/JSON-mode formal:3 authored scenarios ×4 policies=12 rows, exactly9 eligible Qwen calls, retries/actions/effects0, independent offline OrbStack audit PASS_MODEL_FACING_EPOCH_COMPOSITION_SCOPED/23 checks. Typed policy exact3/3: aligned READY, same-epoch40ms skew within50ms READY, cross-epoch conflict ABSTAIN before model. Zero-skew rejector exact2/3; latest-channel and best-effort each2/3, answering BLOCKED instead of required ABSTAIN on mixed epochs. These are three synthetic decision cases, not a population quality/unsafe-action rate. Parent #3204 remains OPEN: held-out missing/reordered/duplicate/stale/malformed/beyond-bound channels, new epoch during delayed response and separately measured recovery/task benefit are not established by this first rung. No GUI effect, general multimodal recovery, efficiency or production reliability; every predecessor STOP remains unchanged. Matrix result.

Selected ordered-event / retained-attempt lineage — 2026-10-02 18:50 UTC:

• #1884, closed; merged #1896 — Preserve admitted event order while reducing logical deliveries; shard exhaustive evidence for durable completion. One logical formal allocation/nine immutable shard processes, no reruns,299,593 sequences/2,054,353 prefix checks, sequence/identity/session-projection/metadata errors0, delivery reduction in299,584 sequences, priority-sorted comparator six reversals, malformed controls5/5, independent audit PASS. PASS_ORDERED_INTERRUPT_BATCHING_A3_SCOPED is representation after arbitration, not live transport/ACK/retry/model fidelity or token/latency/task benefit. Predecessor import-shadowing and outer-timeout STOPs #1871/#1876 contribute0 scientific rows; optional postformal corruption-driver timeout remains separate. Later integration review notes existing ordered receipt lists: eight directed compact-report expansions preserve view/raw identities/authority-none, both select v2, not v3-specific or exhaustive re-audit. Fixed four-event wrapper has no demonstrated call-count benefit; async pending-stream behavior remains untested, no queue implemented. Integration assessment.
• #3830, closed; merged #3839 — Retain complete report despite short delivery; never replay request-only unknown attempts. Distinct Docker29.8 cached-image synthetic characterization: one invocation, writer returns11/115 bytes, delivery_incomplete/exact report readback; child exit73 before report, unknown_or_incomplete/nonreplayable, recorded invocation count1. Independent host audit PASS_SYNTHETIC_SHORTWRITE_AND_UNKNOWN_RECOVERY, three corruptions rejected. Postmerge review explicitly marks this redundant relative to #3752 real attempt-status and #3727/#3729 production checks, not new #3711 completion evidence. No production CLI imported. Same formal01 label also appears in the separately retained #3832 Docker28.5.1 setup STOP with different source/image: precreated /out prevents any case. Preserve both histories; do not infer one result overwrote or repaired the other. Scope correction.
• #3833, closed; merged #3837 — Distinct parent-mount successor to #3832's consumed setup STOP. One Docker28.5.1 formal02, standalone synthetic harness: final-LF-only prefix83/84 remains JSON but INCOMPLETE_STDOUT_WRITE, exact retained report recovered; request-only child exit23 stays unknown/process-unknown/report-absent/replay-false/backend0; complete84/84 control dispatch1. PASS_SYNTHETIC_MECHANICS_SCOPED, independent host audit16 checks/errors0. No production CLI, real OS short write, backend/task effect or broad durability claim; no pooling with #3830's11/115/exit73 corpus or rewriting earlier STOP.
• #3834, closed; merged #3846 — Valid JSON alone does not certify complete JSONL delivery. Real frozen public CLI with only dispatch facade patched, offline OrbStack arm64, five cases. Producer exits0/accepts351 bytes; relay delivers exact350-byte prefix missing only terminal LF. Parser sees completed JSON; framing-aware caller refuses completion, read-only attempt-status/review recover exact retained report, snapshots unchanged/dispatch1. Complete, invalid-prefix, request-only unknown and occupied-destination controls pass; second-container raw audit five rows/errors0, PASS_FRAMING_GUARD_SCOPED. Unlike standalone #3830/#3833 this exercises public CLI, but fault is a synthetic relay and dispatch is mocked: no real pipe/network loss, backend effect, power-loss proof, Docker Desktop parity or broader #3711/#3808 closure. No production code changed. Formal framing result.

Selected closed-Issue diagnosis / CLI integration — 2026-10-02 18:52 UTC:

• #3862, closed not_planned — Check app-owned selected-cell evidence before attributing lost leading text to readiness. Proposed six-session coordinate-only versus selection-gated Calc comparison never ran: one private Docker/Xvfb construction click at89,200 changes UNO selection from Sheet1.A1 to A3 while the same Calc window stays active. STOP_CONSTRUCTION_TARGET_MISMATCH, formal0/6, no text/save/task operation (one pointer click, not zero native input). Hypothesis abandoned rather than repairing frozen condition in place; no explanation of #2704's earlier16-versus116 effect mismatch. Local construction JSON/workbook/source hashes recorded, bounded PR search returned none, so not claimed delivered as a formal/main result. App selection and active-window identity are different checks; no crash-matrix completion, reliability rate, production integration or retry. Construction STOP.
• #3845, closed not_planned — Resolve lazy X11 test mocks through explicit sys.modules, then avoid duplicate implementation. Historical dotted-target resolution failure was identified, but source comparison found current main already contains fake namespace/backend/session module setup and report partial-write/fsync/no-residue tests. This Issue made no new branch/code change and claimed no fresh Docker/CI result; closure is duplicate scope, not new test PASS. Preserve closed/unmerged #3735/#3737's earlier CI failure unchanged. Source presence does not establish broader filesystem durability or close #3711 adoption gates.
• #3835, closed; merged #3838 — Keep Python dependency diagnostics on stderr while retaining JSON-only CLI stdout. Original live Docker/Xvfb observe exited0 and saved raw report/image, but Xlib authentication warning preceded JSON; caller correctly rejected parsing before dispatch, no replay. Integration redirects Python-level diagnostics only during retained observe/dispatch and emits structured response afterward, preserving report/one-call/outcome semantics and direct Python API behavior. Candidate tests: Windows38 with one Linux-only skip; Docker linux/amd64 public CLI94 passed, observe/dispatch warning cases check JSON-only stdout, stderr diagnostic and retained report. This is engineering regression coverage, not a fresh live task/recovery allocation. Native file-descriptor writes are outside the fix; redirection is process-wide, so no arbitrary concurrent CLI isolation claim. Warning contamination is not byte truncation; do not silently strip unknown lines, infer clean Xauthority configuration or erase original failure.

Selected online-LoRA efficacy / measurement lineage — 2026-10-02 18:55 UTC:

• #3807, closed; merged #3819 — Higher adapter rank did not rescue the frozen online schedule. One local Windows RTX3080 five-seed3451–3455 formal, FAIL_ONLINE_RANK_CAPACITY_GPU: rank4-online B .0156/.0149/.0222/.0134/.0193 versus rank2-online .9414/.9326/.9331/.9583/.9368 and rank4-batch .9519/.9299/.9534/.9580/.9531. Mean rank4-online lift −.9233; every batch gap>.03. Feedback p95 33.335ms and base/route/exact snapshot/rollback integrity pass; independent auditor reconstructs20 held-out sets ×4,096 rows. Quality FAIL is not erased by fast updates, integrity PASS or closure. Preallocation disk-full STOP remains separate from later formal; distinct logistic CUDA smoke stopped before first optimizer step at STOP_DETERMINISTIC_CUBLAS_SETUP, no metrics, not a formal LoRA sample. Host CUDA evidence, not container or CPU/GPU comparison; no causal instability diagnosis, real task/transfer or production authority.
• #3820, closed; merged disclosure #3822 and read-only audit #3829 — Per-feedback timings are not per-feedback learning curves. HOLD_LEARNING_CURVE_NOT_RECORDED: #3807 runner evaluates only after16 arrivals; retained final row predictions and16 timing samples per online arm cannot reconstruct unmeasured intermediate accuracy. No model/prediction/optimizer rerun. Preserve valid final-accuracy FAIL while narrowing completeness, not infer when/why collapse arose. #3829's pinned offline OrbStack audit adds source/evidence review, not another training allocation or repair.
• #3826, closed; merged #3844 — One predeclared half-learning-rate intervention also failed its rescue gate. Fresh seeds3461–3465, paired rank4 .04/.02 online arms share initial tensors/support/minibatch schedule; rank2-online and rank4-batch controls,128 updates/16 arrivals. One local RTX3080 formal and independent CPU row/curve audit; required curves retained as declared48 online points per seed with4,096 predictions each (three online arms ×16). FAIL_LR_HALF_RESCUE: LR.02 mean .8977, gain .0111 against .04 control mean .8866 versus required .50; seed3463 also misses batch gap<=.03. Feedback p95 9.9964ms/integrity/routes/full-state rollback pass. No retry/tuning. New seeds and changed shared sampler schedule mean .04 control's noncollapse does not isolate the cause of #3807's collapse; matched-stream #3851 successors are indexed separately with their own STOP/audit limits. Synthetic five-seed efficacy only, not real feedback, runtime benefit, generalization or production latency.

Selected structural-audit / byte-provenance lineage — 2026-10-02 18:57 UTC:

• #3676, closed — Audit strict event schema/order/cardinality and contradictory refusal fields, not just three planted controls. Exact #3675 retained trace passed its original auditor, yet unexpected/duplicate/reordered events, extra fields and would_call_bridge=true with bridge_called=false were accepted; documented six-argument CLI hit usage error. Merged #3687 retains one offline Docker Desktop source-snapshot validation:4/4 tests,9 declared mutations rejected, separate-container CLI PASS_OFFLINE_STRUCTURAL_AUDIT/errors0. Report text also says original3+additional8, so do not silently change its recorded9 into11. Later merged #3683 is a different advanced source revision: Windows construction5/5,21 structural/type mutations rejected, exact raw/freeze-pair binding, direct/CLI parity; still STOP_CONTAINER_VALIDATION in that record. Earlier #3687 snapshot does not certify that newer head. Python True==1/False==0 substitutions motivated exact non-bool integer checks. Preserve daemon/disk/CRLF setup STOPs and original bytes; no XRes/GUI/model rerun, runtime safety claim or arbitrary-corruption completeness. Closed superseded #3684 had unrelated-deletion diff and is not the integration result. Snapshot distinction, newer construction scope.
• #3688, closed; merged #3704 — Parsed-equivalent JSON is not exact frozen evidence. Separate OrbStack arm64 offline provenance matrix PASS_RAW_BYTE_BINDING: original raw accepted,seven raw-byte/digest/source/freeze mutations refused before structural PASS, direct/CLI agree8/8; corrected independent verifier in fresh container reconstructs8 receipts/errors0, formal matrix not rerun. Initial verifier/setup defects remain in AUDIT_ATTEMPTS.md. Earlier explicitly named Docker Desktop allocation stopped before suite/CLI: daemon absent, low disk and uncommitted freeze/README; retain it separately, not relabel as OrbStack parity. Corrected review withdrew the claim that source.sha256 covered only v2 or had Git-blob hash mismatch: declared inputs matched, Windows CRLF caused that observation; unbound local freeze/README limitation remained. These are byte-identity checks for one retained bundle, not authentication, all-mutant completeness, XRes runtime reproduction or production assurance; #3675 observation and every predecessor STOP remain unchanged. Distinct matrix result.

Selected role-graph / streamed-adaptation lineage — 2026-10-02 18:59 UTC:

• #3778, closed as superseded — Compose learned A/B/C adapter nodes with generation-bound fixture receipts. Original single host-CPU formal output exceeded capture limit: STOP_RESULT_CAPTURE_TRUNCATED, complete raw/result hash absent, no formal metric or independent audit verdict. Fixed-seed construction screening is not formal evidence. Never reconstruct missing bytes or call closure a PASS; successor #3780 is another allocation, not recovered original output. Capture STOP and supersession.
• #3780, closed; merged #3786 — Losslessly packed prediction rows make the bounded role graph auditable under output limits. One separate Windows host-CPU formal, same synthetic family/seed: A/B/C .971191/.918213/.922607 over4,096 each; flat/graph predictions equal, two generations traverse A→B→C, eight invalid-transition controls yield without cursor mutation, stale old-generation receipt refused, complete snapshots exact/base unchanged. Both independent audits errors0; separately identified postformal v2 also verifies12,288 rows/configuration, first duplicate-probe ADVANCE then repeated YIELD and raw gzip digest, no candidate rerun. Raw keeps stale embedded allocation label needle-role-graph-3775-v1; successor identity is bound by frozen runner/preregistration, not silently rewritten. Scoped graph composition uses authored edges and synthetic effect receipts, not real app effect/authority/general skill transfer/concurrency/durable process reload. Later #3890 separately tests process reload; its PASS does not retroactively fill #3778's missing evidence. Docker unavailable: host CPU, not container verification.
• #3769, closed; merged #3774 — Stream rank2 role-specific updates as feedback arrives, but enforce absolute competence as well as batch parity. One host-CPU/one-thread deterministic seed3445,16 feedback arrivals ×8 updates=128; A .9465, online B .8877 versus matched batch .8069. Both B arms miss>=.90, so FAIL_ONLINE_ADAPTATION_QUALITY despite online gain+.0808 and update p95 3.205ms. Four invalid/stale/missing routes yield; full state round-trip/rollback and immutable base pass; frozen independent auditor recomputes12,288 final row predictions and agrees FAIL. Retained curve examples at1/2/4/8/12/16 arrivals .0315/.0542/.5522/.8804/.8809/.8877 describe this one task only, not general learning efficiency or real feedback. No Docker/GUI/model-provider/task effect/runtime promotion, retry or tuning; base retention through explicit dispatch is not evidence of unrestricted global adaptation. First outcome.

Selected readiness / abrupt-request persistence lineage — 2026-10-02 19:02 UTC:

• #3633, closed — Absent/ambiguous app identity must stop before geometry/input, and readiness does not certify process cleanup. Two distinct OrbStack arm64 allocations are retained, not one overwritten outcome. Merged #3639: three app identities stable across two snapshots/four negative unit controls pass, geometry/focus/input/model/network0, X socket removed; original runner/auditor PASS is preserved but final HOLD_PROCESS_LIFECYCLE_UNVERIFIED because Calc ownerPID95 differs from tracked launcher66 (exit255), owner exit/reaping receipt absent. Separate identity-missing allocation stops at first Inkscape before transitions/input, no admitted identity/task effect; auditor not run because its frozen contract accepts only full PASS ledger. Closed UNMERGED #3646 publication collided with #3639 path; merged #3788 rescues exact blobs under distinct identity-stop path without rerun or altering HOLD record. Empty-Xauthority/unauthenticated-Xvfb discrepancy is a successor hypothesis, not proven cause. #3645 and Calc lifecycle successors are separate; neither readiness result establishes full persistent #2499 focus/modal/replacement/return session, task benefit or general process safety. Lifecycle correction, distinct early STOP.
• #3752, closed; merged #3766 — An abruptly interrupted request-temp write remains visible, UNKNOWN and nonreplayable. Allocation01 is unchanged STOP_LOCAL_CAPACITY_BEFORE_ALLOCATION/unexecuted. Fresh offline/read-only OrbStack arm64 allocation02: child exits29 after flushed134-byte strict prefix in .request.json.tmp, no final request/report/backend marker; two real public attempt-status reads exit2/unknown/replay-false/residue-visible with identical snapshots. Public CLI reuse exits2/REQUEST_PERSISTENCE_FAILED/operation_invoked:false, backend stub uncalled and snapshots unchanged. PASS_SCOPED_REQUEST_TEMP_UNKNOWN_NO_REPLAY is one injected process-boundary test, not actual backend/task effect or full #3711 adoption. Frozen audit-v2 was never invoked because it would write into read-only input, so no audit-failure outcome is invented. Separate audit-successor03 once returns PASS_AUDIT_V3/errors0, independently binds raw/source/runner/temp/status/exit/snapshots on read-only evidence with separate output. No formal rerun, safe-temp-deletion procedure, power-loss durability, every-filesystem guarantee or general recovery success; inspection preserves unresolved state rather than completing/replaying it. Separate raw-only audit.

Latest reconciliation — 2026-10-02 19:09 UTC follow-up

• Focused German X11 receiver: harness STOP is not erased by a later successor. #3784 proposed a mapped InputOnly receiver with verified focus and receiver-only XTEST control after the earlier xev delivery failure. Its closed/completed state is not its own formula PASS: every formal-01 row stopped before German-map application or candidate planning/emission because the runner incorrectly required empty XLookupString text for KeyRelease, although the receiver correctly returned a for both press and release. Merged #3789 retains raw hash 34f6e18f…9859, frozen audit-v1 FAIL_AUDIT_INTEGRITY, and a distinct post-result forensic audit-v2 with zero errors; the latter confirms the early STOP, not candidate delivery. Another #3784 comment reports an OrbStack formal-01 under raw db8578fa…19b9 and audit 17b9da06…51b8, with eight missing-map findings after the same early gate. These differing raw/audit identities must not be merged into one byte-bound run. The separately recorded #3794 / merged #3800 delivered exact =B2*A2 in three German rows and one US control, with zero-event Euro refusal, empty release state and independent audit PASS. This is private-receiver X11/XTEST evidence only, not Calc, task effects, physical keyboards, IME or general-layout readiness. #3797 is closed as a duplicate: no formal allocation was started there because #3794 already covered the corrected hypothesis; it is not an independent replication. Earlier STOPs remain unchanged.

• Linux selector test: diagnosis changed from package resolution to optional dependency isolation. Closed #3754 initially proposed explicit implementation imports for an Ubuntu AttributeError while Windows/macOS passed. A later Ubuntu run showed X11 files were present but importing the real backend failed on missing optional Xlib; the sparse-checkout-missing-package explanation was not supported by that observed checkout. Merged #3757 instead injects synthetic backend/session modules in one test file, checks display/target forwarding and environment immutability, and changes no production code. Its PR reports the exact Linux test passing in an 88-test Ubuntu CLI suite and a green three-OS CLI matrix; the supplementary duplicate selector job was still queued in that report. Local Docker/WSL fallback was unavailable at the time, so it supplies no container result. Scoped mocked dispatch/test-import repair, not real X11/backend or GUI-effect validation; original failed CI and initial hypothesis remain historical evidence.

• Multifidelity control variates: pilot cost, contrast covariance and stopping validity are separate gates. Open #6155 proposes cheap paired predictions to improve estimation of an unchanged high-fidelity continuous cost endpoint, never to replace real safety/effect scoring. Its initial host-only T0 reported METHOD_PASS_SCOPED, then explicitly withdrew the same-total-cost interpretation because a 20,200-unit pilot was omitted from the 2,100-unit comparison; retain that raw as mechanics evidence, not cost benefit. The separately cost-complete v2 returned METHOD_FAIL_COST_MATCHED: positive-correlation CV MSE .025175 versus Y-only .009824, negative .026889 versus .010662; the arithmetic auditor PASS does not turn this into efficacy PASS. T0c amortized one pilot across a stable synthetic population and observed first ≥5% point crossings at K=4/5; independent-seed T0d simultaneous bootstrap intervals left K=4/5 unresolved, resolved no gain at K=1/2 and gain at K=10–100, with zero-correlation no gain throughout. These are finite synthetic, approximate uncertainty results, not transferable exact thresholds or real GUI cost evidence. T0e tests covariance of within-task route differences rather than misleading arm-level correlations: merged #6291 retains 2,700 reconstructed Docker records and FAIL_METHOD_GATE; the positive-difference K=20 point gain was 21.08%, but its family-wise interval crossed zero, despite the negative control and signed-negative prediction behaving as intended. Its formerly pending delivery is now merged, without changing FAIL.

  Latest T0f asks whether bounded predictable corrections support anytime stopping and whether fitting beta on the current scored row erases a real mean. Open/unmerged #6792 records STOP_PREREG_SOURCE_HASH_MISMATCH and says candidate/auditor/retries 0/0/0 with no raw. A later Issue correction explicitly admits local candidate/auditor ran once each before actual committed blob hashes were checked, reports four source mismatches (two-byte differences, only consistent with—not proof of—line-ending drift), and retains 20,000 rows with passing replay on a separate results branch strictly as forensic exploratory output. The records disagree on execution counts; preserve both and do not describe the allocation as zero-execution or the local replay as formal evidence. Formal disposition remains STOP, not an anytime-validity PASS; no retry or main promotion is established. The literature addendum and proposed predictable/bounded-stream contract are hypotheses/design guidance, not repository experimental findings.

Latest reconciliation — 2026-10-02 19:10 UTC follow-up

• Safe distinguishing probes: a fixture-level successor, not real-app transfer. Open #6147 asks whether safe observation sequences can distinguish visually aliased decision states, or reach a common permitted terminal effect without identifying the initial state. Merged #6787 retains two distinct Xvfb allocations. T1 is STOP / NOT_EVALUATED: adaptive emitted ten rows and the fixture 44 oracle events, but candidate exit 1 at Xlib display teardown prevented the auditor and other two arms from starting; partial rows are not a hypothesis result. Fresh T2 narrowly handles only ConnectionClosedError during Display.close, leaving observation/input and unrelated exceptions as failures. Its three arms produced 30 matched rows, independently replayed with zero errors and five mutation controls rejected. Adaptive had six verified in-memory effects (separable A/B, action-equivalent C/D and convergent I/J at current terminal Z); no-probe and one-step had two C/D effects each; all arms used zero unsafe probes. All seven formal containers exited 0 with no OOM. H_PASS_SCOPED is only for the authored disposable Xvfb fixture, not real applications, model performance, workflow usefulness, latency or production authority. Post-run README edits no longer match the preserved preregistered source manifest; this documentation discrepancy is disclosed separately, while execution-bearing entries and raw manifest verify. Earlier A02 STOP and A03 audit-control-coverage FAIL remain unchanged; T2 does not close the broader transfer question.

• Opaque-model canaries: detect a planted comparability break without proving model identity. Open #6001 proposes a fixed non-action deck around remote-model route blocks, with non-detection meaning only no shift detected in that deck. Allocation-01 STOP_DOCKER_BIND_SOURCE_UNAVAILABLE and allocation-02 STOP_GUEST_MOUNT_LAYOUT_MISMATCH remain NOT_EVALUATED; host rehearsals do not repair their absent formal container evidence. Merged #6318 preserves allocation-02; its administrative publication/branch recovery is not a scientific result. The distinct merged #6088 FCFS fixture shows audit-probe interference: shared canary and equal-cost sham each added 6.0s B-arm queue wait, whereas no-probe/isolated-canary did not, across 3,840 synthetic rows. This is not observed provider queue behavior. Merged #6104 separately retains a seven-scenario/three-policy detection fixture, 21,000 route outcomes and 420,000 probes: in-deck and within-block shifts yield distinct HOLDs; schema, prompt/context and snapshot changes are separated; stationary/out-of-deck clean decks remain MODEL_IDENTITY_UNVERIFIED. Audit errors were zero and nine mutation tests passed; fixed A→B order and out-of-deck blindness remain limits. Hosted checks were queued at that merge, not demonstrated green.

  Open/unmerged #6786 adds PASS_RUNTIME_REPRODUCIBILITY: exact frozen #6104 source reproduced its deterministic raw SHA-256 under native WSLc, with a separate auditor reconstructing 21 groups, zero errors, and 9/9 tests. The changed image/runtime and byte-identical output establish bounded replay only, not new scientific replication, provider stability, latency or actual model identity. Noncanonical construction source/raw is retained separately and excluded. WSLc warned cgroup/swap limits were unavailable, so accepted memory flags are not enforced-isolation evidence. Composed end-to-end canaries and probe-induced service interference remain distinct design questions; component canary success cannot prove route stability.

• Safety-aware backpressure: two finite counterexamples, not a queue-stability theorem. Open #5372 proposes downstream-pressure admission while keeping mandatory safety work separate. Early 12-tick local toy traces suggested less backlog, but delayed telemetry without a reserved safety lane left mandatory work pending/dropped; a reserved lane repaired that toy invariant without clearing stale normal work. These directional unmerged probes do not establish stability. T1 allocation-01 stopped before candidate/auditor because the harness compared canonical image-reference strings rather than the matching digest; retain STOP_HARNESS_IMAGE_IDENTITY_STRING_MISMATCH. The distinct allocation-02, merged in #5810, is PASS_SCOPED_COUNTEREXAMPLE: with 36 fixed offers, BASE/FIFO, FAST/FIFO and FAST/BACKPRESSURE verified 10, 4 and 14 tasks by horizon; stale attempts were 50, 64 and 13. All six mandatory events met the frozen service bound across 16 cells. Survivor-selected p95 (5 versus 30 ticks for the two FAST arms) is a latency/throughput tradeoff, not independent proof; closed-loop offers differ and are not matched-load comparisons.

  Open/unmerged #6791 reports the separate WSLc route-set-expansion A01 method fixture: an optional locally cheaper route reduced total local ticks 12→6 but verified-by-horizon fell 5→3; narrow pressure admission restored five of six offers. Every arm served the three modeled mandatory safety events; audit PASS_METHOD_SCOPED had zero errors and rejected five frozen corruptions. One candidate and one auditor, no retries; cgroup/swap warnings mean requested memory limits are not claimed enforced. This is an authored finite offered-load counterexample, not a Braess-equilibrium proof, real-agent result, production scheduler recommendation, fairness result or general queue-stability guarantee.

• Cutoff-local optional guards: method recovery after lost raw does not establish a live causal effect. Open #6451 proposes regression-discontinuity identification/refusal near an already-permitted optional guard threshold. Its read-only eligibility search did not evidence a suitable implemented cutoff; absence from a bounded search is not proof of global absence, and no safety rule may be relaxed to create a cohort. Original open draft #6476 preserves STOP_RAW_OUTPUT_NOT_RETAINED: candidate exit 0 but transport-truncated stdout left no complete raw/hash, auditor=0, no scientific PASS/FAIL. The original naive comparator also ignored its intended bandwidth. Distinct T0b, merged in #6789, used file-backed raw before a separate read-only-input audit: six cases/72 rows reconstructed, zero errors, five effective corruptions rejected, PASS_METHOD_SCOPED. Local-linear effects were 2.0 for the planted sharp case and 0.0 for the smooth zero-effect control; within-bandwidth naive means were respectively 11.0 and 9.0, demonstrating remaining trend confounding. Noncompliance gave first stage .5 and fuzzy local effect 2.0; sorting, coincident transition and timestamp heaping were refused. This establishes only the frozen synthetic identification gate, not a live GUI threshold, real causal guard benefit, global policy improvement or safety claim; original STOP remains unchanged.

Latest reconciliation — 2026-10-02 19:14 UTC follow-up

• Layout-aware Calc formula symbols: bounded runtime improvement with a saved-file witness. Closed #3668 arose when the public MCP rejected =A2B2 before any emission, forcing explicit layout-dependent key chords. Merged #3670 extends the existing live-keymap punctuation planner to equal/asterisk, including shifted/unshifted mappings, unsupported-level refusal and whole-program refusal on a later unsupported character; 22 relevant tests are reported. A fresh private WSL/Calc use entered =B2A2 directly as one text operation, with A2=7 and B2=13. Final screenshot and independent saved FODS inspection agreed on the formula and value 91. The first post-input image was still redrawing and a later explicit observation was needed. The 42-file evidence package initially described as draft in the Issue is now merged in #3671. This supports only bounded symbol entry and this task transfer, not general layouts/Unicode, matched speed/token savings, direct-host presentation or a formal container result; #3667's refusal/workaround remains unchanged.

• Sequence-monotone resident policy: construction and finite GTK effects are distinct rungs. Closed #3588, merged via #3591, directly exercised the exact upstream research candidate and reproduced delayed-replacement rollback plus delayed same-generation observation acceptance. A separate guarded variant matched an independent oracle over nine traces/27 prefixes, with 6/6 tests and three mutations detected: PASS_CONSTRUCTION_PREFIX_ORACLE, no GUI/model/input invocation. An initial construction summary aggregation failed after writing the ledger; the recorded correction was construction-only, not a formal retry. Closed #3593 then retained separate formal-01 fixture-path STOP (zero invocations) and formal-02 source-root import STOP (one runner, zero GTK rows). Distinct formal-03, merged in #3607, completed 36 rows/108 prefixes: upstream emitted the stale rollback action and incremented GTK once, guarded emitted none; effect-off rows stayed zero, positive controls worked, key releases and process cleanup passed, 72 raw frames were retained, and four corruptions were detected. A later auditor-only bad source-root mount first produced eight source-hash mismatches; corrected isolated replay reproduced the saved PASS/hash without rerunning formal GUI rows. PASS_INDEPENDENT_GTK_SEQUENCE_GUARD_AUDIT is finite synthetic GTK/Xvfb evidence only, not production event-source ordering, broad GUI safety, latency or user benefit.

• A second GTK effect-binding result has a separate committed-evidence HOLD. Closed #3642, merged via #3651, reports one five-prefix allocation with counters 1,1,1,1,2: two admitted Space emits, no effect from stale/rejected prefixes, verified release/cleanup and 33 audit checks with no failures. The saved audit decision string is the inherited PASS_TK_EVENT_WITNESS_AND_PUBLIC_MCP_CONTINUATION_SCOPED, a stale template label explicitly disclosed rather than rewritten; it is not evidence of Tk/public-MCP execution. Open #3655 checks whether another agent can reproduce that result from committed bytes. At the frozen Git tree, eight of ten manifest entries match exactly with zero present-file mismatches, but fixture.log and xvfb.log are absent; exact committed auditor exit 1 reports precisely those missing-log failures. HOLD_COMMITTED_ARTIFACTS_INCOMPLETE is independent of the historical #3642 PASS. Container execution was separately STOP_CONTAINER_UNAVAILABLE, not a container audit. Original #3661 closed unmerged; merged #3836 rescued only its four-file audit record, excluding unrelated runtime history, and independently reproduced the HOLD from the frozen tree. Its retained JSON is a reviewed summary, not byte-identical stdout. No original missing log bytes were recovered in these checks and none are synthesized; no formal GUI allocation was rerun or upgraded.

Latest reconciliation — 2026-10-02 19:16 UTC follow-up

• Public MCP continuation: completed execution with an unmet frozen event gate. Closed #3618 corrected the predecessor auditor's embedded-source revision expectation while preserving archive/runner/fixture and three-allocation design. Early Windows STOP_INFRASTRUCTURE / NOT_STARTED reports are environment-and-time-specific, not proof that the later macOS/OrbStack allocations did not occur. Merged #3621 retains three fresh observe → one dispatch → separate read-only observe allocations: exact independent markers, empty release, fresh post-observation IDs and changed initial/final PNGs were reported. The first audit stopped on an existing output mount; closed audit-only #3620 initially lacked discoverable inputs, then audited the later-published immutable bundle using a fresh child output path. Its 122 checks returned exactly three fixture-received-save-chord failures: Control_L was logged but no s keysym. HOLD_OR_FAIL remains the result, despite correct saved effects and completed execution/closure; no criterion was weakened or GUI allocation rerun. Merged #3627 adds a distinct 65/65 native integration suite, after a smaller image failed 28 tests for absent NumPy; that suite does not upgrade the event-gate HOLD. No registered-host presentation, model-visible acknowledgement, latency/cost or broad reliability finding follows.

• Passive Tk witness: a fresh fixture correction, not retrospective promotion. Closed #3628, merged in #3636, places a passive per-widget audit bindtag before the save class binding. Three distinct formal OrbStack/public-MCP allocations each logged one ordered Control_L+s witness, saved an exact independent marker, observed a fresh post-action image, and verified release/cleanup. Seven construction controls passed and the successor auditor passed 102/102 checks: PASS_TK_EVENT_WITNESS_AND_PUBLIC_MCP_CONTINUATION_SCOPED. The inherited 122-check HOLD and output-path STOP stay unchanged. Supplemental Xvfb/X11 integration passed 9/9; canonical manifest is 96 files/408,499 bytes after supplementation, with an accidental byte-identical nested duplicate removed, not lost evidence. Parallel #3635 is now merged but construction-only: its first three direct-XTest attempts retain failures/STOPs, fourth passes after explicit mapping/focus with 12 events and one marker receipt. Recovery adds the four declared zero-byte fixture.log files to match manifests; this reconstructs declared package bytes, not proof the original process emitted those files. Twelve isolated tests pass; the manual hosted workflow was not dispatched in that record. Construction attempt 04 is not public-MCP delivery or another formal replication. The scoped successor does not establish host/model-visible acknowledgement or close #3370.

• Proxy representations: 28 completed rows still fail evidential gates. Closed #3619, merged via #3623, is a fresh allocation after #3610's startup STOP. Formal-02 completed 28 GTK/Xvfb rows with no runner exceptions, but independent audit returned 44 errors and HOLD_EVIDENCE_INCOMPLETE / HOLD_OR_FAIL_PROXY_BINDING_EFFECT_UNIT. Positive effects, no-effect yields, stale/replacement refusals, unavailable and macro-failure yields were each 4/4; these partial gates are not overall PASS. In all four ambiguous-target controls, the click/effect preceded the second fixture's ready event, so pre-dispatch ambiguity was not tested. Independent release verification was absent for every row because query_pointer was called on the wrong object and the exception swallowed. Static-review limits also remain: PID is a caller hint rather than observed XID owner, all four representation arms emit the same fixed click without consuming their representations, and the output argument is not bound to the frozen path. Neither completion nor closure demonstrates representation superiority, safety, usability, latency, model or production benefit. The separately tracked correction proposal does not repair this immutable HOLD or authorize a repeat.

Latest reconciliation — 2026-10-02 19:18 UTC follow-up

• Stale-source refusal: publication closure is a separate outcome. Closed #3613 addresses 13 manifest-listed logs omitted from the original result merge by the repository ignore rule. Its first read-only recovery check found no originals and retained HOLD_PUBLICATION_INCOMPLETE; hashes alone were not treated as recoverable bytes and no logs were synthesized. Subsequently merged #3616 added the exact original logs at their original paths. The separate committed-tree audit reports all 159 original manifest entries present with matching lengths/SHA-256 and zero mismatches: PASS_PUBLICATION_CLOSED. The original manifest/report and runtime allocation were not changed or rerun. This closes publication integrity only, not model-visible acknowledgement, delayed/no-image, cancellation/disconnect or broader effect claims.

• XRes PID/start-ticks guard: observed refusal, but original provenance and published raw remain qualified. Closed #3555, retained by merged #3568, tests a research wrapper against recycled XIDs. A superseded pre-formal design was not run after its unsafe failure path and release classifier were found defective. Formal-02 reported identical XID/geometry/pixels across different XRes owner PIDs/start ticks; the old alias was refused before any bridge/native emission or effect, and the fresh control produced the expected effect with empty release. The initial auditor failure remains STOP_AUDIT_NOT_REPRODUCED; a later raw-only auditor reported measured-gate reconstruction and eight corruptions detected, but exact original freeze/source-manifest bytes were missing. The qualification is OBSERVED_GATE_PASS / INDEPENDENT_RAW_AUDIT_PASS / PREREGISTRATION_PROVENANCE_INCOMPLETE, not a fully authenticated formal PASS. A still later Issue integrity note reports the committed raw begins with a literal truncated-output warning and matches neither its declared hash nor that hash after stripping the prefix. Thus the published raw cannot be byte-authenticated by that check; do not present the earlier local audit as verification of the currently retained byte string. Earlier reported results stay historical, and neither merging nor later successors cure the missing originals or raw discrepancy. No production/default-runtime or remote-X11 security claim follows.

• X-server lifetime identity: two distinct provenance-complete fixture results. Closed #3574, merged via #3583, separately tests same-display Xvfb restarts rather than replacing #3555's process guard. One formal invocation, four restart pairs/16 classifications, zero input/retries: all pairs reused top-level/root XIDs and exact 153,600-byte pixels while owner PID/start ticks changed. Exact #881 CURRENT_TYPED accepted stale cross-generation identities 4/4; an observation-side opaque lifetime comparison rejected them 4/4 and preserved same-generation acceptance 4/4. Twenty-four negative controls and seven raw-copy corruptions passed independent audit. This is PASS_XSERVER_LIFETIME_BINDING_REQUIRED_SCOPED, not a production token issuer, authority gate or security implementation.

  Closed #3581, merged via #3585, is another separately frozen four-pair/16-row matrix after predecessor source-manifest-binding STOPs. Exact freeze↔manifest and GitHub source readback were checked before its sole run; the independent auditor in a separate container verified source/result hashes, 28 negative controls and nine corruption challenges. Its scoped PASS does not reuse or upgrade #3572/#3578/#3579 STOPs, the older #902 partial row record, or #3555's provenance-incomplete guard. Both successful matrices establish a deterministic Xvfb typed-identity contract requirement only, not production persistence, remote X11/Wayland, real task utility, latency or general reliability.

Latest reconciliation — 2026-10-02 19:21 UTC follow-up

• API/CLI/MCP observation equivalence: separate geometry/source allocations, not a route winner. Closed #3561 follows an unavailable predecessor source freeze and proposes one read-only call through each public route. The current Issue body and first preregistration specify 500×260; a later comment and merged #3565 instead call 320×240 the exact requested geometry. Preserve this documentary discrepancy rather than silently harmonizing it. #3565 retains a distinct frozen-source 320×240 allocation: one API, CLI and stdio MCP observation each, equal decoded RGB/PNG hashes, 36/36 independent audit, no input/model/network. It excludes the earlier 500×260 allocation from its 320×240 claim and retains a harness preflight error before route calls. A separate #3561 comment reports a 500×260 allocation with its own audit-v2 PASS, preserving initial auditor failures; these dimensions/raw identities must not be pooled. Merged #3562 independently records a 500×260 route unit for #3557 with different image hashes and 36/36 audit. Its single timing samples use unequal call boundaries and establish no causal speed/efficiency advantage. All are bounded capture-equivalence observations, not model-visible task comparison, token/cost, routing-policy benefit, recovery or general GUI reliability.

• Recycled XID stale alias: an independently reconstructed experimental failure. Closed #3551 retains an earlier bootstrap preflight STOP (runner accidentally entered without DISPLAY/results, no observe/mint/input) separately from its later formal allocation. In the latter, p1 exited and distinct p2 reused XID 2097152, geometry and exact 153,600-byte pixels. The old alias was accepted and emitted three native events, with independent p2 effect [1,109,118]; a fresh control then produced [2,212,118]. Result is FAIL_STALE_HANDLE_RETARGETED_REPLACEMENT. The separate auditor's PASS_INDEPENDENT_RAW_RECONSTRUCTION verifies that failure; it does not make the candidate PASS. Verified empty releases, clean exits, freeze/hash/PNG checks and five corruption controls are reported, with raw and audit retained on main. Inherited descriptive issue/parent fields are stale and preserved; allocation/source/freeze identify the run. This demonstrates one frozen experimental NativeHandleBridge/private-Xvfb failure, not default-runtime or product-wide behavior, and tests no mitigation. The later #3555 guard's missing original provenance/published-raw discrepancy remains a separate limitation, not a repair of this FAIL.

• Explicit review revokes aliases only under its tested handoff condition. Closed #3467, merged via #3468, tests an explicit review_window call after replacement by a different-XID process. Binding revision advanced 0→1 and scope/store changed; old alias was refused as MISSING, bridge emission counter stayed zero, and no p2 effect/XTEST event preceded the fresh control. The fresh alias emitted three XTEST events and produced the expected independently recorded effect: PASS_REVIEW_REVOKES_PREDECESSOR_ALIAS for this one research path. This is neither implicit handoff nor XID-reuse coverage; it does not contradict or fix #3551's recycled-XID failure and changes no production source/default authority. Broader host/model visibility, task utility, latency and release readiness remain untested.

Latest reconciliation — 2026-10-02 19:25 UTC follow-up

• Inkscape edge admission is not saved movement. Closed #3465, seed 991126, retained STOP_PRE_INPUT_CONTROL_CHANNEL_EOF: a high-contrast source-bound patch was presented, but no decision/input occurred. Separate closed #3470, seed 991127, kept a PTY alive and admitted the unchanged edge point with 15 emissions/empty release; no object was selected and saved SVG stayed x=50, so FAIL_TASK_EFFECT remains. Cleanup flags stayed false: HOLD_CLEANUP_UNVERIFIED despite reported terminal processes/container exit. Later Windows setup STOPs are separate environment/time records, not erasure of this formal failure.

• Selection then explicit save succeeds in one fixture, with qualified audit portability. Closed #3483, seed 991131, visually confirmed selection after click-only before six Right/50 ms/Ctrl+S. Final image and independent saved-SVG parse report x=62/y=50/w=40/h=30/no transform: PASS_SELECTION_GATED_SAVED_MOVE_SCOPED. Prior failures and raw false cleanup flags remain unchanged; producer-auditor image-path correction was post-result. Closed-unmerged #3490 was superseded by merged #3632, rescuing only its eight scoped files. Offline replay/reviews verified 62 manifest entries, action/source/reply links, per-stage releases and saved geometry; lifecycle is producer-attested, not independently portable (no captured inspect receipt). Legacy audit.py mutates audit.json; stdout-only replay was added without rerunning input. Later review findings on canonical path-set pinning, initial-source image-byte hashing and stale legacy-auditor instructions are not established fixed by the cited integration record. Merged #3497 renders relocated historical evidence with explicit path mapping; this is not a new allocation or lifecycle proof. No broad GUI, efficiency or production claim.

Latest reconciliation — 2026-10-02 19:31 UTC follow-up

• Raw-byte repair is not rehabilitation of an old XID result. Closed #3430 found the predecessor's declared hash bound an escaped-newline representation, not attached JSONL; the original bytes were unavailable and remain HOLD_RAW_BYTES_UNAVAILABLE, not reconstructed. Fresh merged #3457 binds an exact 831-byte raw file but retains HOLD_DECOY_EFFECT_UNVERIFIED. Merged #3460 keeps allocation-03 at HOLD_READINESS_UNVERIFIED_STALE_EFFECT_OBSERVED after rejecting a supplied readiness boolean; distinct allocation-04 verifies exact bytes/readiness/cleanup and demonstrates FAIL_STALE_XID_RETARGETED_CURRENT_WINDOW. Old XID input changed p2's DOM to EFFECT-1 before the fresh control advanced it to EFFECT-2; the decoy left p2 unchanged. Publication/byte-gate success does not make this behavior PASS or test a production composite-generation guard.

• Timeout/cancel journal across process restart: finite Chromium controls completed. Closed #3253 originally held for absent restart source/image bundle; a separate six-case fixture claim and later three-allocation packages are distinct evidence identities. Merged #3407 retains six timeout/cancel replays rejected with zero dispatch/effect and fresh effects 3/3. Merged #3411 separately retains missing-release/cleanup-failure replay rejection 3/3; #3415 retains duplicate-terminal and stale-source rejection 3/3 each, with fresh controls effective. Separate raw hashes/auditors are preserved. PASS_CHROMIUM_TIMEOUT_CANCEL_RESTART_FAIL_CLOSED_SCOPED is receipt-journal/process-restart fixture evidence, not production crash durability, graceful browser restoration or general cancellation propagation.

• Old browser receipt admission: real-path evidence follows a weaker contract candidate. Closed #3366 identified #3361's hard-coded old_allowed=false as insufficient. Early six-row contract-candidate audit PASS remained HOLD_OLD_RECEIPT_ADMISSION_UNPROVEN, and a source-missing HOLD was recorded separately. Merged #3384 later exercised generation/process-start comparison across three fresh A→B browser allocations: old receipts had zero dispatch, fresh receipts produced saved:abc DOM effects. Reported PASS_OLD_RECEIPT_ADMISSION_SCOPED follows an initial failed auditor that wrongly expected zero fresh-control dispatch; corrected PR run and manual rerun are disclosed, not described as one untouched audit invocation. No production recovery or cross-app claim follows.

• O3 evaluator preflight is not live transport. Closed #3120 records three Docker tests admitting only current complete bound authority-free evidence and refusing invalid evidence; its original pure-evaluator result stays preflight-only. Later merged #3224 separately adds the #3215 two-app GTK/X11 live matrix, 12 retained rows, mutation controls, denied emissions zero and independent audit. Superseded/closed status does not promote the old preflight into a live O3 result or broaden the later source-window-fix scope.

Historical ideas — reconciled 2026-10-02 19:34 UTC

• #1651 — Does querying at most four samples cost less than fixed eleven? Closed; analytical non-identifiability PASS, not measured savings. The formal comment records PASS_TEMPORAL_SAMPLE_COST_NOT_IDENTIFIABLE_FROM_COUNT_SCOPED: identical counts permit query-cheaper, fixed-cheaper and tie constructions. For one commensurate metric the condition is H + Q(k,p_q) < F(11,p_f). Only equal presentation and constant marginal cost c reduce it to H < (11-k)c (k=4: H < 7c). Exact Fraction frontier 4,913 rows and linear special-case 4,160 rows had zero mismatches; four corruption controls and independent audit passed, source readback 3/3, formal1/reruns0. Wall time, tokens and money need separate/Pareto comparisons; counts alone establish none of them. Matched empirical measurements remain separate downstream work.
• #1671 — Can retained temporal evidence identify that empirical break-even? Closed; PASS_RETAINED_TEMPORAL_BREAK_EVEN_NOT_IDENTIFIABLE_SCOPED. Five pinned families include one actual-model usage family, one local query-latency family and the analytical contract above. Admissible matched F/Q/H plus correctness rows = 0; no causal cost estimate was emitted. Five corruption controls/audit passed, source readback 5/5, formal1/reruns0. The bounded main-branch search had zero model-query hits; this is not proof of absence in private/unpushed work. Local ring latency is not model continuation H, four packed sources are not fixed eleven, and unmatched usage deltas are not causal savings. The audit establishes an evidence limit, not a negative efficacy result or a new experiment.
• #1462 / merged #1473 — Independent audit of retained model branch-cost identifiability; successor to #1199 without rewriting v1. Closed; scoped non-identifiability PASS. The delivered PR, not just the Issue's planned gate, records PASS_RETAINED_MODEL_BRANCH_COST_NOT_IDENTIFIABLE_V2_SCOPED: 24 decisions, 144 grouped/ungrouped candidate pairs, same-image pairs0, fully matched pre-branch pairs0, admissible different-branch pairs0; independent analyzers agree, corruption controls4/4 rejected, frozen source readback4/4, formal1/reruns0. No causal per-branch cost and no new model/provider/GUI/X11 actions. This is not an economic/runtime winner; fresh matched measurements remain distinct downstream evidence.
• #3209 / merged #3203 — Close the formal-artifact gap for staggered subagent orchestration v2. Closed after evidence correction. The earlier STOP_ORCHESTRATOR_RUNNER / HOLD_LOCAL_ONLY_SCOPED recorded inability to dispatch through that connection; retain that historical boundary. Later comments identify successful Docker harness and independent audit in run35469737524, uploaded and independently downloaded artifact: 108 transitions, nine cases, four policies, unsafe control rows42 versus CONDITION_STAGGER unsafe0; integrity/transitions-unique checks passed. Final disposition is PASS_STAGGERED_ORCHESTRATOR_FORMAL_ARTIFACT_SCOPED. This closes publication/formal-artifact absence, not broad production/generalization, model or GUI efficacy; parent #3158 history and deterministic-fixture scope remain unchanged.

Historical ideas — reconciled 2026-10-02 19:38 UTC

• #2794 — Repair mixed-app geometry, focus and replacement identity after #2792. Closed; preserved STOP/FAIL diagnostic, not integrated success. Adding openbox/wmctrl and rejecting 1x1 candidates exposed an Inkscape 710x659 window, but bounded Calc readiness checks found no matching visible identity: STOP_MIXED_APP_LONG_SESSION before transitions. A separate minimal Calc launch progressed to a 15-event ledger with successful focus drift and distinct Chromium replacement, but missing Calc modal/geometry receipts produced FAIL_MIXED_APP_LONG_SESSION. A stricter PID/class/title control again stopped rather than accepting an auxiliary candidate. These are separate allocations, not pooled PASS evidence. Original #2792 is closed unmerged and explicitly retains FAIL. The audit points to #2937 for the remaining identity-admission boundary; no Writer substitution or inferred identity establishes the missing result.
• #2892 — Retain exact raw GTK gate-repair summary and provenance. Closed; evidence-publication improvement only. Merged #2894 retains the #2884 summary.json plus SHA-256/provenance manifest; merged rescue #3114 also preserves the original worker branch. This does not repeat the allocation or broaden GTK acceptance. The subsequent audit explicitly leaves individual event/input/effect/cleanup receipt completeness unresolved here and points to #3184 after #2936/#2895. Summary retention alone is not full formal #2606 acceptance, and prior #2851/#2876/#2885 evidence remains unchanged.
• #3000 — Test target replacement using demonstrably distinct XIDs, avoiding #2994's identity aliasing. Closed; distinguish historical HOLD from fresh scoped PASS. The initial merged #3004 allocation used old2097155/replacement4194307: source mismatch rejected with zero emissions, but old-target focus failure retained possible-partial-input uncertainty, hence HOLD_IDENTITY_FAIL_CLOSED_NOT_PROVEN despite useful control. #3015/#3019 later typed only the narrow pre-emission boundary; they did not retroactively turn that allocation into broad PASS. A fresh Docker/Xvfb :149 allocation in merged #3170, independently audited in the Issue, records PASS_TARGET_REPLACEMENT_DISTINCT_SCOPED: destroyed old2097155 produced zero program emissions, verified empty release and no effect; live replacement4194307 produced four emissions, an effect receipt and verified empty release. A separate comment-only allocation, Xvfb :99 with IDs2097155/2097163 and result SHA-256 146ce1884253adfba7dc361a3612cd87f8b47e0f8c31b65ae90f3c808ca7cbfa, reports PASS_TARGET_REPLACEMENT_DISTINCT_IDENTITY_SCOPED without branch/PR operation; do not conflate its publication status or identity pair with #3170. All positive claims are limited to their GTK/X11 topology; Chromium same-XID replacement failure and cross-platform identity remain separate, and #2994/#2999 historical evidence stays intact.

Historical ideas — reconciled 2026-10-02 19:39 UTC

• #2758 — Freeze a corrected six-task model-IPC allocation and retain its first terminal stop. Closed; runner delivered, allocation still not demonstrated by this evidence. Merged #2764 adds source-hash admission, first-stop retention, partial-report persistence and zero-retry accounting, not a six-task PASS. Source assembly first stopped with STOP_INFRA_SOURCE_MATERIAL_UNAVAILABLE (invalid clone HEAD / corrupted archive extraction); no model/GUI allocation occurred. A separate container smoke HASH_GATE_PASS validated only integrity. A subsequent selective-source preflight stopped with STOP_SOURCE_HASH_MISMATCH on the broker, model_calls0/retry_count0/authority=false and IPC/GUI allocation not started; preregistration was not rewritten. Preserve both STOP records and #2705/#2730 history. The historical audit points to #2807/#2813 for successors; their current outcome is separate, not inferred here. Neither closure nor smoke success proves model efficacy or six-task completion.
• #2540 — Invalidate speculative continuation on obstacle, contradiction, stale/unavailable evidence or target replacement before authority. Closed; PASS_CONSTRUCTION_ONLY. Pinned Python-container fixture covered eight cases × four modes = 32 rows (28 contradiction/obstacle rows), independent oracle32/32, authority grants0/input executions0. Initial remote sparse-checkout stall was cancelled; later integration comments report successful retained-ledger gates. Both #2544 and successor #2567 are currently merged, despite #2567's description calling #2544 a draft; retain that source-description discrepancy instead of treating #2544 as currently unmerged. Integration does not change construction-only scope: no measured model/GUI/task benefit. An adjacent Obstac bridge comment records STOP_TEST_DISCOVERY_IMPORT followed by a corrected single rerun with9/9 tests; it explicitly is not this temporal experiment's completion, and its comment-only publication boundary is distinct from the merged ledger.
• #2459 / merged #2461 — Check X11 shift-adaptation reproducibility across training seeds. Closed; bounded reproducibility result with seed-sensitive utility. The same source-bound split and 354-parameter CNN, confidence threshold0.75, seeds2401–2405 produced accuracies91.25/90.00/91.25/80.00/83.75%, mean87.25%, minimum80%; accepts15/17/14/9/15 and observed aggregate false positives0. A previous single88.75% run is not promoted as representative. These dependent same-split runs do not establish zero population risk, independent-family transfer, semantic correctness, runtime/latency/token benefit or GUI authority.
• #2531 — Separate split uncertainty from training-seed uncertainty after #2459. Open; proposed, still unexecuted according to its retained audit comment. Preregistered idea is three split identities × five seeds (15 cells), unchanged architecture/gate, per-cell accepted true/false/YIELD and coverage, a conservative false-accept upper bound, all-YIELD comparator and corruption controls. Proposed decisions distinguish scoped stability PASS, HOLD_SPLIT_SENSITIVITY and FAIL_FALSE_ACCEPT; none is an observed result here. The parent five-seed observation remains unchanged, and the independent-family question in #2460 remains separate.

Historical ideas — reconciled 2026-10-02 19:42 UTC

• #2388 — Source-bound local CNN gate under X11 distribution shift. Closed; HOLD, not a usefulness PASS. A 354-parameter CNN trained on160 base frames was tested on160 shifted frames (target size/position/clutter/occlusion). Accuracy98.75%, observed false positives0, but accepted0/YIELD160/rejected0 at the fixed confidence gate. The retained audit explicitly separates preserved fixture safety from no suppression value demonstrated. Neither high classification accuracy nor closure establishes runtime authority, arbitrary-GUI transfer or latency/token savings; calibration/coverage needs separate evidence.
• #2419 — Put a source-bound OOD gate before the local classifier. Closed; bounded composition result. With the fixed99th-percentile pixel-distance threshold, all160 fresh-family rows went to OOD YIELD; OOD pass0, classifier calls0, composed accept0, final YIELD160. This shows this fixture's route avoids invoking the classifier on those unknowns, not semantic/application usefulness or production integration. Both stages remain non-authoritative.
• #2429 — Sweep OOD quantiles to find a utility/safety tradeoff. Closed; calibration STOP due to ties. The95th/99th/99.5th percentiles all yielded0.07607439 and identical routing: base accepted160/160, old-shift143/160, fresh0/160. No discriminative calibration value, threshold change or promotion resulted. Preserve the STOP and prior OOD evidence; its audit points to #2519 for order-statistic intervals, tie policies and leave-one-reference-out stability, without asserting that successor's outcome here.
• #2413 / merged #2415 — Audit frame/receipt integrity of the X11 OOD evidence. Closed; published integrity-only result with an execution-evidence caveat. Issue and PR bodies report normal-manifest PASS, one-byte frame tamper rejected and receipt-SHA tamper rejected. A later Issue audit instead flags missing retained execution/mutation evidence and points to #3198; that caveat is not erased by the merged PR or summary claim. Do not infer a full executable tamper audit from manifest construction/file existence, or promote integrity into semantic/runtime safety.
• #3198 — Execute the broader source-bound mutation audit after #2413. Open, but has a retained local result comment; not plan-only. The comment reports Python3.12-slim/Docker network-none, pristine bundle passed and eight directed mutations detected before authoritative output (frame bytes/order, config, threshold, source commit, receipt lineage, summary count, same-name path content), independent audit pristine=true/mutations8/all_detected=true. Reported decision PASS_X11_OOD_MANIFEST_TAMPER_AUDIT_SCOPED; raw SHA-256 cb6bd4fd7b29100852fd673d76926dc3c62f0f12eeda7a3b36155597240b47c5. Execution explicitly occurred without branch/PR operation, with a local report path only; this comment does not establish repository publication of the full raw bundle. Keep result, open status and publication gap distinct. No OOD calibration, GUI/application effect or production claim follows.

Historical ideas — reconciled 2026-10-02 19:45 UTC

• #1602 / merged #1613 — Make lossy multimodal projections read-only proposals, not canonical truth/authority. Closed; consumed formal integrity/accounting STOP, scientific disposition NONE. Despite the later Issue comment describing an execution gap, the delivered PR retains formal1/reruns0 and raw FAIL_SUBSTRATE_SEMANTICS. Its post-result checks identify14 omission false positives from value equality without before/after mutation, and an auditor expecting exactly20k stale rejections despite stale+ABA40k aggregation and a >=20k gate. Candidate/oracle mismatches0 and direct projection mutations0 do not repair the failed accounting/audit or convert the run to PASS. Raw metrics, source bundle, construction PASS and failed audit remain unchanged; source freeze is not runtime evidence.
• #3207 — Execute a projection roundtrip corpus after the design/accounting lineage. Open with a local result comment, not merely a plan. The comment reports120,000 rows/16 entities/10 classes, candidate/oracle mismatches0, fresh revalidation12,000, unsafe SUMMARY_REHYDRATE witnesses108,000 and PASS_MULTIMODAL_PROJECTION_ROUNDTRIP_EXECUTED_SCOPED; ledger SHA-25631db304c072bf8f5811fc1aafaa02959ce41dfff1feca2e004e72f56730d3001. Execution was explicitly without branch/PR operation. Qualification: comment seed0 differs from #1602's frozen159620260918001, and all-class presence/digest-changing mutation checks alone do not prove #1602's >=20k per-category gates or semantic corruption rejection. Do not relabel the consumed #1602 run, assume its exact preregistered corpus was completed, or infer published full raw evidence/model-compaction benefit from this comment.
• #2012 / merged #2019 — Retain bounded-skew contextual evidence without weakening strict authority-critical joins. Closed; PASS_BOUNDED_SKEW_CONTEXT_JOIN_SCOPED. Ten directed finite controls gave COHERENT3/SKEWED_CONTEXT1/STALE2/UNJOINABLE4; all grants_input_authority=false, skewed context never promoted to coherent/action authority, role-corruption control failed closed, formal1/audit1/reruns0. This is synthetic semantic evidence, not live synchronization, model usefulness or recovery. The later execution-gap comment points to #3204 for that distinct transfer, rather than erasing the delivered formal result.
• #3167 — Bind reusable receipts to session/resource/generation/source, then independently observe effects. Closed; several separate scoped results. Initial source audit retained HOLD_SESSION_RESOURCE_PROVENANCE_INCOMPLETE before execution. A separate comment-only temporary-file experiment reports40 rows/4 policies, bound admitted1/unsafe0 versus reduced-policy unsafe4/4/7; it is not the later GTK allocation. Merged #3191 tests real GTK/X11 receipt binding: same identity admits, restart/replacement/generation/source/cross-session/malformed/duplicate deny, PASS_SESSION_RESOURCE_RECEIPT_BRIDGE_SCOPED. A first cross-session construction mistake and corrected execution are disclosed, not silently one-shot pooled. Merged #3200 separately establishes valid-receipt effect transfer (four emissions/effect) and zero-emission denials. Merged #3206 separates transport from postcondition: useful and no-effect arms both have valid receipt/four emissions/completed transport, but only useful is SUCCESS; no-effect is ACCEPTED_NO_EFFECT. One GTK/X11 topology only; receipt validity and transport completion are not task success or universal durability.

Historical qualification for the already-indexed #3204 transfer lineage: preserve the earlier44-row no-model HOLD_MULTIMODAL_RECOVERY_UNMEASURED and distinct10-row join-only PASS; neither establishes recovery. Allocation01 stopped on unreachable host bridge. Merged #3823 preserves allocation02 STOP_MULTIMODAL_RUNNER after one HTTP400 request, zero completed responses, missing response-body limitation and no scientific audit. A later nine-response /api/generate corpus used the same allocation after that STOP; its author withdrew FAIL_MODEL_ANSWER as an admissible formal verdict and recorded STOP_ALLOCATION_COLLISION_CONTINUED_AFTER_STOP. Bytes/hashes remain supplemental post-STOP diagnostics, not a replacement formal result or model-quality estimate. This historical correction does not replace the separately indexed later #3827/#3858 successor evidence or resolve #3204's broader open recovery question.

Historical ideas — reconciled 2026-10-02 19:47 UTC

• #2844 / merged #2847 — Supply Python Xlib for a bounded GTK/X11 adapter allocation after #2836. Closed; dependency-only delivery. The additive image adds python3-xlib and its network-disabled preflight validates GTK/PyGObject/Xlib imports. This does not run input/effect cases or establish formal #2606 acceptance. The Issue audit points to #2851/#2872 for subsequent execution, but does not itself establish their outcome. Preserve #2836 and earlier evidence: successful imports/image preparation are not GUI effect, complete capture-time receipts or independent scoring.
• #2684 / merged #2690 — Build a case-controlled GTK fixture for the eight-case live adapter matrix. Closed; live preflight, not full formal acceptance. The delivered runner records scorer_matches=true with fixed-order SUCCESS/YIELD/YIELD/NONE/PARTIAL/YIELD/YIELD/YIELD; useful/no-effect/partial fixture modes and raw per-case event/effect/adapter artifacts are described. Ambiguous-delivery and terminal-cleanup-failure rows are controlled safety paths, not production-adapter evidence. Do not infer all #2606 receipt/independent-scoring gates from these expected dispositions or closure. The later audit references #2693/#2673/#2850 as distinct downstream execution; preserve #2671/#2672 historical failures/preflight.
• #2436 / merged #2438 — Compile fail-closed temporal-ambiguity recovery semantics. Closed; deterministic policy-table evidence only. DEFINITELY_BEFORE alone permits authoritative DONE; TEMPORALLY_AMBIGUOUS and GUARANTEED_AFTER map to non-authoritative QUERY_EFFECT. Retry requires idempotency plus matching generation; unknown clocks require RECONCILE_CLOCKS; wrong lineage/unrelated semantics ABORT. The standard-library/pinned-container contract does not measure actual clock domains, application consumption, model reasoning or task benefit. Its execution-gap audit points to #3193 without changing #2204/#984.
• #3193 — Transfer temporal recovery policy to executable postconditions. Open with retained result comment; HOLD_MODEL_RECOVERY_UNMEASURED. The comment reports network-none Python-container execution and a real temporary-file postcondition: ten temporal/lineage cases × four policies =40 rows, model_calls0; typed admitted2/unsafe0 versus TIME_ORDER_ONLY5/3, MODEL_FREEFORM_RECOVERY6/4 and DONE_ON_LAST_OBSERVATION9/7. Raw SHA-2568b9fd4408acd95aa1cc588698eb955949c9ed818bef81459abd8a8d7597cf6c8; no branch/PR operation. Reduced-policy labels, including MODEL_FREEFORM_RECOVERY, are not actual model execution. The comment describes typed admission for proven before/after, while #2438 permits authoritative DONE only for before; without raw per-row disposition evidence, do not equate this admission count with DONE or silently redefine the earlier table. Policy safety/temporary-file evidence, open status, model-recovery HOLD and full raw publication remain separate; no universal live-GUI or production safety claim.

Historical ideas — reconciled 2026-10-02 19:50 UTC

• #2344 / merged #2348 — Repair the desktop lifecycle source-identity audit after a reported mismatch. Closed; repair delivered, fresh formal PASS not established by closure. The Issue originally diagnosed main-source drift in run35446022418; the delivered PR instead identifies a false mismatch from reconstructing the Git blob header with literal backslash-zero instead of a NUL byte. It changes that separator only, retaining the four manifest identities and ten lifecycle rows. Preserve the stopped run and revised cause; do not silently describe this as a manifest rebind or infer the required INDEPENDENT_AUDIT_PASS from merged code alone. The later audit points to #3190 for a separate fresh freeze/audit; scope remains static provenance, not GUI/task safety.
• #2345 / merged #2351 — Obtain or record execution of the MAP01 entry-gate workflow after its definition merged. Closed; empty-run observation, readiness HOLD. The inspected head76100d0b31d02d5f2d5b1aedaff3257066462ec1 had workflow_runs=[], no run/job/artifact to score. The merged record preserves that absence without promoting PASS; it does not prove the expected32-vector container output or live matched recovery. The later Issue audit points to #3188 for execution, not an outcome inferred from workflow-definition presence. This records the historical inspected head, not a claim that all later CI remains absent.
• #2818 — Bind the Docker/shared-volume host broker to schema preflight. Closed; retain distinct contract, live preflight and infrastructure results. Initial merged adapter tests passed4/4, with no actual broker. The later BROKER_ENTRYPOINT_MISSING claim was explicitly corrected: the broker was present; that claim must not remain authoritative. A real host preflight returned a lifecycle turn.completed envelope instead of the schema payload: STOP_SCHEMA_RESPONSE_UNNORMALIZED, no retry/task allocation. Subsequent JSONL/parser/validator contract tests passed6/6 and3/3 separately. Merged #2832 records one corrected host-local preflight PASS_DOCKER_IPC_SCHEMA_BRIDGE_PREFLIGHT (exactly one item.completed message, compiled-form-grounding-v1, broker success, authority=false, GUI/task0). This is not the separate cancelled manual CI run35459803394 HOLD_INFRASTRUCTURE/CI_JOB_STALLED, nor the separate local Docker HOLD_DOCKER_IPC_HOST_MODEL_UNAVAILABLE. Merged #3026 later records contract CI run35465482539/job105956911378,5/5 tests and container cleanup, PASS_DOCKER_IPC_CONTRACT_CI_V2 without host model response. Contract green, live schema preflight, stalled CI and held endpoint remain separate; none establishes six-task/GUI effect, release readiness or model quality.
• #3190 — Freeze current-source lifecycle identities and re-audit ten rows. Open; formal audit invocations remain0 in the retained recovery records. Initial three-source/ten-row preflight retained HOLD_SOURCE_DRIFT_UNRESOLVED because the full source set/corruption gates were incomplete. The2026-09-22 four-source prospective freeze allowed only the API-blob rebind, with excluded2/2 construction tests and formal0. Later exact readback found only3/4 matches: frozen API47193afd… versus current a74962ce…; an earlier all-match claim came from a failed shell comparison and was withdrawn. Merged #5540 preserves all six preformal files and STOP_STALE_FROZEN_SOURCE_BEFORE_FORMAL. Separate merged #5571 preserves STOP_STALE_SOURCE_AND_LOCAL_TEST_FIXTURE (construction20 passed/1 sparse-import error, formal0). Its publication/index repairs do not change science. Later archive/ref recovery likewise reran no formal experiment. Preserve the old freeze and genuine later source drift; neither construction tests, archival integration nor the earlier NUL repair establishes a current-main formal PASS.

Historical ideas — reconciled 2026-10-02 19:53 UTC

• #2753 / merged #2754 — Make eight-case GTK results derive from immutable input/effect/release/cleanup receipts instead of hard-coded tuples. Closed for preregistration only. The delivered PR explicitly adds no execution result, Docker claim or #2606 acceptance. Required stale-observation/fresh-repair lineage and independently scored cleanup remain gates, not outcomes inferred from the plan. Its audit points to #2936/#3184 for later raw-bundle work while preserving #2690/#2709 preflight and review findings.
• #2916 — Compose fresh visual admission, source identity and separately correlated effect scoring around golden-v3 dispatch. Closed; do not infer integrated task success. Its audit points to distinct #2972/#2979/#2996 executions; merged rescue #3119 retains evidence, not a new run. Checked successor #2996 records live-observation provenance composition: matching session/target/sequence/binding accepts for ordinary admission, stale/authority contradiction refuse, authority=false/model0/input0/task-effect0. The explicit no-task-effect receipt is not an application score. Thus this closes a typed provenance bridge only, not visual-target discovery, actual input authorization, independent application success or six-task acceptance.
• #2050 / merged #2138 — Measure class/anchor-specific temporal query cost at the model boundary. Closed; HOLD_PRE_MODEL_COST_BOUNDARY. Delivered ten-case construction fixture reports canonical serialization totals universal5,359/class-only2,099/class+anchor1,709 bytes with unknown/shifted inputs retained. A queued container attempt was not used as evidence; local construction passed. Fewer serialized bytes do not establish tokens, query/capture/transport/provider/model costs, held-out correctness or end-to-end benefit. The audit points to #2090/#2068 for distinct model-facing work; symbolic break-even stays a planning bound, not runtime policy.
• #2105 — Freshly replicate scoped MAP01 physical occupancy without borrowing the prior execution identity. Closed; PASS_MAP01_V12_PHYSICAL_OCCUPANCY_R1_SCOPED. Retained run35443575048/job105898557254: import/preflight PASS, construction1 then formal1, three fresh sessions all PASS, six confirmed DOWN/UP edges, max censor width0.666226ms (<5ms), terminal empty ownership in all sessions; reruns/replacements/tuning/authority expansion/v3 promotion0. Artifact10585310389 SHA-256c5daa78298b6a1b642905de5ec75a44878b1d0e04e9089e284588604da49a4bc. #2114 and result-bearing #2127 are currently merged, though #2127 calls #2114 a draft historically. Pending duplicate run35443579255 was cancelled before execution, not a second replication. Physical-occupancy scope only: no gameplay/MAP01-clear, model, human-tempo, cross-platform or production claim; #1928 remains unchanged.
• #1358 / merged #1361 — Require fresh readiness generation plus separate ordinary authority at final admission. Closed; original source-first synthetic composition PASS. Formal1/reruns0 covered600,000 transitions/four scopes/eight states: stale stress25k and ABA25k each old-receipt effects0; fresh effects71,380, candidate/oracle state/result mismatches0; snapshot-only discriminator stale effects25k/25k, malformed8/8, corruption5/5, source rehash exact. Readiness itself grants no authority. Keep separate later comment-only experiments: missing-positive-control HOLD_FORMAL_STRESS_COVERAGE_INCOMPLETE; corrected schedule with shared state/authority modeling HOLD_INDEPENDENT_ORACLE_OR_AUTHORITY_MODEL_INCOMPLETE; then650k-record shared-fold HOLD_ORACLE_IMPLEMENTATION_NOT_INDEPENDENT. A final comment reports separately implemented folds and scoped PASS (stale0/fresh50k), but repeats the prior shared-fold result digest2ce902f… and supplies no branch/PR operation or independently inspectable auditor source here. Treat that as a reported follow-up, not proof that the earlier held audit passed or that the original71,380-effect corpus was rerun. All results remain synthetic semantics, without production ABI, real GUI/model/task benefit or authority implementation.

Historical ideas — reconciled 2026-10-02 19:54 UTC

• #2851 — Reach GTK/X11 dispatch after adding Xlib. Closed; completed but unscored, not formal acceptance. Two retained runs35460321582/35460315665 are separate allocation records. The first artifact audit reports exit0/empty stderr and COMPLETED_UNSCORED_GTK_XLIB_ALLOCATION: six independent cases ready, but ambiguous delivery remains UNKNOWN/no replay and terminal cleanup CLEANUP_FAILURE. Backend availability and completed execution do not establish all eight #2606 receipt/scoring gates; retain the original STOP/completed-unscored labels rather than pooling them as PASS.
• #2872 / merged #2876 — Repair evidence admission for ambiguous delivery and terminal cleanup failure. Closed; contract repair, fresh allocation separate. The delivered change requires explicit delivery-not-observable/no-replay evidence and cleanup failure reason plus release truth. Its comment cites an earlier eight-row PASS_GTK_RECEIPT_AUDIT_CONTRACT_SCOPED, explicitly schema/safety distinctions only. Neither that audit nor merged gate code establishes capture-time provenance, application effect or formal #2606 acceptance; #2851 stays unchanged.
• #2884 / merged #2885 — Fresh allocation after the independent gate contract merged. Closed; scoped receipt-emission/gate-repair evidence. Initial runs35461131317 and35461291363 failed with zero jobs/logs: dispatch infrastructure, no scientific allocation. Distinct run35461294079 completed unscored despite all eight ready flags. Run35461309046 is reported as PASS_GTK_INDEPENDENT_GATE_REPAIR_ALLOCATION in one comment, but a later comment reports its retained decision COMPLETED_GATE_REPAIR_ALLOCATION; preserve this label distinction, not a broad PASS. Separate run35461310655 also retains COMPLETED_GATE_REPAIR_ALLOCATION, fixed expected dispositions, scorer agreement and eight ready rows. Ambiguous/no-replay and cleanup-failure/release evidence became explicit, but capture-time provenance and application-bound causality remain unknown. Multiple artifacts are not one pooled one-shot result; none automatically satisfies formal acceptance.
• #2895 / merged #2897 — Evaluate immutable GTK evidence without rerunning the allocation. Closed; fail-closed verifier delivered, not formal acceptance granted. The verifier refuses when raw event traces are only path strings rather than retained event/input/effect/release/cleanup bytes. This preserves #2884 behavior and evidence while isolating raw-bundle completeness. Closure/merge does not establish PASS_FORMAL_2606_ACCEPTANCE; subsequent raw-file retention and later acceptance audits are separate evidence, including the already indexed #2892 summary-retention boundary.

Historical ideas — reconciled 2026-10-02 19:57 UTC

• #2896 / merged #2969 — Find a live caller-visible residual beyond #934's frozen rule VM. Closed; conditional Chromium-scoped PASS, LibreOffice HOLD. Fresh isolated Chromium data-navigation outside the frozen scale/OpenTTD clauses retained request, before/after epoch, same PID/window binding and independently observed title effect: PASS_LIVE_RESIDUAL_BOUNDARY_SCOPED. Separate stale control detected PID9→137 even with reused XID and submitted no old request; ambiguous two-candidate control likewise yielded without submission. First scorer allocation retained HOLD for the platform title suffix rather than being pooled with the repaired allocation. LibreOffice remains HOLD_EFFECT_NOT_INDEPENDENTLY_OBSERVED, with no save used to manufacture evidence. Model/provider/network/authority0. This is finite live residual/binding evidence, not demonstrated learned-backend need, model benefit, second-application success, product acceptance or broad generality; #934 replay remains distinct.
• #2558 / merged #2962 — Connect local Docker observations, a host model boundary and guarded native effects into the golden route. Closed after the scoped experimental record, not roadmap completion. Early native/model-free slice and optional non-authoritative GPU smoke did not establish model-in-loop success. Subsequent replay-only plans, one-task successes and six separate fresh fixture allocations likewise did not satisfy the preregistered six-task session route. Preserve NO_INTERACTIVE_DISPLAY, source/dependency-fetch and broker/timeout STOPs, schema rejections and STOP_COMPLETE_PLAN_NOT_ADMITTED (submit coordinates outside the image while a harness used Ctrl+S). A separately declared keyboard contract addressed that gap; its effects are not retroactive complete-coordinate-plan success. Final integrated-efficiency-live-01 at9c9d7cf…, seed991028, gpt-5.6-luna low, retained disposition RETAIN: plain/ephemeral/persistent arms each6/6 exact submissions, cumulative input tokens131,517/136,473/56,403, observed break-even task2; all releases verified, old-target pointer admissions0, persistent task4 stale-handle refusal plus repair. The delivered broker records actual codex exec usage and supplied schema/instructions. These are this allocation's retained counts, not price/latency savings, population reliability or automatic proof of persistent model-thread reuse; earlier two-call ephemeral and transport-only studies remain separately scoped. Intermediate failures are not success rows or pooled into the final denominator. The Issue's own update marks its original execution gap satisfied for this allocation; broad integration/production/generalization remain unproven.

Historical ideas — reconciled 2026-10-02 19:59 UTC

• #2836 / merged #2837 — Try the bounded GTK matrix after fixture startup/PyGObject repair. Closed; two retained STOP_GTK_ALLOCATION records. Runs35459760232 and35459765526 uploaded separate artifacts10588847843/10589612719; their existence and companion green checks do not make the allocation a PASS. The delivery explicitly closes this bounded recording unit only and preserves existing #2690/#2709/#2754 receipt-integrity defects. No formal #2606 acceptance follows from wrapper completion, startup repair or uploaded stdout/stderr. The already indexed #2844 Xlib dependency and #2851 dispatch successors are distinct later evidence, not a relabel of these STOPs.
• #2833 / merged #2840 — Bound and diagnose host broker response waiting after #2818. Closed; timeout behavior recorded, not a schema-valid response success. Configurable subprocess timeout now records TimeoutExpired, request ID, timeout, stderr and start/exit timestamps with authority=false. The local no-GUI five-second diagnostic terminated as HOST_BROKER_SUBPROCESS_TIMEOUT; broker contract tests3/3 and bridge tests6/6 passed. Those tests establish bounded failure handling, not PASS_HOST_BROKER_RESPONSE or a resolved CLI/provider cause. Preserve PATH/response-missing historical stops and the separate later #2818 preflight evidence; no GUI/input/task, model-quality, latency improvement or six-task claim.

Historical ideas — reconciled 2026-10-02 20:01 UTC

• #2796 — Freeze transitive GTK/runtime imports instead of only direct imports. Closed; source-closure and preacceptance results remain distinct. Merged #2801 adds runtime.backends coverage and an Xlib-backed image/import gate. A subsequent eight-case Docker run reports PASS_GTK_MATRIX_PREACCEPTANCE_SCOPED, expected dispositions/scorer order matched, useful/partial effects, stale emissions0/no replay and terminal cleanup failure as non-success. This is explicitly below formal #2606 acceptance. A later current-main manifest mismatch stopped before another matrix; no old manifest was rewritten. Separate merged #2835 freezes a new15-file/100,887-byte closure and reproduces its source audit read-only. That provenance PASS is not a rerun or upgrade of the earlier GTK effects, and #2780 failures stay unchanged.
• #2803 / merged #2808 — Independently verify retained GTK matrix evidence without rerunning it. Closed; PASS_ACCEPTANCE_BRIDGE_SCOPED. Eight retained raw rows and the verifier check order/dispositions, authority0/release, useful/partial effects, stale/ambiguous no replay and cleanup failure non-success; rerun=false. A subsequent retained-data check reports the same scoped result. This is an audit of that bundle, not a new live allocation, broad #2606/product promotion or evidence that another lineage's missing raw files are complete. Model utility, human tempo and cross-platform generality remain unmeasured.
• #2804 / merged #2812 — Explain why the formal runner timed out while a manual GTK launch succeeded. Closed; readiness fix revealed a different dependency HOLD. The runner used /usr/local/bin/python3, which could not import Debian gi under its path; the fixture exited before metadata. Selecting /usr/bin/python3 allowed one bounded allocation to reach all eight cases, but adapter dispatch remained backend_unavailable/No module named Xlib: HOLD_X11_XLIB_DEPENDENCY_UNAVAILABLE. Source-bundle readiness and case execution are not observed application effects or formal acceptance. Preserve the original case1 timeout; later Xlib successors are separate.
• #2805 — Capture Xvfb/fixture startup evidence before any GTK input. Closed; historical diagnostic STOPs, then narrowly scoped startup PASS. Initial /out already-existing FileExistsError stopped the diagnostic itself; separate artifacts with only exit1 or no files did not establish readiness. Output-directory repair still left HOLD_INFRASTRUCTURE_DIAGNOSTIC_OUTPUT_EMPTY. Merged #2827 preserves Debian dist-packages in PYTHONPATH; subsequent run35459593665/artifact10589178021 reports PASS_GTK_FIXTURE_STARTUP_DIAGNOSTIC (socket ready, metadata present, fixture alive,304.608ms, retained stderr and bounded cleanup exit−15). This is no-input startup/evidence retention only. A separately completed existing matrix remained unscored; do not pool it with the diagnostic or claim GTK task effects/model/general correctness. Every output-directory/empty-artifact STOP remains historical evidence.

Latest reconciliation — 2026-10-02 20:07 UTC: attribution HOLD, corrected coverage-gate contrast, and wait-cost contract

• First-failed-boundary attribution: #5694 A03 in PR #6799 retains nine authored rows, zero audit errors and five rejected corruptions (PASS_METHOD_SCOPED; candidate/auditor once each, no retries). Its scientific disposition is HOLD_PHASE_CONTRAST_NOT_MATCHED: phase-hit used capture 10 ms/horizon 50 ms, while phase-miss used captures 10/50 ms/horizon 60 ms. The descriptive boundary difference therefore does not isolate cue phase. PR was open/unmerged at this check despite the historical “[Draft]” title (GitHub draft flag false). Successor #6803 prospectively holds both arms at captures [10,50] ms/horizon 60 ms; no result was present in the fetched Issue/comments. Preserve A03 HOLD and earlier allocations separately; no live-control or causal-prevalence claim.
• Coverage-gate refinement — original comparator confound and separate successor: #6645 T1 PR #6800 is merged, retaining five synthetic rows and zero raw-only audit errors (PASS_COVERAGE_GATE_SCOPED). An append-only author erratum withdraws its isolated causal interpretation: the candidate also checked hidden modal occlusion later, whereas legacy comparator omitted that check, so the same candidate without the gate would already REFUSE the harmful row. Raw legacy ADMIT/candidate UNKNOWN is preserved, not regraded. Separately frozen T1b PR #6804, merged, varies only the gate boolean in the same observation-limited function, with no oracle mount or hidden-modal value available to the candidate. Hidden harmful state: ungated ADMIT, gated UNKNOWN; observationally identical safe state: gated UNKNOWN too. Complete controls retain ADMIT/REFUSE; independent audit zero errors, five authored rows, candidate/auditor once each in separate pinned network-disabled OrbStack containers, zero retries: PASS_COVERAGE_GATE_ISOLATED_SCOPED. This demonstrates conservative abstention in the fixture, not hidden-state discrimination, unknown-unknown discovery or live safety/performance. Both allocations assume a complete external family registry; memory configuration is not enforcement proof. Author records local checks passing but GitHub Actions queued at merge; integration is not remote-CI success. Parent remains open.
• Wait-cost monotonicity: #5428, PR #6802, open/unmerged, adds a fresh T0 without replacing historical T1 failure/T2 fixed-trace threshold repair. Rule after hard gates: V_now = commit_value - lost_flexibility_cost; V_wait = future_gross_value - wait_cost; COMMIT on ties, unpriceable cases HOLD. Independent reconstruction covers 235/235 rows (225 numeric + 10 corners), zero audit errors and zero wait-cost/lost-flexibility monotonicity violations — PASS_METHOD_SCOPED. Windows standard-library CPU finite-contract check only, not calibrated utility, optimal timing, live benefit, or container/GPU evidence. Pre-freeze construction failures remain retained.
• Integration/check-status correction only: merged PR #6801 records that #6680's PR #6724 merged at 19:43:30 UTC while six GitHub checks were still QUEUED. Passing local checks are separate evidence; do not infer remote checks green/completed from merge. Frozen experimental bytes/results were not rerun or changed.

Historical reconciliation — 2026-10-02 20:09 UTC: mixed-app discovery and GTK receipt/fault prerequisites

• Stable typed identities, not mixed-app acceptance: closed #2714 succeeded #2687's identity-discovery HOLD. Invalid Inkscape arguments/readiness repair in merged #2762 yielded only Inkscape/Chromium: HOLD_IDENTITY_DISCOVERY. A distinct local v3 removing LibreOffice window-suppressing flags in merged #2773 obtained a scoped three-app PASS; do not confuse it with the separate CI v3 backend-only #2784, closed/unmerged, which remained HOLD (LibreOffice process alive, no visible identity). Merged v4 #2795 retained both run 35458826748's stable three-app PASS and run 35458837488's temporal-startup HOLD/stable=false; neither replaces the other. Merged v5 #2809 waits for all expected app families and deterministically selects primary records, excluding raw-window-count/Tip-window instability. Run 35459125406: three distinct stable window/PID pairs across two observations, PASS_IDENTITY_DISCOVERY_SCOPED (author's normalized-scoped summary). Input/model/network operations all zero. Closure establishes this readiness seam only, not task effects, recovery, eight-case acceptance, long-session correctness or performance; earlier HOLDs remain separate.
• Real target disappearance: closed #2693 externally destroys a fresh GTK fixture immediately before dispatch in local Docker/Xvfb. X BadWindow yields BACKEND_EXECUTION_FAILED/execution_failed, failed op 0 (focus), no completed ops, program incomplete, partial-input effect UNKNOWN, verified empty release; success disallowed. Real adapter fail-closed observation, not the preregistered terminal-cleanup acceptance case and not full #2606 PASS. Source describes locally retained output; this index does not independently certify raw publication.
• Artificial transport ambiguity after real dispatch: closed #2701 preserves native execution of four ops and verified empty release, then explicitly injects delivery=ambiguous before adaptation. Bounded fault preflight PASS: refused, program_completed/task_success/replay_allowed/authority_granted all false; effect file absent (program did not use the application receipt). Open #2697 retains STOP_AMBIGUOUS_BACKEND_EMISSION_UNAVAILABLE: production GTK/X11 emits no natural ambiguous delivery field. #2701 tests injected boundary behavior, not natural backend ambiguity, application effect or full #2606 acceptance.
• Capture-time receipt schema rather than post-hoc labels: open #2706 retains HOLD_LIVE_RECEIPT_EMISSION_INCOMPLETE; its four contract tests and reported merged #2710 are structural prerequisites, not receipt provenance. Closed successor #2708 reports additive live GTK/X11 runner emission of session/window/revision/input/effect/cleanup fields and direct matrix gate invocation: fixed order/scorer agreement, six classified outcomes ready, ambiguity UNKNOWN/no replay, terminal cleanup CLEANUP_FAILURE, zero authority grants. This is receipt-emission preflight only: ambiguity and terminal-cleanup remain controlled fault/safety paths, not naturally emitted production cases. Parent open state and original HOLD are not silently rewritten; no full eight-case acceptance follows.

Historical reconciliation — 2026-10-02 20:11 UTC: fresh GTK observations/effect fixtures and advisory handoff

• Observation identity without effect proof: closed #2660 replaces Xvfb-root capture with each fresh GTK window's own X drawable. Author-reported local Docker/GTK3/Xvfb preflight: 8/8 cases, eight distinct SHA-256 observations, dispositions SUCCESS/YIELD/YIELD/NONE/PARTIAL/YIELD/YIELD/YIELD. Root-capture negative control retained only three distinct hashes out of eight. No model; replay false and safety cases grant no authority. This separates authored fixture observations, not independent effects, production integration or formal #2606 acceptance.
• Fresh post-effect window, with failed rendering attempts preserved: open #2671 retains STOP: deprecated background mutation produced no useful/partial drawable change, and draw-callback retry failed with missing cairo.Context foreign converter; scorer_matches=false. Closed successor #2672 instead creates a fresh post-effect GTK window, avoiding that bridge. Author reports local Docker eight-case scorer agreement with dispositions SUCCESS/YIELD/YIELD/NONE/PARTIAL/YIELD/YIELD/YIELD and before/after hashes, event/input/authority/cleanup/scorer receipts in a local retry artifact. Bounded fixture effect-receipt preflight only: replacing a window is not proof of production app effect detection, model utility or full eight-case acceptance. This index records the linked author report, not an independent readback of its locally held raw file. #2671's failed attempts and open state are unchanged.
• Advisory handoff capsule, never live authority transfer: closed #2630 proposes a bounded capsule requiring matching task/session, uncertainty, replay prohibition, expiry and fresh authority acquisition; malformed/stale/contradictory capsules UNKNOWN. Current GitHub state confirms #2633 merged, correcting the earlier intermediate comments that still described it as held/unmerged. Its disclosed local pinned network-disabled Docker contract reproduction was 3/3, not a live handoff experiment. Parallel #2634, closed/unmerged, is reported in Issue comments as eight-case formal container PASS_HANDOFF_CAPSULE_CONTRACT_SCOPED (run 35455330928); #2635, closed/unmerged duplicate, retained local 8/8 but checkout cancellation before source gate. Fresh-main v3 #2650, closed/unmerged, retains STOP_RUNNER_CHECKOUT_BEFORE_MATRIX (run 35455691931: initialization succeeded, checkout cancelled, unittest skipped, cleanup succeeded); prior full-archive no-space stop was setup failure, not contract failure. Do not pool these matrices or infer v3 formal PASS from another branch. Integrated result is pure validator scope only; planner/model/GUI/input, fresh acquisition in a real handoff, two-domain recovery, latency and token benefit remain unverified.

Historical reconciliation — 2026-10-02 20:14 UTC: O3/O4 gate contracts and GTK scorer layers

• Relevant-region suppression contract: closed #2612, merged #2619, proposes fallback unless observation/intent/region/coverage/freshness/effect binding are current and authority-free. Local Docker three-test PASS and later author-reported run 35454982132 three-test PASS_O3_RELEVANT_REGION_FAIL_OPEN_CONTRACT_SCOPED are distinct from the earlier retained checkout HOLD/STOP comments; do not erase that chronology. This is a pure evaluator, not pixel capture/region selection or live multi-app benefit. Review flags missing authority/ambiguity keys defaulting to safe assertions and bool/float epochs matching integers. Read-only current main gate (blob 50c27bd4e8651da74163c786daeb654928b08958) still uses those defaults and equality without exact integer validation, although a later optional trusted-source-window check exists. Thus the finite PASS is not proof of the complete malformed-record contract; no new experiment or failure regrade was performed here.
• VERIFY before model escalation: closed #2595, merged #2598, requires current-bound TRUE/FALSE only to suppress; UNKNOWN/stale/mismatch/invalid/authority-bearing receipts escalate. Local network-disabled pinned Docker 3/3 and later reported run 35454651426 3/3 (historical label PASS_O4_RUNTIME_VERIFY_FAIL_CLOSED_CONTRACT_SCOPED) are contract evidence only; the behavior falls back to escalation, not execution authority. Earlier queued/checkout HOLD remains history. Review points out non-mapping payload exceptions and bool/float epoch equality. Current main gate (blob dba9fbd1d0af152a039d1bb924b5c58d9b3a1df7) still calls receipt.get without a mapping guard and compares epochs without exact-type validation. These source-observed coverage limitations do not overwrite the three-test PASS; broad malformed-input handling, live verifier quality, latency/token savings and runtime integration remain unproved.
• GTK adapter contract versus live/synthetic coverage: closed #2557, merged #2563, retains PASS_GTK_ADAPTER_CONTRACT_CONTAINER_SCOPED (local 3/3 and reported dedicated run 35453895958 successful after checkout HOLD). Separate no-network local GTK/Xvfb click/text/Ctrl-S bridge reports USEFUL/ready with authority_grants=0: one live fixture case only. Another eight-case matrix classified USEFUL/UNAVAILABLE/REFUSED/NO_EFFECT/PARTIAL/REPAIRED/UNKNOWN/CLEANUP_FAILURE with ambiguity/cleanup not ready, but is PASS_SYNTHETIC_SECOND_DOMAIN_MATRIX_SCOPED, not live eight-case acceptance. Initial python3-gi runtime mismatch remains a construction FAIL, separate from a system-Python four-case preflight. Merged additive #2581 addresses post-merge P1 findings using explicit authority/cleanup/effect/ledger/repair evidence and seven focused adversarial tests; no live transfer or #2492 acceptance follows.
• Adapter-independent scoring: closed #2586 aims to reconstruct outcomes from receipts without calling/importing the adapter. Local pinned Docker 3/3 PASS is a contract result; separate actual GTK fixture click/text/Ctrl-S receipt scores USEFUL/ready/authority_grants=0 in the independent scorer — PASS_GTK_LIVE_TO_INDEPENDENT_SCORER_ONE_CASE_SCOPED, not all eight live cases. Earlier checkout HOLD remains retained. Current GitHub API says #2590 merged=true, while its body calls the simpler scorer superseded by merged #2581 and reports dedicated run 35454445535 success. Preserve both the actual integration state and that supersession note, not the literal body wording “closed” as evidence of non-merge. None establishes provider/model utility, authority admission or full second-domain acceptance.

Historical reconciliation — 2026-10-02 20:17 UTC: MotorState seams and calibration support readiness

• Public pure MotorState, not a backend: closed #2508 proposes packaging an already verified research transformation under runtime without side effects. #2512 is closed/unmerged despite “Closes” wording; an independent local checkout stopped with disk exhaustion before testing. Current main public package exists and exports SCHEMA/validate (blob a0e7179cbca9f4948ac222d48b5976ce724aab77), but this does not make #2512 merged. Later #2541 comments retain cancelled dedicated tests/checkout STOP and only platform CLI success. #2579, now closed/unmerged/superseded, retained green checks yet had a reviewed pointer-validity gap: key presence admitted null/string/bool as certainty. Merged #2592 separately requires finite numeric x/y and excludes bool (reported local 10/10); merged #2608 handles huge exact integers without math.isfinite overflow (11/11). Contract repairs only: no native emitter/backend, authority or task-effect evidence; green checks on predecessor are not full semantic correctness.
• Explicit-context dispatch-result bridge: closed #2530, merged #2536, maps synthetic accepted/rejected/failed/released dispatch results only with caller-supplied observation/surface/frame context; release uncertainty retained and accepted dispatch not task success. #2535 invalid image-digest-length failure happened during container initialization before tests; #2536/#2543 cancelled checkout runs retain STOP_RUNNER_STALL_BEFORE_BRIDGE_TEST. Later author reports new #2536 dedicated tests/platform CLI/repository gate PASS and integration — not a replacement of those cancelled runs. Merged #2550 adds only missing workflow trigger; separately reported current-main bridge checks passed. Pure synthetic contract/CI coverage, not proof that live native sessions supply the context or that CLI composition improves tasks.
• Read-only receipt validation versus distribution readiness: closed #2521, merged #2524, exposes optional motor_state_validation in non-raw receipt views, absent/malformed/rejected states explicit; raw/dispatch unchanged, no authority/freshness/effect inference. Initial three-platform ModuleNotFoundError was an integration dependency/sparse-checkout defect, not semantic failure; after including the public package, CLI and X11 checks passed. Build/gate were pending at the recorded merge checkpoint; subsequent portable zipapp builds failed on Ubuntu/Windows/macOS (doctor nonzero; empty stdout/JSONDecodeError in malformed-target test, archive-closure passed). Preserve HOLD_RECEIPT_SEAM_DISTRIBUTION_BUILD alongside actual merged state; this is not clean all-platform distribution readiness.
• Guard calibration needs real negative support: closed #2509 identifies #1793's zero-count stale/fresh-refusal and failure/recovery strata as a support gap, not proof its expectation estimate is wrong. It proposes a separately allocated same-route block with negative/censored support, independent labels/recovery, disjoint calibration/heldout population and uncertainty; no historical rows pooled. Merged #2517 closes construction/preflight only: dependency-free synthetic ledger readiness rejects duplicate IDs, mixed clocks/populations, historical rows and missing strata. Scientific disposition remains HOLD_SYNTHETIC_READINESS_ONLY; real episodes, costs, recovery and break-even stability unmeasured. Earlier workspace-index failure is STOP_INTEGRATION_INDEX_STALE, not calibration FAIL. Maintenance #2523, closed/unmerged, failed public relative links due untracked target; also bookkeeping, not a scientific result. Closure/merge does not satisfy PASS_GUARD_CALIBRATION_SUPPORT_CLOSED.

Historical reconciliation — 2026-10-02 20:20 UTC: semantic differences, observation-only recovery, and pure state mapping

• Semantic differences need identity/epoch/source binding: closed #2004, successor to #1938, tests typed movement/text/enabled/focus/appearance facts without action authority. Merged #2011 reports an initial four-pair finite PASS_SEMANTIC_DIFFERENCE_EVIDENCE_SCOPED; its unrelated concurrent-directory index failure is integration evidence, not scientific validation. A distinct formal run 35443718133 retains FAIL_ORACLE_MISMATCH: 12 cases plus 54 exhaustive cases, 65 matches/one mismatch; ambiguous-frame control expected UNKNOWN/unbound_or_stale but emitted KNOWN/moved. Independent audit did not run after the nonzero formal exit; authority/GUI/model/network zero. Merged separate successor #2507 reports ten-case PASS_SEMANTIC_DIFFERENCE_EVIDENCE_SCOPED, zero authority; parent comment records analysis-index/gate still IN_PROGRESS at the merge checkpoint. Do not overwrite, pool or reinterpret the 66-case FAIL from the later ten-case claim. Model usability, real observation reliability, GUI benefit, tokens/latency and runtime promotion remain unproved.
• Observation-only recovery without authority: closed #2452, merged #2455, models bounded same-session/current recovery observations, rejecting input operations, lease transfer/extension and effect/task-success claims. Issue retains checkout-cancellation STOPs and a separately reported four-test classification/schema failure (valid schema omitted from allowed fields; expected rejection reasons unreachable); some comments give differing descriptions of run 35451287547, so this index does not resolve that run's chronology from comments alone. A local mirrored 3/3 check was explicitly not exact-head evidence. Later exact-head run 35451709630 reports finite non-authority PASS; run 35451963400 reportedly fetched exact source by SHA and passed five tests/compile/cleanup: PASS_OBSERVATION_RECOVERY_CONTRACT_SCOPED. Separate merged successor #2488 fixes forbidden-field classification ordering and reports independent ten-case PASS_OBSERVATION_ONLY_CONTRACT_SCOPED plus 5/5 tests, zero authority. Its analysis-index failure remains integration-quality evidence, not a clean repository-wide gate. All PASS claims are finite contracts: no live observer/GUI/model/transport recovery, input authority, or planner-boundary benefit.
• Native records to MotorState, preserving uncertainty: closed #2490, merged #2496, maps synthetic native-shaped observation/acknowledgement/binding/held-input/release fields without inventing focus or effect confirmation. Missing observation/focus stays uncertain; failed release stays an event. Parent reports native-motor-state-adapter check SUCCESS but detailed logs were rate-limited at that checkpoint. Scoped record mapping only, not live backend completeness or task-effect proof.
• Reusable pure mapping seam: closed #2502, merged #2505, packages the preceding mapping with tests of malformed/authority-bearing input, uncertainty/release preservation, immutability and no backend imports. Reported test check PASS supports the pure module; no lease mutation, dispatch, GUI/model or application-effect result. Distinguish this research-to-module step from later public packaging (#2508), explicit-context bridge (#2530) and certainty-type successor repairs already indexed above; those later results do not make the original synthetic map a live backend test.

Historical reconciliation — 2026-10-02 20:22 UTC: direct-engine control HOLDs and MotorState contract

• Unblocked rotation control did not establish delivery: closed #2468, merged #2470, tests whether the earlier forward HOLD was merely blocked geometry. basic.wad/MAP01, TURN_LEFT only for four tics: angle 0→0, reward zero, identical pre/post frame hash, cleanup complete — HOLD_NO_MEASURABLE_TURN (reported formal run 35451534111 and separate local Docker reproduction). Earlier STOP_SOURCE_IDENTITY_UNVERIFIED recorded workflow/head source mismatch; do not erase it or infer source verification from merge. Review warns frame changes could be autonomous rather than rotation and tee could mask Python failure; infrastructure success is not action-effect PASS. Both forward and rotation unchanged leave blocked-forward-only explanation insufficient, not proof that all direct action delivery fails. No X11/controller/task benefit.
• Known-good fixture reproduction still HOLD; history hypothesis is separate: closed #2472, merged #2475, repeats freedoom2/MAP01/PLAYER/RGB24 with four-button TURN_LEFT [0,0,1,0] for four tics. Reported run 35451688220: zero reward, identical frame, HOLD_FIXTURE_NO_FRAME_CHANGE. Separate local #2482, closed/unmerged, additionally records angle 0→0: HOLD_NO_MEASURABLE_TURN; its closure did not integrate it. Another Issue comment concerns a different 20-seed/320-sample action-effect-history reproduction: FAIL_ACTION_EFFECT_HISTORY_HYPOTHESIS, 128 test samples, current accuracy .6015625 versus delta-history .5859375; 17 alias pairs but history accuracy .5882352941 below .95 and one exact cross-label history collision versus required zero. Keep this distinct from the one-action HOLDs and from prior ATTACK observations; no model/OS input or general product claim.
• ATTACK ammo control did not reproduce historical change: closed #2477, merged #2478, tests one direct-engine ATTACK with explicit AMMO1. Retained formal/local outcome HOLD_NO_ATTACK_EFFECT: AMMO1 0→0, reward zero, identical frame, cleanup complete. Local reproduction reports seven declared buttons and vector [0,0,0,0,0,0,1], while proposal described a four-vector setup: retain those different descriptions, not a matched identical-setup replication claim. Source-only checkout independently returned HOLD_MISSING_RUNTIME before execution, not another attack result. No X11 transport/controller/end-to-end evidence; this does not negate a separately retained historical AMMO2 response.
• Commanded/observed state is not authority or application effect: closed #2481, merged #2484, defines MotorState uncertainty and retained critical release events. Author reports native and Docker seven-case PASS_MOTOR_STATE_CONTRACT_SCOPED: confirmed/explicit OS uncertainty/verified release accepted; missing transition, authority promotion, implicit confirmation and missing surface rejected; model/GUI/input/lease/network zero. Review nevertheless identifies untested contradictory release event/status and empty acknowledgement identifier acceptance. Record these as original review limitations, not automatically repaired by merge or by the seven-case PASS. Only a pure JSON contract prerequisite; later mappings/public seam results remain separately indexed, with no live backend or task-effect promotion.

Historical reconciliation — 2026-10-02 20:24 UTC: occupancy frontier, partial delivery, and direct-action calibration

• Occupancy interval-refinement frontier, not realized telemetry: closed #1592 asks how much inward information is needed to meet retained #443's width/upper gate. Preformal correction notes independently rounded W=2151.534 ms versus U-L=2151.533 ms; do not conflate them. Gate-native quantized formula is bdW-adU >= bW-aU, with q=a/b=1/4. It states integer-microsecond requirements 38.351 ms width-only and 51.135 ms upper-only. Initial harness/path STOP ran no scientific rows; a separate claimant yielded at formal0 on collision. Read-only review flagged the zero-upper convention, so six-boundary/random checks alone do not prove complete edge-domain equivalence. Merged #2251 reports PASS_OCCUPANCY_GATE_FRONTIER_SCOPED, 200006 rows, zero direct/frontier mismatches, under immutable-source Python container runs. Distinct merged Decimal analysis #2469 uses L/U-derived width and reports 200000 refinements/zero mismatches, 38.34975 ms deficit and 51.133 ms upper-only requirement. Preserve these as different input/rounding analyses, not a numerical replication or replacement of the gate-native correction. Index staleness STOP, a recorded repair verification, and subsequent new index drift are bookkeeping history, not frontier scientific FAIL. Analytical planning scope only: no live occupancy narrowing, MAP01 efficacy, useful feedback, recovery authorization or threshold change.
• Partial delivery is not success despite task/program flags: closed #2457, merged #2465, reproduced confirmed_partial + nested task_success=true incorrectly mapping to success, then minimally preserves typed partial and raw partial_effects. Native/local Docker regression PASS_CONFIRMED_PARTIAL_NON_SUCCESS (one partial case/one confirmed positive) and eight adapter-contract tests in each environment. program_completed/task scoring/delivery completeness remain distinct; unknown/ambiguous fail-closed behavior preserved. This is semantic regression evidence, no actual task effect, GUI/model benefit, latency/token or authority result.
• Forward direct-engine calibration retained no change: closed #2458, merged #2463, follows observation readiness with one fresh basic.wad/MAP01 declared movement action for four tics. Reported run 35451412153: reward zero, identical position variables and frame hashes, cleanup complete — HOLD_NO_MEASURABLE_STATE_CHANGE, not PASS_DIRECT_ACTION_EFFECT_BOUNDARY. Earlier STOP_SOURCE_IDENTITY_UNVERIFIED records workflow audit-source/PR-head mismatch and missing head-associated run; later infrastructure success/merge does not silently erase that provenance limitation. Scoped calibration artifact only, not X11 transport, action-target admission or general utility. Subsequent rotation/ATTACK controls (#2468/#2472/#2477) and their HOLDs are separately indexed above.

Historical reconciliation — 2026-10-02 20:27 UTC: lineage freshness, observation readiness, and duplicate ROI proposal

• Current runtime lineage, not merely internal consistency: closed #2428, merged #2433, adds caller-current observation/binding checks before dispatch. Preserve initial checkout STOP_RUNNER_STALL_BEFORE_FRESHNESS_TEST, later test-completion/cleanup-cancellation notes and reported EVIDENCE_DIGEST_MISMATCH instead of expected stale reasons; those records are not clean terminal freshness PASS, and comments reuse run identities across attempts. Duplicate #2443 was not promoted. Separate merged #2456 supplies direct tests with a deterministic digest-bound program/receipt/sidecar and counting dispatch stub: local three-test PASS, stale observation/binding return STALE_OBSERVATION/STALE_BINDING with dispatch0; current lineage delegates once. Coverage only, no new implementation/live experiment. Caller-currentness authority and composition with real focus/target invalidation remain unverified.
• One game observation is not action consumption: closed #2446, merged #2449, probes import/launch/reset/observation/cleanup in private Xvfb. Retained author record: run 35451201198/job 105918548338, pinned Python3.12, ViZDoom1.2.3, /usr/bin/Xvfb, screen shape (3,240,320), PASS_REAL_OBSERVATION_BOUNDARY. Earlier source-identity/missing-log STOP/HOLD remains historical. This admits only an observation-readiness prerequisite, not input delivery, effect, visual-target admission, game success, model behavior or latency/token benefit. Later one-action calibration HOLDs are separately indexed.
• ROI proposal closed as duplicate, with wall-tail HOLD retained: closed #2444 proposes replacing #1299 full-root event-source capture while holding effect/timing gates fixed; collision review found canonical #1346 already completed it, so #2444 adds no experiment. #1346: one64-pair/128-XTerm allocation, effects128/128, timing/ROI64/64, audit zero errors, corruption7/7; observation p95 .940ms and active acquisition1.761ms, but paired controller-wall p95 31.185ms versus <10ms — HOLD_XTERM_TERMINATION_TAIL_PERSISTS, not speedup from negative median. Separate #1601 decomposition retains PASS_XTERM_NATIVE_EXIT_TAIL_LOCALIZED_A4_SCOPED, one64-session allocation: effect→child-ready p95 21.037592ms versus child-ready→native-exit p50 164.064920ms/p95 214.750334ms; audit/corruptions/source checks passed. It localizes this fixture's terminal interval, not internal XTerm/kernel cause or production acceleration, and does not relabel #1346 HOLD. Earlier interpreter/wrapper STOPs remain scientific NONE with no partial-row pooling; unrelated temporal-buffer ROI A6 normalization STOP is a different apparatus.

Historical reconciliation — 2026-10-02 20:30 UTC: dwell-bound provenance, authority receipts, and mixed OOD routing

• Finite timeout bound needs provenance, not just a value: closed #2418, merged #2424, adds a trusted/missing/untrusted/invalid-bound certificate contract after #2279/#2402. Initial missing oracle/source-path preflight and later fetched-file working-directory mismatch retain STOP_WORKFLOW_PATH_MISMATCH_BEFORE_AUDIT, not contract FAIL. Author reports corrected pinned-container runs 35450801899/35450804146 four cases/zero external calls, cleanup successful: PASS_DWELL_BOUND_PROVENANCE_ORACLE_SCOPED. Preserve earlier source stops separately. This validates a provenance contract, not an empirical upper bound M, censoring assumptions, live GUI timeout policy or model/task effects.
• Receipt validation does not provision provider authority: closed #2382, supporting live adapter #2337, leads to merged schema #2385 and pure validator #2389. Pinned run 35449900799/job105915123206, four tests: declared valid receipt accepted, missing/undeclared/wrong fixture-source rejected — PASS_RECEIPT_ADMISSION_FAIL_CLOSED. Separate merged #2410 adds environment-independence regression; earlier checkout stall STOP ran no tests, later reported terminal job105917040813 passes five tests: PASS_AUTHORITY_ENV_INDEPENDENT. AGENT_INTERFACE_LIVE_AUTHORITY=true cannot turn missing/malformed receipt into authority. No actual provider/model/fixture/scorer ownership was provisioned and no live calls occurred; HOLD_NO_MODEL_AUTHORITY remains the preparation's live-allocation disposition, despite closure/merge.
• Mixed known/unknown X11 routing, corrected aggregation only: closed #2425, merged #2427, reports fixed base99th-percentile OOD distance gate plus positive-target prediction/confidence>=.75. Three160-frame blocks: base classifier160/target-accept21/accepted-false0; old shift OOD-YIELD17/classifier143/accept9/false0; fresh OOD-YIELD160/classifier0/accept0/false0. The retained main report (blob96e86d0b0316cbc7b7397097eeb9dfb17437cf7a) discloses first aggregation wrongly counted high-confidence negative predictions as accepts and was discarded; corrected target-class accounting is the published result, not an untouched first-pass success. Training used base160 plus old-shift adaptation80; no independent held-out population claim follows from these totals. CUDA RTX3080/source-bound fixture routing only, non-authoritative: no general safety rate, semantic accuracy, runtime promotion, task/gameplay or latency/token benefit. No experiment was rerun for this index.

Historical reconciliation — 2026-10-02 20:32 UTC: V5 source-lineage duplicate and fresh-family gates

• V5 executable references already repaired elsewhere: closed #2422 proposes replacing stale V1 runner/result references in the Chromium V5 preregistration after #2375/#2337. It closed as a duplicate: parent comment says existing golden_v3_cli_live_chromium_2337_v2 README already points to authoritative V5 lineage; local main144ef669 check found all22 preregistered files, SHA-256 matching retained live-05 preregistration, and compile PASS for runner/preregistration/audit/adapter/API. This is reported static source/readiness verification, not a new live allocation or authority provisioning. Earlier V5 results remain unchanged; no model/GUI/input/network/external workflow ran for this duplicate check.
• Source-bound distance gate, not target admission: closed #2409, merged #2411, uses base160 reference frames and base non-self nearest-distance99th percentile threshold .07607. Reported fixture result: base passes160/160; old shift passes143/160, YIELD17; fresh family passes0/160, YIELD160. “Accept” here means distance-gate pass to subsequent processing, not positive target classification, action permission or task success. All-base pass is reference-population behavior, not an independent generalization estimate. Fixture-only OOD routing; no arbitrary GUI semantics, safety rate, tokens/latency/gameplay or runtime promotion. Later mixed routing (#2425) and integrity audit limitations are separately indexed; no new test was run here.
• Within-family adaptation did not transfer: closed #2404, merged #2405, trains the fixed354-parameter CNN on base160 plus old-shift adaptation80, evaluates fresh160 unique-hash X11 frames with grid background/ellipse/clutter/occlusion. Accuracy50%, accept/YIELD/reject0/160/0, accepted true/false0/0: HOLD fresh-family transfer, no promotion. Zero accepted false cases is abstention, not demonstrated useful accuracy or population-level safety. Old adaptation frames are not fresh evaluation; retained prior within-family results stay separate. No model authority, arbitrary GUI transfer, runtime effect or efficiency claim.

Historical reconciliation — 2026-10-02 20:35 UTC: X11 adaptation, failed augmentation, and exact relevance transfer

• Same-family adaptation, not fresh-family generalization: closed #2399, merged #2401, trains the354-parameter CNN on base160 plus first80 shifted frames, evaluates remaining80. Reported accuracy88.75%, false positives0, accept/YIELD/reject56/24/0; parent separately reports accepted true16/false0. Preserve both summaries: do not call all56 “accepted” rows verified useful positive targets (later #2425 distinguishes positive-target acceptance from confidence alone). It is a bounded positive same-generated-family adaptation result; no population-wide safety, novel-family transfer, authority or runtime promotion. Fresh-family successor #2404 remains HOLD and is not replaced by this result.
• Training augmentation was not useful: closed #2394, merged #2395, adds four translated/noisy copies per original training frame (800 rows), leaving the160 shifted evaluation frames unchanged. Accuracy drops from retained unaugmented98.75% to50%, false positives0, accept/YIELD/reject0/160/0 — FAIL augmentation configuration; no promotion. This is a specific combined translation/noise configuration, not evidence that all augmentation fails or that either factor is causally isolated. All-yield abstention is not useful classifier success; previous HOLDs/results remain unchanged. CUDA/source-bound fixture evaluation only, no model authority or GUI/efficiency claim.
• Exact pixel differences with supplied relevance: closed #2378, merged #2380, extends capture-only #2377 to72 rows (12 fresh Docker/Xvfb/Tk/Xlib sessions×6 cases), fixed320×240 BGRX and40×30-pixel tiles. Author reports nonempty intended changes72/72, oracle mismatches0, relevant false suppressions0, stale/missing full-current fallback12/12 each, fresh irrelevant suppression12/12, critical forward12/12, avoided compressed bytes9048. This is scoped evidence-substrate/relevance-receipt transfer over real captured fixture pixels; relevance is authored, not inferred from arbitrary GUI or a model. Descriptive avoided bytes are not token/latency savings. No gameplay, live runtime authority, broad semantic relevance or task-quality promotion; prior #1635/#2377 retained scope unchanged.

Historical reconciliation — 2026-10-02 20:37 UTC: readiness-oracle mismatch and audit-harness repair

• Execution of the readiness gate failed before a result: closed #2336 proposes running #2296's exact32-vector MAP01 matched-recovery classifier with five laundering controls, retaining current snapshot HOLD. Reported pinned-container run35446160139/job105905325664 fetched1597-byte source and compiled, then first assertion failed: decide(current) included authority:false while oracle(current) omitted it. This is classifier/oracle shape mismatch, not container failure or measured MAP01 readiness. No PASS/live allocation or authority was claimed; old #1866 endpoints remain unchanged.
• Separate shape repair proves a HOLD-classifier contract: closed #2346, merged #2349 and separate merged #2350, minimally includes the existing non-authorizing authority:false field in oracle. #2350 reports native/Python3.12 container output PASS_MATCHED_RECOVERY_ENTRY_GATE_HOLD_SCOPED vectors=32 authorize=1 current=HOLD controls=5/5. authorize=1 is a synthetic candidate count, not granted live authority; actual current readiness remains HOLD because fresh same-run physical/TASK_EFFECT endpoints are absent. Preserve original assertion failure and both PR identities; do not pool parallel repair records or infer a live matched-recovery result. No thresholds, GUI/model/input/network authority or historical evidence changed.
• Import repair uncovered an independent blob-hash defect: closed #2323, merged #2328, follows #2265/#2277's incomplete path-covered lifecycle audit: run35445769575 stopped with NameError because hashlib import was missing/commented by a literal escape. #2328 turns it into a real import, but merge/pending check is not PASS. Later run35446022418/job105904965330 compiled/printed manifest then asserted before lifecycle audit: computed0176b943… versus expected26db03b… for the same source. Review identifies git_blob_sha header using literal backslash-zero instead of Git's NUL byte, so the import-only repair merely exposes the next unconditional mismatch; do not treat that mismatch itself as proof of altered source or mere base drift. Subsequent NUL-header successor #2344/#2348 is separately indexed. Original import STOP and failed hash audit remain retained; all are static harness/provenance evidence, not adapter execution, task completion, authority or performance results.

Historical reconciliation — 2026-10-02 20:39 UTC: path-covered lifecycle audit and CLI adapter rungs

• A covered audit workflow is not a completed audit: closed #2265, merged #2277, addresses empty-check ambiguity for retained #2259/#2260 static contract payload. Adds a path trigger/copied payload, seeking exact ten rows/four source blobs and INDEPENDENT_AUDIT_PASS rows=10 authority_grants=0. Parent explicitly records merge while formal check pending and other legacy checks failing; no PASS follows from merge. Later missing hashlib and Git-header NUL defects/repair stops are separately indexed under #2323/#2344. Static workflow coverage only, not adapter execution, GUI/model/input or end-to-end lifecycle completion.
• Synthetic adapter trace progresses beyond static mapping, not into live tasks: closed #2309, merged sole-owner #2314, pins four current-main source blobs and evaluates exact nine-state synthetic setup/observation/dispatch/refusal/effect/stale/repair/release/cleanup trace. Author reports run35445836055/job105904478666 native and pinned Python3.12-slim compile/JSON/audit PASS: ADAPTER_TRACE_AUDIT_PASS attempts=9 authority_grants=0 unknown_fail_closed=1 cleanup_failures=1. Cleanup failure is deliberately retained as a typed outcome, not hidden or counted task success. Duplicate #2319 closed/unmerged and its checkout-stalled run35445892910 remains execution history; no pooling. Deterministic side-effect-free integration trace, no live GUI/model/OS input/network task or latency/token benefit.
• Explicit golden-v3 wrapper at CLI boundary: closed #2292, merged #2303, exposes dispatch_golden_v3 and typed result mapping without changing underlying dispatch admission. Parent reports immutable-source pinned-container run35445767101 compile/six tests/cleanup PASS and platform CLI/portable-build checks passing; initial harness failure omitted selector_v1 dependencies and is retained separately. Tests cover returned/refused/unavailable/failure/stale/partial/cleanup/unknown forms, preserve diagnostics and authority false; completion is not application-effect success. These are source/runtime contract and packaging observations at that checkpoint, not live model/desktop evidence or permission to infer later distribution health.

Historical reconciliation — 2026-10-02 20:41 UTC: lifecycle source contract, offline mapping, and retained replay

• Source-bound lifecycle contract before implementation: closed #2259, merged successor #2283, freezes golden-v3 launcher/report, CLI API and core lifecycle mapping with explicit partial/failed calls and authority0. Parent reports current-main run35445513369: four source Git blobs match, ten lifecycle mappings agree, authority/model/GUI/input/network zero. PASS_DESKTOP_VERTICAL_SLICE_CONTRACT_AUDIT_SCOPED is a static artifact/source-audit claim, not adapter or task execution. Earlier #2260/#2270 base-tree/audit failures and #2178 HOLD remain unchanged; path-trigger/harness repair history is separately indexed.
• Fail-closed result-schema translation: closed #2269, merged #2273, maps frozen golden-v3-result-v1 to CLI without silent coercion. Reported immutable-source Python3.12-slim run35445434792 compile/eight tests/cleanup PASS: missing/unknown status and lifecycle rejected, diagnostics/partial effects/usage retained, task_success distinct from program_completed, cleanup failure not success, authority=true rejected. Initial checkout-stalled run35445263341 retained separately. Offline finite candidate/oracle agreement only; no live backend/desktop/model effect, efficiency or production claim. Later CLI wrapper (#2292) is a separate integration rung.
• Retained-result replay, while live dependency boundary remains STOP: closed #2274, merged #2282, translates a hash-bound retained golden result into ten lifecycle rows, preserves raw reference/task-program distinction/partial effects, cleanup failure runtime_failed, and rejects malformed/missing provenance. Local Docker/Python3.11 reports ten-row independent audit plus three tamper/provenance controls PASS, all authority/model/GUI/input/network counters0. Separate merged current-main workflow successor #2301, run35445663166, reports compile/replay-audit/controls PASS with retained artifact10585103493: PASS_GOLDEN_V3_CLI_ADAPTER_REPLAY_SCOPED. PR body historical “draft #2282” wording does not change its fetched merged state. A distinct later live rung stopped before GUI/model/input because sparse hydration of research/live_control did not complete and left Git locks; runtime-only closure could not import golden-v3 dependencies. This environmental STOP does not revoke replay PASS or become a live semantic FAIL. No new task input was replayed and no live completion/latency/token/production claim follows.

Historical reconciliation — 2026-10-02 20:44 UTC: offline effect-contract integration and container provenance

• Offline task-effect admissibility, not measured MAP01 effect: closed #2241 follows verified #2217 after concurrent-base merge rejection. #2217 is now merged and reports an eight-control candidate/oracle contract separating physical/state-feedback/plan-bound TASK_EFFECT with authority false. Initial successor file-create API HTTP422 wrote no files; not scientific FAIL. Later parent notes main already held the four-file contract and describes #2262 closed/unmerged; fetched #2262 API instead has merged=true, despite its body still saying closed without merge. Preserve this source-description/state disagreement and later reported workflow35445109126 eight-case PASS; do not invent a clean collision-resolution chronology. Occupancy-frontier STOP/PASS comments on this Issue concern #1592's different experiment, not evidence for this effect contract. No live action, occupancy correctness, gameplay/recovery efficacy or model authority follows.
• Provenance completeness after independent stops: closed #2196, merged #2202, adds exact parent source/blob, resolved image digest, stdout and result hash to generation-bound revalidation. Earlier run35444414669 retained FileNotFoundError (audit opened RESULT.json from wrong working directory); run35444562604 retained STOP_PROVENANCE_IMAGE_DIGEST_CONTRACT (tag rather than immutable digest). Later author-reported run35444807157 verifies parent blob/resolved digest/output/result and independent audit, artifact10585252213: PASS_PROVENANCE_COMPLETE_SCOPED. Analysis-index drift failure stays separate, not hidden by formal-job PASS. This verifies deterministic container evidence lineage only, no live X11/PNG/model/task value or transfer.
• Canonical digest syntax, not changed scientific harness: closed #2247, merged #2250, follows #2243's overly restrictive docker.io/library/python@sha256 prefix requirement. Actual inspect value python@sha256:2f17fc… is preserved; audit uses canonical digest equality rather than accepting a mutable tag. Reported fresh run35444857681 matches source/blob/stdout/result/counters and prints INDEPENDENT_PROVENANCE_AUDIT_PASS, artifact10584313318, PASS_CONTAINER_PROVENANCE_COMPLETE_SCOPED. Separate merged result publication #2258 records this same artifact; it is not another independent replication. Earlier failures unchanged, analysis-index failure independently retained. No GUI/model/performance/production claim.

Historical reconciliation — 2026-10-02 20:46 UTC: model-facing readiness STOP and emitted-result scopes

• Typed readiness did not establish real recovery value: closed #2234, merged #2242, proposes transferring #1247's readiness vocabulary to two held-out application-like routes with the same model/policy, incomplete/aliased signals and independent effects/costs. Actual disposition STOP_EVIDENCE_ACQUISITION_GAP: routes, model allocation, independent task-effect scorer and token/latency accounting unavailable. A container-checked stop ledger is not a recovery experiment PASS; model/GUI/input/runtime mutation zero. Private classifier correctness is not substituted for model use/task benefit; ordinary admission remains separate from readiness.
• Source-emission audit retains missing field provenance: closed #2227, merged #2236, pins v3 launcher and frozen v2 report-construction source. It finds machine-readable schema/usage/routes/task rows/release/independent evaluation, but not source-backed program_completed/task_success/partial_effects/status/cleanup-error fields required for the full contract: HOLD_FIELD_PROVENANCE_INCOMPLETE. Independent stdlib source audit only, no runtime/model/GUI/input/container execution. Existence of a report does not prove every proposed result field is emitted; no production schema or adapter admission is inferred.
• Read-only mapping of one actual retained report: closed #2226, merged #2233, validates pinned golden-desktop-app-server-v3-live-01/golden-report.json blob e168a9bdc84fc6b807f4e90806ec7c501da89689. Reported Python3.12-slim validator PASS_EMITTED_GOLDEN_V3_MAPPING_SCOPED: six exact submissions/six independent evaluation records, verified release, old-target admissions0, doctor/repair evidence; task_success derived from independent evaluation rather than program completion, authority false. This is a retained-artifact interpretation/check, not a fresh GUI/model allocation or proof that every underlying source constructor emits the full proposed schema. Keep #2227's broader field-provenance HOLD and earlier #2215 HOLD separately; no latency/generalization/production claim and no rerun of historical tasks.

Reconciliation — 2026-10-02 20:52 UTC: new successors and historical contract audits

• Matched phase control, separate from A03: #6803 now records PASS_METHOD_SCOPED; PR #6805 remains open/unmerged. Identical captures [10,50] ms and horizon 60 ms isolate the authored cue-phase contrast. One native-Windows candidate and one independent auditor (exit 0/0, retries 0) produced/reconstructed 9 rows, zero audit errors and 5/5 corruption rejections. This supersedes only the earlier index observation that no A04 result was available; #5694 A03's HOLD_PHASE_CONTRAST_NOT_MATCHED is unchanged. Finite synthetic method evidence, not live causal prevalence, GPU or controller benefit.
• Image-only looming assumptions: #6808, open, proposes a 12-case rendered-image gate, including a pixel-identical approach/object-growth pair that must remain UNKNOWN. Current disposition STOP_PRE_CANDIDATE_RESOURCE_HOLD / NOT_EVALUATED: host construction 3/3, formal candidate/auditor/retries 0/0/0; the proposed exclusive private OrbStack assignment was not available. No scientific result, no existing container modified. Preserve #5905's prior STOP separately; a proposed observability gate is not live safety evidence.
• Mandatory obligations after early negative: #6809, open, succeeds #6689/#6706 rather than rewriting their FAIL/STOP. Canonical PR #6812 is merged: the reported WSLc CPU-only synthetic PASS covers five prefix cases, independently reconstructed, with five mutation rejections and no retries. The full mandatory vector stays pending after an early negative; classification counts are separated from stability metrics. #6811 was a duplicate publication, not a second allocation. The Issue was reopened after premature closure; its latest comment keeps it open while merged-head analysis-index/replay-gate checks finish (pending in that observation). Merge is not proof of those checks or production correctness.
• Independent focused-observation raw-row audit: closed #2137 succeeds #2058. PR #2219, merged, records run 35444300481: 256 rows independently agree, accepted 8, mismatches 0, authority-positive 0; artifact 10585066487. A separate Issue comment records run 35444413889, artifact 10584437623 and INDEPENDENT_RAW_ROW_AUDIT_PASS, also 256 rows/8 accepts. Keep these run/artifact identities distinct, not pooled as one execution. Request-validation only; no region-discovery, GUI/model, latency or utility claim; #2058 and #2154 HOLD unchanged.
• Golden-v3 schema freeze with validator caveat: closed #2186, merged PR #2194, retains PASS_GOLDEN_V3_RESULT_SCHEMA_FREEZE_SCOPED: five lifecycle fixtures accepted and cleanup-failure-as-success rejected. Post-merge audit says the stdlib validator does not enforce all schema const/enum/field-specific constraints and was not compared with an independent JSON-Schema implementation. This limits validation strength, not a retraction of the schema itself. Distinct run 35444335830 retains STOP_TEST_HARNESS_MISSING_PYTEST; compilation invocation is not adapter-test success.
• Schema-to-CLI mapping lacked source-backed fields: closed #2203, merged PR #2215, records HOLD_ADAPTER_CONTRACT_NOT_SOURCE_BACKED: five pinned source blobs, 3 accepted/4 rejected synthetic controls, but program_completed/task_success/status/partial_effects remain unverified. CLI returned is not task success; invalid_request/backend_unavailable/runtime_failed and cleanup_error require explicit diagnostics/fail-closed handling; historical receipts grant no freshness. PR #2209 is closed/unmerged, superseded with findings preserved. Later report-mapping/emission successors indexed separately do not retroactively convert this HOLD into PASS; no production adapter or live execution was established.

Preservation reconciliation — 2026-10-02 20:55 UTC: symmetry and clock units

• Typed identity symmetry: open #6251 proposes a role/authority-preserving state quotient with concrete counterexample witnesses. Merged PR #6257 preserves the Windows-host-only T0's historical PASS_METHOD_SCOPED, 312 named states → 166 verifier-swap classes, 624 equivariance and 624 property checks, and four identity-breaking controls. The main preservation qualification limits independent witness replay to baseline named traces: mutation traces, distinct class coverage, quotient membership/inverse mapping and lifting a separately explored quotient trace are not independently established. Class counting followed full enumeration, so 46.8% fewer classes is not measured exploration-cost reduction. A one-nibble preregistration/FREEZE candidate-hash discrepancy and ambiguous freeze counters remain unresolved; large raw was not fully retrieved in preservation review. No rerun or full-PLAN revalidation; earlier local readiness HOLD remains separate, with no real-worker, live, fairness/liveness, authority or container claim.
• Exact-runner clock/cancel T3V: open #6261, merged PR #6266, retains HOLD_D_CLOCK_RETURN_BOUND_AMBIGUOUS despite two mock-case sub-results and the historical raw-only audit passing. The main preservation qualification identifies a factor-1,000 prose/raw mismatch: nanosecond-labeled case-A fields subtract to 55,791 ns = 0.055791 ms after clock return and 36,166 ns = 0.036166 ms after cancel request, not the original 55.791/36.166 ms. Therefore the prose's claimed 5.791 ms miss is not established by those fields. This static arithmetic finding neither replaces the audit nor regrades the historical HOLD; threshold-origin ambiguity and original evidence remain visible. T3V has no cancel-first runner arm, so cannot inherit #6236's different A/C comparison or demonstrate a real-runner benefit. Preserve #6252/#6254 STOP separately; host fake-session/backend evidence only, no natural transport, live MAP01, task effect or #59-completion claim.

Preservation reconciliation — 2026-10-02 20:59 UTC: predictive histories and runner STOP lineage

• Observable predictive tests / within-route history: open #6258 asks whether finite future-test vectors preserve effect-relevant history without turning predictions into authority. Merged PR #6263 preserves the Windows-host allocation OBSERVABLE-PREDICTIVE-TESTS-6258-T0-HOST-20261002-01 and its historical PASS_METHOD_SCOPED: one candidate/one auditor, retries 0, nine cases/66 bounded transition outputs, zero reported mismatches; construction 16/16. Silent-effect alias is held despite equal observable vectors, a fresh independent receipt separates histories, the null pair merges, and step-3 divergence stays UNKNOWN beyond horizon 2. The main preservation qualification records committed LF raw versus recorded CRLF sizes/hashes: a diagnostic newline conversion reproduces the recorded hashes but is not contemporaneous artifact recovery. Original bytes/manifests/PASS labels are unchanged; preservation is not a new audit or endorsement. Separately, Issue comments report a local OrbStack allocation observable-predictive-state-6258-t0-v1-20261002-01 with five pairs, ten audit gates, two rejected mutations and construction 5/5; its result comment says no branch/PR/push/merge. Do not combine that local comment-only packet with #6263's nine-case host packet or infer its later integration. Neither establishes learned PSR, complete live vocabulary, long-horizon safety, route conformance or real effects.
• Runner clock/cancel T3, executed before hash registration: open #6242, merged PR #6244, preserves the original protocol STOP: candidate was invoked before candidate/auditor hashes were recorded; raw traces and arm summaries were retained, with no rerun or scientific PASS. The main preservation qualification notes different standalone/embedded construction-summary timestamps, unestablished common provenance, narrative-only initial construction failure, and declared rather than independently logged invocation counts. Timestamp disagreement alone does not prove a particular count. The clock-return marker differs from the separately returned planner-end value; the reported 0.111 ms measures marker-to-backend cancel observation, not cancel-request delay. Packaging/index checks and posthoc arithmetic do not repair the consumed allocation or validate live transport.
• Same Issue, distinct T3S pre-candidate design STOP: #6242 comments preserve MAP01-V6-RUNNER-CLOCK-CANCEL-59-T3S-20261002-01 at frozen d3839dbce01da43f9663ca9214355181b206f603. Construction smoke and six pre-freeze audit mutation tests passed, but a final source/harness check found clock_return_ns sampled before the timestamp returned to the runner, defeating the frozen exact end-clock equality; backend Lease used a different real clock domain. Candidate invocations 0, retries 0, no formal result; frozen sources/gates were not edited. This is an untested hypothesis under a design STOP, not candidate failure, not T3's executed protocol STOP and not later #6261 T3V's HOLD. No GUI/game/model/input or production benefit follows from any of these mock-clock records.

Historical reconciliation — 2026-10-02 21:02 UTC: golden contract audit and generation reproduction

• Golden desktop → CLI contract, reported PASS versus actual audit coverage: closed #2172 asks for source-backed lifecycle mapping, task/program separation, partial effects and non-success cleanup semantics. Closed-unmerged PR #2178 retains HOLD_INTEGRATION_ADAPTER_GAP_PENDING_V3_SCHEMA_FREEZE. Separately, merged PR #2182 historically reports PASS_DESKTOP_VERTICAL_SLICE_CONTRACT_AUDIT_SCOPED, 11 mapped states and digest 24861e56…; that label remains visible but is not independent proof of the required gate. P1 review says the auditor never reads the six frozen source blobs/identities and hardcodes authority_escalation=false. Current main audit.py, blob 0b9788109a3f93b53716009a990bc49b46bc3c4c, confirms its presence checks use manually assembled frozen_text and its authority flag is constant. Thus the PASS validates that authored fixture, not source-backed mappings or authority regressions. The Issue's independent source comment retains the mapping gap; later #2186/#2198/#2203 successors remain distinct. No adapter implementation, live desktop or product acceptance follows from either PR's closure/merge.
• Source-first schema-freeze duplicate: closed #2187 proposes pinning actual launcher/v3/CLI sources and marking absent result fields unmappable. Source comments distinguish v3 doctor vocabulary and delegation to v2 from a typed live execution-result schema; the v3 evidence document is not that schema. PR #2193 is closed/unmerged as a duplicate of pre-existing #2186, preserving its evidence without making it canonical. Closure is coordination state, not successful schema verification; #2186's proposed-schema PASS and later #2198/#2206 HOLD_PROPOSED_SCHEMA_NOT_SOURCE_BACKED must not be conflated with this duplicate.
• Generation-bound evidence, container-only reproduction: closed #2166, merged PR #2177 and mechanical result-record PR #2191, retains PASS_CONTAINER_REVALIDATION_SCOPED from run 35444087832. One python:3.12-slim formal execution and independent audit: five cases, ENCODED/OBSOLETE/CACHE_HIT/ENCODED sequence, stale rejection and byte-identical exact-generation reuse pass, model/X11/input counters 0. Artifact 10585146161, ZIP SHA-256 a8634e1f92013b904da8e153c42d586487b76c745b0afe6c946e92daff7d9db6; repeated Issue comments cite the same run/artifact, not independent repetitions. Parent #2047/#2159 HOLD remains unchanged. Container reproducibility/bookkeeping only, not PNG/X11 capture, model-boundary usefulness, latency or task correctness; result publication is not a rerun.

Historical reconciliation — 2026-10-02 21:04 UTC: preflight, evidence preparation and import readiness

• Model-independent AltGr preflight: closed #2171, merged PR #2176, retains PASS_ALTGR_PREFLIGHT_CONTRACT_SCOPED for six authored states: only DIRECT_LEVEL0_3 yields DIRECT_TYPE; dead/compose, stale map, unknown modifier and partial delivery abort, alternate text path is not direct typing. Two P1 findings limit reproducibility/audit strength: experiment.py overwrites RESULT.json with rows_detail only, losing status/counts/scope metadata, and audit coverage checks six rows plus one DIRECT_TYPE but not complete unique state membership. Current main experiment and auditor still show those behaviors; duplicated valid abort rows can hide missing states. Historical PASS is not regraded, but does not prove adversarial coverage or full-schema reproduction. Analysis Index run 35444087989 separately failed on missing entries at that time. No physical receiver, model policy, multilingual application, clipboard, latency or platform-transfer claim; #355/#2151 acceptance remains separate.
• Generation-bound early preparation, model value still untested: closed #2047 proposes WAIT versus early-prep/reuse versus stale-unchecked comparison, including packaging/model cost and task quality. Merged PR #2155 records PASS_GENERATION_BOUND_PREPARATION_CONSTRUCTION_SCOPED, five reuse/rebuild/wait cases only. Separate merged PR #2159 retains HOLD_PRE_LIVE_X11_MODEL: deterministic PPM sequence ENCODED/OBSOLETE/CACHE_HIT/ENCODED, obsolete bytes absent, exact-generation reuse identical and stale control rejected, digest 9d4e4bb5…; Docker execution stopped before invocation amid shared workloads. Neither rung tested live X11 PNG/model decisions/end-to-end latency. Later #2166 container reproducibility PASS is already indexed separately and does not close this model-facing hypothesis or turn local preparation into task utility.
• MAP01 import namespace repair: closed #1962 succeeds #1950/#1954's preserved failure. Diagnosis: Doom session_v7's bare import loads live_control session_v8, whose bare session_v7 resolves back to the partially initialized Doom module; switching Doom v7 to v6 alone repeats the collision. Merged PR #1980 adds a private namespace adapter/import-only smoke. Retain distinct first stops: not-yet-implemented source audit, run 35442055216's missing transitive NumPy, and bounded checkout/container acquisition termination. Later run 35442280826 after Pillow/python-xlib/NumPy/openpyxl dependency closure reports PASS_IMPORT_GRAPH_REPAIR_SCOPED, compile and intended live_control module origins only. A separate current-main partial import check encountered openpyxl and did not complete the full bounded smoke. PR #2158 is closed/unmerged as superseded by #2038/#2049, preserving corrected diagnosis/partial evidence; not another integrated PASS. No retained v12/fixture rewrite, GUI/X11 session, model, MAP01 formal run, occupancy or gameplay result follows from import readiness or Issue closure.

Historical reconciliation — pair/arm binding and registered frame identity (2026-10-02)

• Pair metrics bound to arm evidence: closed #1632 succeeds #1609/#1627's CONFIRMED_V6_SUMMARY_ARM_BINDING_GAP. Merged PR #2199 reports PASS_V6_BOUND_AUDIT_SUCCESSOR_SCOPED, six independently mutated pair↔arm numeric bindings rejected; original v6 audit/allocation/thresholds remain untouched. However P1 layout finding says real v6 writes summary.pairs and pair-01/coast_control or bounded_recovery, while the wrapper/test use summary.pair1 and pair1/coast or recovery; truthful real results therefore fail rather than retain the base decision. A separate entrypoint finding says the new module has no CLI and existing workflow calls the old audit. Current main wrapper, blob 1fb5143086784e86565057c3d1e5bd9ca6d05b7e, still uses pair1-style paths and only defines functions, confirming those implementation boundaries without execution. Preserve reported synthetic PASS, but do not treat it as verified deployment on actual v6 artifacts or workflow enforcement. Workspace Index SUCCESS is packaging evidence, not repair coverage. Even correct numeric binding would not validate arm occupancy measurement or MAP01 efficacy.
• Registered source-frame identity before semantic facts: closed #2135, merged PR #2142, separately succeeds #2004's failed #2128 allocation. Run 35443853558 on head 8198ca633840849bf29bbdfc2397b931a43b95cf reports PASS_SEMANTIC_DIFFERENCE_IDENTITY_REGISTRY_SCOPED: one formal invocation, five controls, exact oracle agreement; unknown/stale/surface-replaced identities fail closed. Artifact 10584811228, digest 24ab5a4747973e4f8179585fd0fbd0c7736f2607a790d680e2deac05151329ff; authority/calls/task-input/rerun/replacement/tuning counters 0. This finite synthetic registry boundary does not regrade predecessor failure or prove model usability, GUI truth, latency/token value, transfer or runtime promotion. Keep other #2004 successor allocations separately identified.

Reconciliation — pre-live hypotheses and new oracle-blind successor (2026-10-02)

• Resident reaction against GUI predicates: closed #2055 proposes controlled live predicate/effect comparison with polling, stale/open-loop controls and terminal release. Merged PR #2149 retains HOLD_PRE_LIVE_GUI, a twelve-case authored harness, digest 8a913418…; Docker formal invocation stopped before execution amid concurrent workloads, GUI/model/input counters 0. Its reported one-shot/duplicate/transient behavior is not live GUI evidence. P1 condition finding says invalidity is selected by scenario-name blacklist rather than authority/timing state in event traces, so it cannot establish handling of revoke/deadline/stale/focus/target/gap/already-effected transitions. Release finding says released=true is unconditional with no release lifecycle, leaving reported all-cases release unsupported. Keep those review limits with the original HOLD; no deployment, planner-delay task benefit or authority guarantee follows from closure/merge.
• Temporal-ring history for model disambiguation: closed #2045 succeeds #1084's transport timing, asking whether task-specific history improves decisions without stale/window/surface confusion. Merged PR #2053 reports PASS_TEMPORAL_RING_PROVENANCE_CONSTRUCTION_SCOPED, six controls/five UNKNOWN. Separate merged PR #2145 records HOLD_PRE_MODEL_TEMPORAL_RING_EVALUATION, ten controlled ledgers, eight oracle agreements, four UNKNOWN, model/authority/input 0, digest 16971dfb… . First-draft temporal-role/surface conflation and failed oracle assertion are retained; correction preceded the single recorded result. Eight agreements are not full validation. No live X11 capture, model decisions, history cost or ring utility measured; transport speed is not attention/reasoning benefit. Earlier index failure was a separate packaging gate.
• Oracle-blind return-cue packets: new open #6819 succeeds #6492/#6499/#6520 without rewriting their results/STOPs. Proposal isolates hidden truth in an auditor-only file/mount, requiring identical outputs for byte-identical public views despite conflicting hidden return labels; fresh user-authored cues remain context, not target authority. Pre-execution correction fixes the frozen total to 14 rows (four equivalence pairs = eight rows, plus six distinct controls), not the body’s original eight. Planned independent raw-only audit/eight mutations cover truth exposure, row integrity, stale/agent cues, wrong window and release loss. No candidate/auditor result is reported in fetched Issue/comments; treat proposed/unverified, not PASS. Even successful abstention/invariance would validate only the information boundary, not human resumption, GUI usefulness, latency, accessibility, privacy or product benefit.

Historical reconciliation — authorized contracts, collateral outcomes and independent evidence (2026-10-02)

• Authorized effect-time contract generations: closed #2072 extends #532's temporal binding with issuer/scope/target/session/epoch admission. Merged PR #2089 records PASS_EFFECT_TIME_CONTRACT_AUTHORIZATION_SCOPED for authored matching/reject controls and later-generation evaluation. Retain that historical label alongside P2 structural finding: admission can accept records missing generation or value, with evaluate failing on missing payload. Snapshot finding says the original admitted dictionary is not copied/frozen, so in-place value mutation can rewrite historical evaluation; the control only creates a separate later dictionary. Authority finding says zero promotions derives from a constant false flag, not exercised authority semantics. Review also notes D inventories artifacts rather than defining a decision rule and C repeats the outcome. Finite control PASS does not prove immutable admission, complete malformed-input rejection, real issuer trust or full contract integrity; #532's older result is unchanged.
• Primary restoration versus complete collateral success: closed #2076, merged PR #2087, records PASS_TYPED_PARTIAL_COLLATERAL_OUTCOME_SCOPED: complete success needs primary+collateral, primary-only restoration is PARTIAL_PRIMARY_RESTORED, missing/stale evidence UNKNOWN. P2 retention finding says output rows drop primary/collateral/generation inputs and set raw_retained=true unconditionally, so the published digest does not bind receipts or generation provenance. Keep typed classifier result distinct from the unsupported raw-retention claim. This finite reducer does not establish model completion behavior, application semantics, live compensation or task success, and does not rewrite #464's prior boundary.
• Independent effect evidence under scorer compromise: closed #2078, merged PR #2082, reports PASS_INDEPENDENT_EFFECT_EVIDENCE_BOUNDARY_SCOPED: legitimate pair ACCEPTED, forged/wrong-target REJECTED, contradictory effect CONFLICT, missing evidence UNKNOWN, authority promotions 0. P2 identity-order finding identifies an omitted combined control: differing effect_id plus differing outcome reaches CONFLICT before identity rejection, misclassifying unrelated effects as contradictory evidence. Preserve the reported finite controls but do not infer complete identity handling, causal independence of real evidence roots, compromise recovery, verification cost or runtime authority. #1295's compromised-root negative result remains separate; multiple signatures alone are not independent provenance.

Historical reconciliation — syntax and expectation repairs without regrading predecessors (2026-10-02)

• Typed-deadband syntax successor: closed #2062, merged PR #2080, preserves #2017/#2020 run 35442719440's STOP_HARNESS_SYNTAX_ERROR separately. Successor run 35443071825 reports formal_mechanics_pass=true for three coast/recovery pairs, six decisions/arm and fixed planner_wait=0.34/cover_budget=0.24/deadband=0.06; balanced keys, empty terminal states, five recovery guards, stale repress 0, maximum guard→app release 0.279750 ms, decoder error 0.01760504, lower unsafe time than coast in three pairs. Review limits the interpretation: missing RECOVERY_BASELINE means the declared primary deadband-versus-baseline comparison was not run; exposure gate conflates deadband and source-bound invalidation, so both required types are unproven. Rendered frames were discarded, preventing pixel-level reconstruction; failure upload would be skipped on nonzero mechanics exit. Historical bounded synthetic X11 result is not complete preregistered comparison, latency generalization, model/DOOM efficacy or production readiness.
• Focused-observation first container expectation STOP: closed #2046, merged PR #2048, records STOP_EXPECTATION_COUNT_MISMATCH: 256 rows, candidate/inline expectation accepts 8, mismatches 0, but harness asserted 16. This is a counting/preregistration stop, not a request-contract mismatch, with no retry under that identity. Review flags authority refusal reason labeled ok despite rejection and vacuous authority count, sum(ok and False), with auth absent from raw rows. Thus reported authority admissions 0 is not measured input/output evidence in this predecessor. Keep its original raw hash and STOP; no region-discovery, GUI/model utility or runtime promotion.
• Separate expected-count repair: closed #2052, merged PR #2058 and current-main reconstruction PR #2075, historically report PASS_FOCUSED_OBSERVATION_REQUEST_CONTAINER_REVALIDATION_SCOPED, one formal invocation, 256 rows/accepted=expected=8/mismatches=0/authority-positive=0, digest 00ef9286… . Independent-audit finding says audit reruns experiment.py and trusts its own counts rather than independently reconstructing retained decisions/hash; #2075 retains the same correlated-oracle limitation. Working-directory finding says root invocation cannot locate the sibling script. Preserve the reported finite PASS with these audit/reproduction limits; #2075 is reconstruction, not a fresh scientific replication. Later #2137 raw-only independent audit is indexed as a distinct successor, not a rewrite of #2046/#2052 or #2009.

Latest reconciliation — 2026-10-02 16:12 UTC follow-up

Selected proposals, setup outcomes and preservation changes; no experiment or resource operation was performed by this indexing task.

• #6749, OPEN — Decisive negative verdicts must retain pending obligations; disposition counts must exclude derived metrics. New successor preserves #6689/#6706 FAIL_METHOD and #6704 FAIL_AUDIT / STOP_CLASSIFICATION_COUNT_CONTRACT. Proposed independent finite oracle audits every legal prefix, pending mandatory/source-frontier obligations and five corruptions. Prefreeze construction amendment adds CLEAR→CONFLICT→COMPLETE after the initial fixture lacked future conflict: 32 authored worlds / 2,112 interleavings / 4,064 unique prefixes, host checks 5/5. Temporary construction smoke is not retained formal evidence; candidate/auditor containers zero at latest comment. No method result, useful latency, GUI/source truth or authority claim. Amendment.
• #59, still OPEN — Local WSL model-route readiness has split outcomes, not a formal live-control result. Open/unmerged #6752 records one exploratory qwen3:4b request to temporary Arch WSL Ollama using the existing Windows model store: reported CUDA route PASS, 37/37 GPU layers, peak utilization 92%; response-contract FAIL, empty response because the 32-token cap was consumed by thinking, sentinel absent. No retry/tuning. Full 31,264-byte server log is absent from the commit; its digest alone is not independently verifiable, so this index records the published report, not a raw-log audit. One-shot durations are not a benchmark. Older #6747 is closed UNMERGED; do not confuse it with integration. Formal R134 candidate/auditor/game remain unrun; no MAP01 or live-threat-control completion.
• #59 separately has open/unmerged #6750, read-only model-store mount prelaunch STOP: unassigned allocation/shared bridge-state uncertainty, candidate/auditor 0/0, not evidence that mount access fails. Host synthetic tests 5/5 and three manifest/sixteen descriptor metadata fixtures do not test container/store access; no blob contents read. Earlier host-network option unsupported before container creation is route-specific, not impossibility of every host route. Bridge-ID change causation remains unknown. Passing publication workflows do not turn setup STOP into a mount/model/game result.
• #5352 — Effect-qualified hysteresis discriminator remains preparation. Open Draft #6753 proposes independently auditing benign versus harmful mode divergences, unsafe prefixes, critical deadlines and authority/effect envelopes rather than substituting final task labels. T15 candidate/auditor/formal containers 0/0/0; idle resource snapshots are not allocation or result. Preserve T0 FAIL_SAFETY_OR_REFERENCE_GATE (hysteresis final-mode divergence 14/155 eligible traces; minimum dwell 54/155); neither task harm nor task-equivalent benefit follows from that mode-only gate. No retrospective waiver or live efficacy.
• #6081 — Invocation is not scientific execution. Open Draft archive #6754 preserves S1 STOP_RUNNER_ARGUMENT_BINDING / NOT_EVALUATED (empty-argv wrappers exit 0 in interactive Python; scripts not run) and S2 STOP_RUNNER_SOURCE_NOT_MATERIALIZED / NOT_EVALUATED (one CLI attempt exits 2 before module load; candidate module/auditor 0/0). Result JSON remains absent; planned FREEZE budgets are not actual counts. Twenty-four original files/65,969 bytes and historical CRLF receipts retained without importing/running science; static identity verification is not method PASS or historical-host authentication. Distinct #6087/#6091/#6357 outcomes unchanged; no pooling, missing-result reconstruction or Issue completion.

Selected proposal/provenance follow-up — 2026-10-02 16:14 UTC:

• #6342 — Safe-stop framing T0 is preregistered, not a human-reactance result. New REACTANCE-SAFE-STOP-6342-T0-ORB-20261003-01 freezes four authored cases × directive/autonomy-supportive variants (eight cards) plus unsafe-retry/false-success controls. Candidate and independent auditor are planned once each in separate pinned offline OrbStack containers; at the latest registration, neither formal stage started. Six construction tests are excluded. Full PASS additionally requires an independent blinded semantic reviewer; missing review means HOLD even if field parity passes. Equal eight-word framing lines/structural facts do not prove equal comprehension, perceived autonomy or reading time. No participants, reactance, audio/GUI/model/safety/product measurement or T1 claim.
• #3442 — Historical generation-fencing v1 remains NOT_EVALUATED / STOP_AUDITOR_SYNTAX_ERROR_RAW_CAPTURE_INCOMPLETE. New source-retention review confirms five predecessor source/STOP files already match main, but FREEZE says formal candidate/auditor 0/0 while STOP says candidate exit 0/auditor exit 1; raw stdout is absent. Preserve both recorded claims rather than invent execution counts or a scientific result. Exact source/history preservation and redundant-ref cleanup are not raw recovery, a rerun or resolution of this discrepancy. Separate v2 archive #6688 retains its own HOLD (auditor rejects valid null operation_index rows); v2 does not regrade v1.

Latest reconciliation — 2026-10-02 16:10 UTC follow-up

Selected new outcomes and delivery changes; this indexing task ran no experiment. Earlier timestamped snapshots remain historical.

• #6147 — Safe distinguishing probes: A03 is FAIL_AUDIT_CONTROL_COVERAGE, not method PASS. Merged #6748 retains one host-CPU candidate and raw-only auditor, exits 0, agreement across nine finite policy-tree layers. The frozen auditor did not inject an unsafe-probe policy and assert rejection, so the preregistered five-control gate was unmet despite its PASS_METHOD_SCOPED label. No postfreeze fix/rerun; A02 STOP_CANDIDATE_RUNTIME_ERROR and its raw-byte uncertainty remain unchanged. Finite authored host-stdlib method only, no container/model/GUI or safe-probe benefit. Qualification.
• #5424 — Severity-sensitive error budgets can reverse count-only rankings, with fallback costs still visible. New T4 candidate/auditor once each on Windows host CPU independently reconstruct eighty history/128 continuation rows, zero mismatches, five corruption controls rejected: PASS_METHOD_SCOPED. Count-only freezes A with eight recoverable failures but not B with one catastrophe plus one recoverable; ordinal/separate hard-catastrophe gates freeze B. Held-out authored primary catastrophes fall 2→0, while incident-correlated fallback severe events rise 2→3 and completed offers fall 6→5. No calibrated severity weights/SLO/live safety/model/runtime benefit. #6755 is OPEN/unmerged (currently non-draft, superseding comment draft wording); review/CI not inferred from local tests. Audit-digest transcription correction and sparse-checkout/CRLF validation limitations retained; no rerun after main advanced. T2/T3 unchanged. Result and correction.
• #6413 — Evaluation-cue construction now has a static synthetic method result, not model-reactivity evidence. Merged #6751 retains separate network-disabled WSLc candidate/auditor once each: METHOD_PASS_SCOPED, twelve valid rows/six cue pairs, three planted invalid controls rejected, zero audit errors; construction suite 7/7. Exact freeze/raw/audit/commands retained. Earlier pre-allocation resource HOLD remains historical, not repaired execution. No model behavior, evaluation awareness, empirical cue effect or T1 authorization; cue-versus-opportunity-salience separability remains an explicit interpretation constraint. Outcome.
• Selector-reproduction #6746 is now merged, superseding the 15:41 open-PR status. This changes evidence integration only: #6740 A01 import-collision FAIL and #6744 separate A02 synthetic ranking reproduction remain distinct; no real research-productivity/roadmap-utility finding follows.

Latest reconciliation — 2026-10-02 15:41 UTC follow-up

Selected new outcomes and delivery changes; no experiment launched by this indexing task.

• #6740 → #6744: A01 remains FAIL_RUNNER_IMPORT_COLLISION_NO_METHOD_RESULT, method auditor0; a separate posthoc trace auditor validated only the exception chain. Fresh namespace-safe A02 candidate/auditor each ran once in separate pinned OrbStack arm64 containers, exits0/errors=[]; cheapest and nominal entropy selected A, robust decision reversal selected B (minimum scenario score0.60); null UNRANKABLE and frozen entropy reversal matched the host predecessor. Local analysis-index equivalent20 steps/114 tests passed; PR #6746 is OPEN/unmerged, hosted checks not replaced by local results. Synthetic selector reproducibility only, no real roadmap utility. A01 failure unchanged.

• #6691: A01 remains STOP_AUDITOR_CONTAINER_LAUNCH (candidate1, completed audit0). Separate A02 auditor-only allocation copied immutable A01 source/raw/result, candidate0/auditor1/retry0; errors=[], 37/37 rows per case. At B=off, equal-effect control pooled/standardized5.0/5.0; version-interaction challenge3.2/5.0. Structural-zero cell excluded, missing/unknown unresolved, safety separate. A02 receipt and PR #6745 label PASS_AUDIT_ONLY_SCOPED; integration comment reports retained PASS_METHOD_SCOPED, both limited to this saved synthetic artifact, not new-data replication or empirical causal attribution. PR #6745 merged; unrelated #6500 checkout job was cancelled, not PASS. A01 STOP unchanged.

• #5274/#5282 preservation-only PR #6741 is now merged. This changes archive delivery status only; historical duplicate-scope STOP and missing #5282 historical raw/process/audit evidence remain as previously recorded.

Latest reconciliation — 2026-10-02 15:25 UTC follow-up

Selected outcome/provenance changes; previous snapshots remain historical. This indexing task launched no experiment.

• #6616 — Invocation-lifecycle classifier successor, not a human-reliance finding. Merged #6743 preserves A01 STOP_CONTAINER_SOURCE_MOUNT_EMPTY (candidate container exited before input, auditor zero/no retry) and distinct A02 HOLD_ALLOCATION_ID_MISMATCH. A02 candidate/auditor each ran once in separate pinned OrbStack containers, both exit 0; fixture auditor reconstructed 5/5 and emitted PASS_METHOD_SCOPED, but input embeds A01 ID while freeze names A02, so overall HOLD takes precedence. Raw/audit unchanged, no repair/rerun. This does not resolve original local T0-01 HOLD_PROVENANCE_PENDING, authenticate historical host receipts, or test people/GUI/model reliance. Integration is not scientific clearance; closeout explicitly does not claim PASS for an adjacent method-contract still in progress. Outcome.
• #6740 → #6744 — Selector container reproduction now has a failed launch-context allocation and a new pending successor. #6744 documents A01 FAIL_RUNNER_IMPORT_COLLISION_NO_METHOD_RESULT: select.py shadowed standard-library select during platform.platform(), candidate ran once but wrote no method result; failed trace is retained, not support for selector efficacy. New A02 prospectively isolates import/search path and keeps frozen selector/auditor bytes, pins the same image, and plans one candidate/one mathematical audit with independently retained outputs. At review time #6744 is open, has no result comments and its latest amendment says no A02 process had started before the base check; no reproduction PASS is yet evidenced. Preserve #6003/#6042 host-only partial evidence and A01 failure separately. This is synthetic reproducibility, not calibrated priors, useful research decisions or broad portability.

Reconciliation history — 2026-10-02 15:09 UTC follow-up

Selected result changes since the previous reconciliation; older snapshots below remain historical. This index-maintenance task ran no experiment.

• #6536 — Bind a short demo excerpt to master/trace evidence and the exact scope of its public claim. Merged #6739 retains synthetic nine-case T0 PASS_METHOD_SCOPED: claim-scoped arm D accepted two honest scoped excerpts and rejected seven unsupported/forged claims; false accepts for clip+caption / clip+master-link / hashed-EDL-only / claim-scoped arms were 7/6/2/0. Candidate and independent raw-only auditor each ran once, retries=0; five frozen corruption controls rejected, focused tests 5/5. Execution was native macOS CPython 3.14.5, not a container replication. This evaluates authored method cases only: no real video/trace, C2PA signature or validator, GUI/game/model, user media or actual task outcome; no claim that real demos are truthful or complete. Result.
• #5841 — T2 independently reconstructs truth labels for the immutable T1 negative-control fixture. Merged #6738 records PASS_FIXTURE_DERIVED_RAW: one raw-only auditor reconstructed 56/56 rows across seven cases, zero errors; five specified mutations rejected, including a label swap that preserves aggregate counts, and six construction tests passed. T1 candidate/original auditor were not rerun or modified. Mutation tests occurred before the one-shot raw audit, a recorded sequence deviation. This strengthens support for those exact synthetic labels, not detection of every bias, production scoring, real-route results or #57 completion. Prior T0 exploratory overlap, T1 SHARED_PATH_HOLD and out-of-scope undetected sentinel/all-attempt silent bias, and separate 0/0 collision STOP remain unchanged; audit qualification does not erase those boundaries. T2 receipt.

Selected proposal/preservation follow-up — 2026-10-02 15:11 UTC:

• #6740, open — Reproduce #6003 decision-reversal selector in pinned OrbStack arm64 containers. New cross-runtime allocation proposes one candidate and one independent checker, metadata-only scope-label changes with algorithm/FREEZE unchanged, and exact predecessor ranking/control agreement (cheapest/entropy A, robust reversal B, null UNRANKABLE). No result comment or outcome is posted at review time; protocol wording is not execution evidence. Preserve #6003/#6042's host-only partial result and unmet container boundary. Synthetic authored probabilities do not establish calibrated priors, research productivity or useful roadmap decisions.
• #5274/#5282 — Correct the overly broad missing-evidence lineage claim below. Open/unmerged Draft #6741 reports static reconstruction of the canonical #5274/#5283 current-main six-part archive: 68,036 bytes/44 members with original REQUESTS/RESPONSES hashes. Thus absence must not be attributed indiscriminately to canonical #5274. Distinct generic vj01 remains STOP_DUPLICATE_SCOPE_BEFORE_FORMAL (zero invocations); #5282's consumed namespace allocation still lacks raw/process/audit/construction packet and remains HOLD_RECOVERY_INCOMPLETE despite historical owner-reported PASS. The rescue preserves 14 source files, not that missing result packet, and ran no research code. Archive/source identity is not a scientific replay; no old STOP/HOLD or historical report is rewritten.
• #5752 — A04 STOP preservation only. Open/unmerged #6742 retains twelve exact original files: STOP_INSUFFICIENT_DISK_SPACE, candidate/CUDA/auditor 0/0/0, NOT_EVALUATED. Static JSON/AST/index checks are not experiment execution; an existing unittest import failed for missing torch before test bodies, a validation limitation rather than a failed CUDA hypothesis. A05/#5979 is separate; its published-byte HOLD_INTEGRITY and unequal complete-output timing endpoints are not repaired by A04 archiving.

Reconciliation history — 2026-10-02 14:52 UTC follow-up

Selected updated Issue/PR evidence; earlier snapshots below remain historical, not current status. No experiment was run by this index-maintenance task.

• #6530, open — Typed temporal-effect identity: OrbStack A01 now METHOD_PASS_SCOPED, integrated by merged #6737. The frozen synthetic eight-case New York/TZDB 2026c fixture ran candidate 1/1 and separate raw-only auditor 1/1, both exit 0; audit PASS, eight rows, errors=[], retries=0, VM/pulled output hashes matched. Typed checks accepted ordinary and explicit fold choices, returned AMBIGUOUS_UNRESOLVED with both instants for unspecified fold intent, classified the spring gap, and rejected the planted fixed-UTC recurrence accepted by display-string and offset-only baselines; no-save/duplicate controls returned NO_EFFECT/DUPLICATE_EFFECT. A malformed-digest auditor container-create command failed before any scientific process and is retained, not erased or counted as a scientific retry. The older merged #6666 WSLc pre-start STOP and formal counts 0/0/0 remain unchanged. This replaces A01's earlier unrun status only: synthetic schema/method discrimination, not a real saved-calendar/GUI result or reliability claim. Formal result and integration receipts.
• #6539, open — Fault-exposure proposal training remains unrun; preparation expanded. Open/unmerged Draft #6697 now carries a frozen candidate, independent auditor and protocol rather than only the earlier paired-twin draft. Five equal-budget arms, three fixed seeds, 256 training examples/40 epochs per fitted arm and 2,242 disjoint held-out episodes per arm/seed are planned, not measured. Local construction 21/21, failed construction audits and later UNCERTAIN_NO_DECISION_MARGIN are retained/excluded from formal inference. Allocation remains PENDING_UNASSIGNED; formal candidate/auditor/container/retry 0/0/0/0. Missing effect-confirmation evidence requires YIELD; the post-action oracle records outcome, not a multi-step recovery trajectory. Generic visual perturbation is explicitly not a #2526 replication. Paired-twin identifiability/source-review evidence does not establish trained-policy usefulness, safety or transfer. Latest preparation receipt.
• #5887 — Preservation PR #6736 is now merged, superseding the older open-PR status below. The finite-trace temporal-preservation idea still has NOT_EVALUATED, no raw and candidate/auditor/container 0/0/0; source drift/storage/runtime STOP reasons are unchanged. AST/index checks and integration preserve prepared source, not a temporal-semantic experiment or method PASS. Archive.

Reconciliation history — 2026-10-02 14:35 UTC follow-up

• #6613, open — Service-debt/deadline A01 formal outcome: FAIL_HYPOTHESIS, now retained in merged #6735. Candidate and independent auditor each ran once; PASS_RAW_REPLAY reconstructs 288/288 rows with zero errors, not hypothesis success. Across 32 holdout seeds × three strata × three policies, asymmetric mean per-trace worst wait was debt 33.28125 ticks versus FIFO 31.875 and shortest-service 32.28125: debt failed to improve over both. On-time completion floors passed all strata (debt versus weaker baseline 117/109, 129/121, 103/102). This formal result replaces the earlier “unrun” status only for this separate A01; predecessor STOP_PREREGISTRATION_HASH_MISMATCH, excluded construction failures and #6347 results remain unchanged. Synthetic scheduler-method evidence is not human/live-GUI fairness, authority or product benefit; no retry/tuning followed.

• #5841, open — Same-cohort negative controls for route-comparison bias. Merged #5872, not closed-unmerged #5858, is the durable T0/T1 source: T0 six cases/48 assignments formalizes an overlapping exploratory probe, not independent replication; T1 seven cases/56 assignments detects shared export-path drift with SHARED_PATH_HOLD. A primary-only classifier fault changes primary rate 0.50→0.75 while sentinel/all-attempt signals stay silent: OUT_OF_SCOPE_UNDETECTED, not a clean all-clear. Independent raw audits and omitted-case/false-all-clear mutation controls pass, but authored adjudication is not an operational truth signal. Separately merged #6734 preserves eight files from a parallel-collision STOP, candidate/auditor 0/0 and NOT_EVALUATED; this is neither replication nor another method PASS. Existing scientific limits remain unchanged.

• #5887, open — Predicate-relative finite-trace preservation checker. Proposed comparison distinguishes property-preserving stuttering from deletion of warnings/edges/order/deadline evidence, returning UNKNOWN for missing valuations/timestamps/coverage. No method result exists: source/main drift and storage below the frozen reserve stopped before execution; later zero storage, unresponsive Docker inventory and permission-gated service start remain environment records, not temporal-semantic findings. Candidate/auditor/container 0/0/0, raw absent, NOT_EVALUATED. Preservation #6736 is open/unmerged; AST/index checks did not execute either experimental entrypoint. The old STOP is not retried or superseded by packaging.

• Preservation integration status only: #6503's #6733 and #5836's #6732 are now merged. #6503 remains STOP_PREFORMAL_SCORER_GATE_MISSING (formal 0/0, no vigilance result); #5836 remains pre-candidate resource STOP / NOT_EVALUATED (0/0, no trusted-path result). Integration does not upgrade either scientific status; earlier open-PR wording below is timestamped history.

Reconciliation history — 2026-10-02 14:24 UTC follow-up

• #6503, open — Time-on-task vigilance during quiet human supervision. Compare continuous monitoring with matched evidence-triggered/fixed review checkpoints, keeping mandatory machine alerts unchanged. No human study or vigilance result exists. Preformal review found opportunity/display-row auditing but no hit/miss/false-alarm/no-response scorer: STOP_PREFORMAL_SCORER_GATE_MISSING, before formal freeze, candidate/auditor 0/0, not a formal method FAIL. Six construction tests passed; initially wrong CI execution contexts and their corrected passes were disclosed separately. Evidence-delivery #6733 is open/unmerged. Scripted response scoring would still not be human data.

• #6530, open — Temporal-effect identity across DST folds/gaps and local recurrence. Preserve merged #6666's unchanged pre-container WSLc access STOP, formal candidate/auditor/container 0/0/0. A separately preregistered OrbStack A01 copies its source, eight-case synthetic fixture and independent UTC truth byte-for-byte to a fresh package; cached digest-pinned Python 3.12.15 arm64/TZDB 2026c preflight and 5/5 construction tests are not formal evidence. At the latest posted freeze, formal candidate was not invoked. Intended comparison remains typed persisted-effect identity versus display-string/offset-only checks; no live calendar, GUI or reliability benefit is established.

• #3446, still open — Evidence-bound adapter routing, then a version-validated competence-location map. Merged #4291: allocation 02 PASS_SKILL_ROUTER_ADAPTER_SELECTION_SCOPED, 48/48 authored rows correct, zero wrong routes/YIELD violations/authority grants, 288 independent checks and 12/12 mutation rejection; confidence-only violates 30/48 directed cases. Allocation 01 remains STOP_FORMAL_EVIDENCE_WRITE_PERMISSION with no retained scientific rows. Distinct merged #5879: PASS_COMPETENCE_LOCATION_MAP_CONSTRUCTION_SCOPED, 48/48 known locations versus 16/48 for each control, all eight novel cases YIELD, audit 168 rows and 4/4 mutations. Map lookup proxy cost is higher (1,064 versus 72 static/160 recent-success), not a latency win. Latest recovery reports frozen source preserved on main; it changes no result. Neither probe establishes learned calibration, real adapter effects, production readiness or routing authority; prior storage STOP remains retained.

Reconciliation history — 2026-10-02 14:13 UTC follow-up

• #6613, open long-horizon service fairness: fresh service-debt/deadline A01 preregistration targets variable service durations and on-time completion beyond #6347 delay-swap/boundary fixtures. Proposed 32 held-out seeds × three strata × three policies, formal unrun. Construction seeds reconstruct 36/36 rows but fail wait-gain and burst/recovery on-time-floor gates; the excluded hand-authored trace also has debt worst wait eight versus baseline seven. These negative development observations do not decide the held-out hypothesis and authorize no tuning. Preserve earlier STOP_PREREGISTRATION_HASH_MISMATCH; no human, GUI or fairness efficacy claim.

• #6576, timer-support A02 in merged #6728: PASS_SUPPORT_RULE_SCOPED / AUDIT_VALID, candidate and independent auditor once each exit 0. Frozen cutoff twenty distinct tail values adds holds to continuous 0/21, q=0.25 7/24 and q=1.0 30/30 baseline-eligible synthetic fixtures. This tests one cutoff, not a general production threshold or tail calibration.

• #6576, distinct A03 in merged #6731: fifty seeds per arm on four quantization levels; smallest tested qualifying cutoff eight holds q=1.0 48/50 while adding no holds to continuous 0/31 or q=0.25 0/36. Cutoff twenty instead holds 14/36 q=0.25; q=0.5 has only 2/50 baseline-eligible cases. First candidate/auditor each once, AUDIT_VALID; an additional isolated CI candidate/auditor rerun is explicitly disclosed and separate from the immutable first-run record, not absent execution or pooled replication. Generator/grid/sample-size only: no EVT fit, calibrated p99, actual timer/release endpoint, safety or worst-case bound. Formal six-case T0 remains uninvoked; A01 counterexample and A06/A07 STOPs unchanged.

• #6558, open shared-referent idea, merged #6729: distinct allocation-02 PASS_METHOD_SCOPED, candidate and independent auditor once in separate pinned offline OrbStack containers, exit 0; nine cases × four policies, zero errors. Scoped restatement seven correct / zero wrong / two UNKNOWN with three re-asks; raw coordinates six correct / three wrong, fresh rebind and clarification-without-regrounding each five/four. Stable/zoom-only cases do not re-ask. Candidate cannot see hidden oracle; allocation-01 Docker CLI exit-125 before container creation remains STOP. Synthetic protocol only, no human/GUI grounding or T1 authority.

• #6600, open faded-demonstration learning idea, merged #6730: PASS_METHOD_SCOPED over nine authored protocol cases, independent audit zero errors, candidate/auditor once each exit 0. Only one optional teaching step eligible with verified effect, freshness, release, reversibility and explicit opt-in; uncertain/false-success/opt-out cases do not become demonstrations. Five construction/mutation tests; no participants, delayed learning/retention, GUI/model or implemented teaching feature. Human hypothesis remains untested and T1 separately governed.

• #6616, open history-conditioned reliance idea: latest handoff requests HOLD_PROVENANCE_PENDING for local T0-01. Pre-run freeze says zero invocations while result records claim candidate/auditor once; contemporaneous lifecycle receipts are not reconciled. File mtimes alone do not establish creation chronology or invalidate a normal immutable pre-run freeze, so record the reported uncertainty rather than infer a scientific failure or PASS. Preserve all bytes; no participant result or independently verified run follows from the handoff.

• #59 is open/reopened after an erroneous completed closure. Live-03 telemetry-only measurement/score agreement in #6725 has no MAP01 exit/episode finish or positive useful events; neither explicit exit alternative is satisfied. Reopening restores workflow consistency without changing raw outcomes or authorizing a fresh allocation.

• #6723 is now closed completed after #6727 merged; T7's one-case synthetic composition gain and ineffective requested CPU-cap caveat remain unchanged. Parent #5370 roadmap is not complete. Open #6732 separately preserves #5836 pre-candidate resource STOP, candidate/auditor 0/0, NOT_EVALUATED; preservation tests are not trusted-path or human evidence.

• #6693, open migration-cost idea: current task inspection finds no usable Docker CLI/Engine; an empty residual directory is not proof of historical installation/removal. Candidate/auditor/containers zero and prior STOP/HOLD unchanged; no Dockerless cost comparison. Closed-unmerged #5806 was superseded by merged #5821 preserving the same #5797 portfolio T0, not rerun or relabeled. #4817/#5791 stale-ref reconciliation likewise reports no change to already integrated outcomes; no branch operations are performed by this index.

Reconciliation history — 2026-10-02 13:30 UTC follow-up

• #6500, open origin-to-effect binding idea, merged #6718: local PASS_METHOD_SCOPED over 24 metadata rows, candidate/auditor once, zero errors, six mutations rejected. Origin-bound has 0/6 false allows and 2/2 legitimate accepts; visual-only/similarity-alarm each 6/6 false allows. Truthful complete metadata is stipulated, not measured from an OS/GUI. Recipient-side correction separately retains HOLD_PUBLICATION_FIDELITY: two manifest-declared logs absent from the source tree; producer-local digests are not recovered bytes. Raw-only replay matches the published audit without candidate rerun. Merge does not resolve missing-log fidelity or establish spoofing/security efficacy.
• #6576, separate optimizer A08 in merged #6715: six fresh fixtures, independently audited default R/Python parameter parity 6/6 and R optimizer sensitivity 4/6. Consistent with, not proof of, optimizer-path explanation for A05. Preserve A06 output-permission STOP and A07 classification STOP; no retry, formal six-case T0 or physical-release conclusion.
• #6576, distinct timer-quantization A01 in merged #6726: COUNTEREXAMPLE_A01 / AUDIT_VALID, candidate/auditor once exit 0, retries 0. Coarse q=1.0 is ELIGIBLE_REFERENCE with 333 tail exceedances but only six distinct values; q=0.25 has 357/21. Continuous control fails a separate block-median ratio gate (1.545 >1.5), limiting causal comparison. Exposes an omitted resolution/tie-support eligibility condition in a finite synthetic fixture, not a fitted tail, p99 calibration, host timer characterization, key-up or safety bound.
• #6723, open T7 composition successor under #5370, open #6727: PASS_COMPOSITION_METHOD_SCOPED, candidate/auditor once, 36 reconstructed rows, zero errors, nine corruptions rejected. Inversion-subset deadline misses 7 deadline-only, 6 composed bounded inheritance, 1 unsafe unbounded control. One-case synthetic gain only; requested CPU quota was not effectively enforced. Historical T6 STOP_PROTOCOL_DEVIATION unchanged; no parent-roadmap completion.
• #6680, open Draft #6724: retained v2/v3 temporal constructions now adjudicated FAIL_PREREGISTERED_GATE (ineligible fixed-arm probes / nondiscriminating mixed-case probe), despite their raw auditor labels. Distinct v4 twenty rows passes construction temporal audit and posthoc gates, suite 28/28. Same-author synthetic effect oracle; formal candidate/auditor/container/retry still 0/0/0/0. Neither construction PASS nor hosted publication checks establish empirical causal benefit.
• #6558, open purpose-bound shared referent idea: allocation-01 terminal STOP retained; fresh allocation-02 fixes mount syntax and excludes hidden oracle from candidate inputs, with independent auditor receiving it separately. Five construction tests pass, formal candidate/auditor 0/0 at latest preregistration. No referent-validity result yet.
• #6716, closed duplicate, updateability certificates for action-bearing GUI summaries: proposal moved as a refinement to open #5368. Require intended realizable effect and no forbidden collateral effect across every state compatible with the summary; otherwise disambiguate or UNKNOWN. No distinct experiment/result; existing four-trace T0 unchanged. Database view-update analogy is a test-design transfer, not GUI safety evidence.
• #6645, open counterexample-qualified guard refinement: owner reports distinct WSLc T0-01 PASS_METHOD_SCOPED, eleven cases / zero violations, candidate and independent auditor once each exit 0; original guard three harmful admits, refined zero harmful/three safe, exact blacklist two harmful across six held-out states. No rerun or change to merged #6661 host HOLD. Exact local packet is not committed/published: C: storage has zero free bytes and Git index-lock creation failed. Treat this as outcome notification, not GitHub artifact verification or live skill safety.
• #5274, open typed verifier reducer: delivery audit finds the claimed distinct formal REPORT and exact REQUESTS/RESPONSES raw absent from both retained branch and main. Frozen source/tests alone are not published formal evidence; #5477 exploratory detector T0 is separate. No rescue-as-result, experiment replacement or branch deletion was performed. Exact historical packet recovery remains unresolved.
• #5846, open guest/host path-mapped runner: recovery search across 41 readable local Git object databases and GitHub found neither referenced unpushed construction commit nor branch. Useful proposed fix remains unrecoverable from attached sources; #5831 Draft and #5795 STOP unchanged, no reconstructed replacement or formal run.
• #5309, WSLc migration audit STOP_ENVIRONMENT_ACCESS_DENIED before invocation under the sandbox identity, not a user-account WSLc defect or scientific failure. Candidate/WSLc auditor/containers zero. Separate host read-only exact-Git-blob audit reproduces retained T6 432 rows and failed-probe wrong-target counts TASK 27 / YIELD 0; PASS_READ_ONLY_GIT_BLOB_AUDIT is corroboration, not WSLc portability or timing evidence. #6693 likewise has checkout/runtime preflight only, comparative iteration-cost run unstarted.
• #5516, open path equivalence: translation-validation review consolidates under existing route/bisimulation and coalescing work. Proposed refinement validates exact transformation stages and their aligned intermediate/end-to-end contracts; individually valid stages do not automatically compose. No new experiment or live equivalence result. #5309's decision-relevant effect-witness preservation control likewise remains a proposal, not a new efficacy result.
• #59, open real-time control, merged #6725: archives live-03 GitHub-hosted Ubuntu measurement-integration/terminal-score agreement PASS, thirty-two files and reported source/checksum matches. Episode did not finish/exit MAP01 and had zero positive useful events; no recovery-versus-coast efficacy. Candidate and two audits each once, retries 0; packaging reran nothing. Distinct from separately authorized WSLc route, instrumentation only.
• #5730 remains closed not-planned. Merged #6722 restores thirteen original A01 STOP files; declared stderr sizes/hashes do not match retained CRLF bytes (52/2278 versus 54/2280), explicitly unresolved, not hash-verified stderr. Merged #6713 separately preserves #5737 FAIL_CONSTRUCTION_CLEANUP_GAP: digest-mismatch STOP leaves tampered output visible. Old #5746/#5737 closed-unmerged history stays; archive is not GPU evidence, repair or PASS.
• Evidence-only integrations: merged #6720 preserves #4817 STOP_DIAGNOSTIC_COMPLETE_NO_FORMAL_RUN (45/45 finite differences, balanced task lacks competence, formal fits zero, invalid positive-only record still invalid); merged #6714 rescues #6679 HOLD_APPLICABILITY_UNRESOLVED under #6669, no pressure/cap result; merged #6719 preserves #5776 pre-candidate main-advance STOP; merged #6711 preserves #5665 STOP_SOURCE_HASH_PREFLIGHT. All distinct later allocations/results remain separate; no reruns. Closed-unmerged #5787/#5788 retain their synthetic recovery-sentinel FAIL and ledger/resource STOP histories, not superseding successful claims.

Reconciliation history — 2026-10-02 12:32 UTC follow-up

• #6710, closed completed, controller-law replay successor to #6650/#6670. Merged #6717 retains allocation-01 pre-input auditor STOP and distinct allocation-02 exact 36 replay matches / 9 mutation rejections in pinned OrbStack. This is finite audit conformance, not policy usefulness; original FAIL_HYPOTHESIS (six versus three stale optional requests) and #6673 serialization correction stay unchanged. No simulator rerun or threshold tuning.

• #6680, open safe discriminating observation idea: new temporal construction retains source/receive times, input lease/release and a separate post-recovery effect oracle, addressing the earlier label-only repeated-residual proxy. Host tests 18/18, twenty authored rows, raw-only construction audit PASS_METHOD_SCOPED / zero errors. Two identifiable cases give diagnosis zero wrong/zero recurrence/two useful effects; reset/reobserve each one wrong/one recurrence, YIELD zero wrong/zero useful. Formal candidate/auditor/container remain 0/0/0; these authored oracle counts are construction only, not the hypothesis result. Earlier proxy/raw retained.

• #6565, open source-bound user-constraint preview idea: merged #6712 retains PASS_METHOD_SCOPED for authored-card factual/provenance parity only. Candidate and independent auditor once each in separate pinned OrbStack containers, both exit 0, twelve unique rows, zero errors, all six mutations rejected, retries 0. Preference revision stays unscorable. No participants, rendering, comprehension, approval, burden, preference or causal presentation effect was measured; neither H-pass nor T1 authorization follows.

• #6576, distinct A07 STOP_AUDITOR_CLASSIFICATION_MISMATCH after preserved A06 STOP_OUTPUT_PERMISSIONS: two candidate arms each once exit 0, six rows each; sole auditor emits AUDIT_INVALID (payload exit 2), retries 0. Auditor incorrectly required all exploratory starts to converge rather than comparing converged fits under the frozen rule. Candidate-only convergence/NLL observations remain explicitly unaudited; no formal optimizer-sensitivity or numerical-parity conclusion. A05 FAIL and earlier A02 scoped method result unchanged. A07 source.

• Delivery clarification only: #6706 is now merged, retaining the distinct #6689 allocation-01 FAIL_METHOD adjudication and original raw audit PASS without repair or retry. #6709 is merged; it corrects #6695 README status text only, not frozen gates or the twelve-trace scoped outcome. #6333 is closed unmerged: #6315 temporal-coalescing packet reports a retained first audit defect followed by corrected audit-only PASS_METHOD_SCOPED (14 rows, 4/4 mutations) on the same raw; candidate not rerun. Preserve that audit chronology and synthetic-only limits; closure is neither main integration nor failure of the scientific idea.

• #6501, open scope-typed singleflight idea: share concurrent read-only verification only across matching target/generation/provenance scope. Correction to the earlier intake: a stopped duplicate intake (workspace not a checkout / WSLc unavailable) did not establish that the idea was untested. The owner reports retained T0-01 STOP_OUTPUT_SERIALIZATION and a separate T0b-01 PASS_METHOD_SCOPED, ten synthetic cases, zero violations, candidate/auditor once, retries 0, construction 24/24 in WSLc. They inspected the retained main artifacts and stopped duplicate work before execution. This index records that evidence report, not an independent raw re-audit; no real demand, latency, GUI, runtime or product result. Correction source.

• #5686, open surrogate-endpoint validation idea; merged #5736 retains separate T2 allocation-02 PASS_SELECTION_FEEDBACK_GATE_SCOPED with candidate/auditor exit 0 and byte-matching local re-audit. Allocation-01 NOT_EVALUATED remains separate. Authored selection-overfit, coverage/endpoint and non-compensable safety gates do not empirically validate a real interface metric; later ref archival is not a rerun.

• #4000, currently open despite merged #4018: retrospective retained pause-coverage v2 PASS_PAUSE_COVERAGE_COUNTEREXAMPLE_SCOPED, 12 cases / 3,600 timing tuples. Equal 300-sample counts hide 20/40 ms interwake gaps and empty 2 ms bins: catch-up cardinality is not uninterrupted observation. First procfs metadata STOP, separate repaired v2 and no-pause outlier retained. Fresh publication raw audit is not a timing rerun. Linux execution container without Docker identity; no actual visual miss, release-safety, model or human-tempo claim.

• Current formal outcomes for #6695/#6707, #6691/#6708 and both distinct #6689 allocations are in their entries below. The #5407 archival STOP PR #6703 is now merged, without changing STOP_BASE_ADVANCED_BEFORE_RUN.

• #6689 now has FAIL_AUDIT / STOP_CLASSIFICATION_COUNT_CONTRACT in merged #6704; the current idea entry below records the actual executed failure, not an unrun status or method PASS.

• #5407, open heterogeneous task-allocation idea: scarcity-first T2b hypothesis FAIL on 50 synthetic seeds, safe value 1,794 versus value-first 1,981 (shortfall +187), cost 116.5%; both constrained policies have zero unsafe admissions, unconstrained control 236/250 unsafe. Raw audit PASS is distinct from hypothesis FAIL. Closed-unmerged #5723/#5727 have their eight original evidence blobs preserved on main through merged #5903, with reported 8/8 matches per tip. Host CPU construction only; frozen-byte mismatch qualification remains and container reproduction is unrun. Merged #6703 separately archives the exact T2 allocation-01 STOP_BASE_ADVANCED_BEFORE_RUN, runner/auditor 0/0 and no raw; archive is not a new experiment or PASS/FAIL.

• Public Python input-dependency guard, open #6705: opt-in source checking before new presses with scope/freshness revalidation and refusal-safe release. Two fresh Calc cases retain normal persisted 731/864 and wrong 731/863 stopped before Save with empty cells. 412 protocol + 205 harness checks and 22 semantic mutations pass; mid-chord visual mutation is inert-control evidence, not live testing. RETAIN_EXPLICIT_SCOPED_PYTHON_INPUT_GUARD; HOLD_EFFICIENCY_GAIN / HOLD_GENERAL_READINESS. Trusted synchronous callbacks lack hard interruption, sandbox or atomic visual emission; no matched latency/cost or human-tempo result. Reported joint usage 6,782,867 input (6,606,208 cached subset) / 30,481 output covers preparation through audit, not free publication.

• Delivery status only: #6698, #6700, #6701, #6704, #6681, #6677 and #6702 are now merged; their scoped outcomes and predecessor failures are unchanged. In particular, merging #6704 preserves FAIL_AUDIT / STOP_CLASSIFICATION_COUNT_CONTRACT, and merging #6681 preserves source-only FAIL_APPLICABILITY and installed-runtime/pressure HOLD; neither is success or Issue completion. #6539 remains training-unrun; its root import failure and zero-test sparse-checkout discovery are harness errors, not PASS; direct restored-path invocation ran 4/4. #6691 has a finite version-intervention preregistration, not an executed result.

Reconciliation history — 2026-10-02 12:00 UTC follow-up

• #6684, additional open non-draft #6701: allocation-06 PASS_METHOD_SCOPED over nine cases / 98,414 concrete states, false admissions 0, candidate 1 / independently exhaustive auditor 1 / retries 0. Three exactly safe common-mode cases have BOX refusals 3 versus RELATIONAL 0; independent-error controls show no advantage. This is local arm64 CPU execution without a container, not the separate #6696 WSLc STOP/INFRA package or GUI/runtime/authority evidence. Prior source-head/hash STOPs, allocation-04 stop and allocation-05 preformal KeyError remain distinct; local construction 11/11 was checked after formal execution. Synthetic finite geometry only, no product benefit. #6696 was closed as duplicate after separate #6701 merged; no WSLc STOP is retroactively changed.
• #5882, A12 evidence now retained in merged #6702, replacing the earlier PR-pending status. The inspected PR includes report, frozen protocol, raw/audit outputs, logs and evidence manifest: PASS_CPU_PREFERRED_NO_GPU_CROSSOVER_SCOPED; at batch 1,024, CPU 125.0 ns versus CUDA 920.4 ns median end-to-end per row. No tested batch achieves the frozen 20% improvement. Candidate/auditor 1/1, retries 0; audit zero errors and five corruption controls rejected. Host launcher continued after a non-terminating missing candidate-exit-receipt read error; the report says direct exit 0 and one result were observed before the sole auditor, and receipts were later recorded. Preserve that wrapper defect, cgroup/swap enforcement caveat and A11 STOP. PR evidence inspection is not an independent re-execution or checksum audit; no general GPU, model, GUI or Docker-removal conclusion.

Reconciliation history — 2026-10-02 11:57 UTC follow-up

• A12 outcome report in #5085, transfer-inclusive CPU/CUDA supervisor comparison: the owner reports candidate 1 / independent CPU raw auditor 1, both exit 0, retries 0; PASS_CPU_AUDIT_SCOPED, zero errors and five corruption controls rejected. Six batch sizes with 30 alternating pairs each retain hint/admission parity and no unsafe admissions. Reported PASS_CPU_PREFERRED_NO_GPU_CROSSOVER_SCOPED: CUDA still 7.37 times slower per row at batch 1,024, no transfer-inclusive crossover. This is the tested allocation, not general GPU inferiority. Raw evidence/host auditor-launcher note are being collected into a PR; this index has inspected the outcome comment, not independently verified those raw artifacts. Effective memory enforcement is unqualified. A11 and earlier STOPs remain separate and unchanged; no allocation, resource-management or experiment action is made here.
• #6698 is now open non-draft, with unchanged PASS_METHOD_SCOPED scope. #6689 refreshed its base-main freeze without changing source bytes; candidate/auditor remain unrun in the latest prelaunch comment. These are workflow/provenance changes, not new efficacy results.

Reconciliation history — 2026-10-02 11:52 UTC follow-up

• #6539, open fault-exposure training idea; open Draft #6697 delivers only a paired-twin finite identifiability check, bounded prior-art inspection and non-executable protocol draft. Identical policy-visible bytes with different hidden authority/effect truth cannot support both authorized completion and safe non-YIELD behavior. Host tests finally pass 4/4; first-draft oracle error (3/4) and correction remain visible. Completion, unsafe raw proposals and unnecessary YIELD are separate outcomes. Training/candidate/scoring/audit containers remain 0; no model/GPU/GUI or efficacy/transfer claim. Preserve #2526 STOP_SOURCE_MANIFEST_UNAVAILABLE; synthetic data is not its replacement.
• Intake outcomes: #2084 records a generic lost-acknowledgment GUI retry successor screened but not opened because of overlap; physical transient-event multiplicity is an unvalidated possible gap, not a new finding. #6604 similarly records event-trigger timing/delay and generic mode-change ideas as not opened, retaining them as prior-art context for existing topics. Neither note is an experiment or a GUI transfer result; #6604 remains HOLD_NO_ELIGIBLE_TRACE.
• Current outcome changes for #6678/#6700, #6655/#6686/#6698, and the preformal boundaries of #6680/#6689 are reflected in their entries below. All cited Issues remain open; PRs #6697/#6698 are Draft and #6700 is open non-draft at review time.

Reconciliation history — 2026-10-02 11:40 UTC heartbeat

Idea/outcome indexing only. Earlier entries are dated history of distinct allocations, not evidence that a later successor was run; this index performs no experiments or resource actions.

• #6664, open observable quiescence for control handoff / merged #6692: finite protocol METHOD_PASS_SCOPED, 1,191 schedules/3,573 checks, independent audit zero errors/seven corruptions rejected. Accounted policy has 128 horizon certificates, zero false certificates and 1,063 horizon UNKNOWN; epoch-only/delay policies have 3,445/2,254 false certificates. Candidate/auditor once; no actual GUI/backend drain, effect completion or human takeover-safety evidence.
• #6684, open successor to #5577/#18: preserve common-mode target/action coordinate relations with DBM versus independent BOX bounds and exact enumeration, without replacing semantic identity/freshness/authority. Closed-unmerged #6696 records 11 authored construction rows/5 tests and raw-only checker PASS_METHOD_SCOPED. Formal WSLc STOP/INFRA before candidate, no image identity; host construction is not GUI/action or formal runtime evidence. #5577 outcomes unchanged. #6696 was closed as duplicate after separate #6701 merged; no WSLc STOP is retroactively changed.
• #6678, open Blackwell observation-channel idea; merged #6700 now retains METHOD_PASS_SCOPED: construction 5/5, one OrbStack candidate and one independent read-only-input auditor, exit 0 / zero errors, retries 0. Exact rational garbling checks cover 16 ordered pairs of four channels over three states; equal-marginal-entropy partitions are incomparable and task-loss preferences reverse (2/3 versus 1). Only this frozen synthetic fixture is established; no actual GUI/model, safety, latency or user benefit. Requested isolation/limits are documented, effective per-container cgroup enforcement not independently observed.
• #6680, open residual-cause ambiguity idea: one safe discriminating observation versus fixed reobserve/reset/YIELD; mixed/unidentifiable causes remain UNKNOWN/YIELD. Construction 5/5 and 20 retained synthetic rows only. On two identifiable cases, diagnosis gives 0 wrong / 2 useful actions, reobserve and reset each 1 wrong, YIELD 0 wrong / 0 useful; all rows release at tick 0. The metric audit found repeated_residual duplicates wrong_recovery and is not a recurrence endpoint. A root-level import failure is retained alongside the supported passing invocation. Formal candidate/auditor/container/retries remain 0/0/0/0; hypothesis and real recovery benefit untested.
• #6689, open prefix-stability certificate idea: every legal future continuation must preserve disposition; stability does not grant truth/freshness/authority. Merged #6704 retains formal FAIL_AUDIT / STOP_CLASSIFICATION_COUNT_CONTRACT: 40 worlds, 2,592 legal interleavings, 4,707 prefixes; candidate 1 exit 0, raw-only auditor 1 exit 2, retries 0. Per-prefix fields matched and four corruption controls were rejected, but classification_counts mixed disposition classes with EARLY_STABLE_* metrics. No post-freeze repair/rerun or method PASS; local four passing construction tests are separate. Any correction requires separate frozen evidence, not rewriting this failure. Distinct merged #6706 allocation PREFIX-STABILITY-6689-T0-20261002-01 has candidate 769 rows / auditor 768 states, both exit 0, retries 0. Original audit PASS_METHOD_SCOPED is preserved but final posthoc preregistration adjudication is FAIL_METHOD: 192 STABLE_FINAL_FAIL states drop incomplete mandatory-check-vector obligations. No fix/rerun; separate packet/raw hashes from #6704, not pooled or a rescued PASS.
• #6691, open version-defined attribution idea; merged #6708 retains STOP_AUDITOR_CONTAINER_LAUNCH: one candidate exit 0, sole OrbStack auditor launch exit 125 from mistyped bind path before Python. No method audit, retry, host substitute or causal finding. Five local construction tests do not repair the STOP. Preformal control clarification restricts primary equal-effect/version contrast to B=off common support; B=on is joint-support diagnostic excluding infeasible A2×B2 without imputation. No old experiment is alleged biased.
• #6693, open, unrun migration-cost successor: paired same-host Docker/WSLc replay with exact architecture/image/dependency/raw/audit parity, separate cold/warm latency and memory attribution. Proposed >=10% improvement with other endpoint <=5% regression; portability alone is not cost benefit. No measurement, host pressure, Docker removal or broad replacement finding; #6389 distinct.
• #6695, open reversibility-horizon idea; merged #6707 retains PASS_SCOPED for 12 unique authored FSM traces: native WSLc candidate 1 and raw-only auditor 1, both exit 0 / zero errors, retries 0. Wrong commit 1/1→0/1 only with informative signal and a correction route; waiting for signal 2 helps only where that signal adds information. Uninformative/no-route controls show no benefit; deadline misses 0/12. Construction 6/6 is separate. Finite fixture counts, not natural rates, real GUI reversibility or broad #5428 option-policy validation; compensation is not reversal, effective cgroup/swap caps unqualified.
• #6669, open WSLc control-plane survivability question: open #6679 HOLD_APPLICABILITY_UNRESOLVED (Arch distro cgroup snapshot is not WSLc utility-VM evidence); separate merged #6681 FAIL_APPLICABILITY for inspected source path only—WSLC_ROOT_INIT does not use ordinary-distro isolation hook. Installed-binary source binding and pressure hypothesis remain HOLD. Zero workloads/pressure blocks; no cap/OOM/survivability or general installed-runtime defect established.
• #6668, open audit-only successor to #6640 / merged #6677: PASS_ARCHIVE_RECONSTRUCTION on immutable 8,192 rows/32 seeds. Posthoc Ochiai/count/rate MRR = 0.96875/0.93750/0.93229 versus matched 1.0; old FAIL_METHOD/0.10 threshold unchanged. Existing four metrics reconstructed within 1e-12; script hash was recorded after exploratory execution. No predecessor rerun, causal finding or fresh formal comparison.
• #6590: merged #6663 T1 v1 HOLD_AUDIT_INTEGRITY—candidate omitted W1 update, auditor cannot reconstruct. Separate merged #6682 corrected v2 H_FAIL_SCOPED: independently reconstructed 30,720 predictions/10 models, random/block ACCEPT 1218/2560 vs 1119/2560, gap 0.038672 below frozen 0.20. Candidate/auditor once, zero retries. Rejects that magnitude only in the frozen synthetic envelope, not equivalence/general absence of spatial effects. All prelaunch STOPs and v1 HOLD preserved.
• #6350 / merged #6685: allocation03 candidate retained, first auditor exit1 lacked writable /out mount (STOP, no retry). Distinct audit-only04 reads exact raw and PASS_METHOD_SCOPED on 10 tasks/30 records, four controls rejected; candidate not rerun. Finite contract only, no live useful feedback/efficiency. #6410 open Draft #6676 has WSLc construction 5/5, formal auditor still zero/gated; old environment-path STOP unchanged.
• #6530 / merged #6666: access-denied STOP before container, formal candidate/auditor 0/0. Five host tests/eight construction rows only, not calendar/TZ outcome. #6645/#6553/#6576 packages #6661/#6659/#6662 are now merged without expanding their construction/synthetic scope; #6222 rescue #6658 merged, HOLD_ARTIFACT_HASH_MISMATCH remains.
• #5348 merged #6683: 25-cut/six-control first audit reconstruction FAIL, post-run correction not frozen auditor, no audit02 receipt; five correction construction tests error on undefined false, no promotion. #3442 merged #6688: 390-row candidate, auditor FAIL_OR_HOLD on operation_index, scientific HOLD; rejecting ten mutations does not rescue accepted-row failures. Original outputs untouched.
• #5537 merged #6665: T7 STOP_AUDIT_MISMATCH, only 5/6 mutations rejected; false_status no-op, 135 rows, runner/freeze absent from packet. #5156 merged #6687: A07 STOP_RESOURCE_COORDINATION_BEFORE_CANDIDATE, no execution/result. #2437 merged #6672: A04 source archive NOT RUN/NOT EVALUATED, no formal raw/audit. Archives are not scientific PASS.
• #6315 merged #6674: 22 exact missing originals restored without rerun, first apparent PASS remains HOLD_AUDIT_IMPLEMENTATION, audit-only correction distinct. WSLc successor is portability only, not a rewrite or iteration speedup.
• #57 / merged #6671: second-domain Calc composition persists 731/864 in both normal routes; wrong 731/863 controls stop before Save with persisted cells empty (correct refusal, not completion). Equal command/image/capture counts, one case per condition, HOLD_EFFICIENCY_GAIN/HOLD_GENERAL_READINESS. All development/failure/modal/usage retained; overlapping usage windows are not additive or per-route cost. Trusted local recipe only; full goal remains open.
• #3700 / merged #6114: one app cue-wait versus immediate-feedback pair uses 4 versus 5 public requests but input reply 3169.757 versus 85.756 ms; review error/correction retained, no general latency/efficiency conclusion. WSLc replay two tests PASS is not GUI performance/parity. #5256: closed #6101 read-only-after-STOP API proposal is not on main; deleted stacked base/ref-head mismatch prevents treating it as adopted.

Reconciliation history — 2026-10-02 10:39 UTC heartbeat

GitHub-reported evidence only: this index runs no experiments. Earlier entries preserve distinct allocations; a new result does not rewrite its predecessor.

• #6645, open counterexample-qualified skill-guard refinement: refine only independently reproducible guard misses within a declared observable predicate vocabulary, invalidate dependent siblings and fallback on ambiguity. Parent #4260 cache/deopt evidence is not guard-synthesis evidence; #5504 hidden-family false passes limit generalization. Merged PR #6661 records host construction only: 11 authored states/6 events, 13 tests, zero audit errors; harmful admits 0/5 but valid admits only 2/5 (others UNKNOWN). Formal WSLc 0/0/0 and RESOURCE_HOLD, no live safety or least-guard/generalization proof.
• #6650, open optional-work persistence throttle / merged #6670: FAIL_HYPOTHESIS on nine synthetic host traces. Sustained-overload stale deliveries fixed/queue-length/age-persistence = 13/3/6; descriptive tier changes 11 versus 16 for queue-length/age. Transient-burst avoidance and mandatory-outcome agreement do not override the adverse primary comparison. Accounting audit passed, but independent full controller-law replay was absent. Merged #6673 adds strict-JSON corrected audit with identical fields; malformed original/hash remain, no rerun or hypothesis repair. No live freshness/safety/session benefit.
• #6655, open incidental-state legacy idea: inherit/reset prior incidental changes while preserving required task effects and restoration accounting. Open Draft #6694 preserves consumed Windows-host candidate/auditor 1/1/0, 224 synthetic rows and raw PASS_METHOD_SCOPED as a protocol deviation: missed pre-run WSLc directive, not conforming formal evidence. Later line-ending normalization leaves fixture.py/FREEZE.json/PREREGISTRATION.md inconsistent with frozen hashes; raw outputs unchanged. No rerun/relabel or real-interface benefit. Separate #6686 / merged #6690 preserves the original FROZEN_NOT_RUN preparation. Merged #6698 adds a separate WSL-native WSLc successor result: PASS_METHOD_SCOPED, construction 9/9, candidate 1 and raw-only auditor 1 (both exit 0), 224 rows / zero audit errors, retries 0. Pinned local image, network none, CPU 1; WSL warned cgroup/swap enforcement unavailable, so effective memory limits are not established. Original Windows-host deviation and hash mismatch remain unchanged and unpooled; neither Issue closes and no real-interface benefit follows.
• #6657, open notification-induced sampling reactivity idea: merged #6721 retains PASS_METHOD_SCOPED over eight audit gates and two rejected mutations. Exact clock-time exposure and independent uniform-epoch expectation both 0.8; scripted notification-conditioned estimates 4/7 and 10/11 differ, while nonreactive control matches silent 4/5. No humans/live GUI; these are authored method values, not human behavior or causal notification effects. Human/T1 work remains separately governed and unauthorized.
• #6219 / open Draft PR #6644: separate WSLc allocation-01 METHOD_FAIL, 36 rows, candidate once/auditor once, no repair/retry. Independent audit finds ORACLE_CONTRACT:case_addition: replacement of an irreversible download with archive was labeled compatible addition without specifying cumulative versus replacement semantics. Eight other matching cases do not rescue the failed gate. Historical ten-scenario PASS in another package remains distinct; no real intent/authentication/effect result.
• #6553 / merged PR #6659: claim-scoped postdominator observation placement T0 PASS_METHOD_SCOPED, seven synthetic cases, zero audit errors, candidate/auditor once. Typed shared check cost 3 versus route-local 6; loop null 6/6; graph-only cheaper cut 1.5 is semantically unsound. Incomplete oracle finds 3 claim paths versus candidate 2, UNKNOWN/HOLD retained. Eight construction tests; no real graph/oracle, runtime benefit, memory-enforcement guarantee or new generic cut algorithm.
• #6576 / merged PR #6662: separate isolated OrbStack/Docker stationary pilot A02, 4,000 train + 4,000 holdout, one candidate/auditor, retries zero, PASS_METHOD_SCOPED/PASS_RAW_ONLY. ELIGIBLE_REFERENCE; gated p99 4.6663 (42/4000), TailID-equivalent 4.5825 (45/4000); both exact 95% intervals contain nominal 1%, no superiority. Not the still-unconsumed six-case formal T0, physical release, hard safety or worst-case evidence; earlier censor-probe/T1 HOLD unchanged.
• #751 / merged PR #6643: finite top-k skill-candidate starvation METHOD_PASS_SCOPED, 8×3 outputs independently reconstructed. Fixed post-filter misses eligible ranks 3/6; six-record prefix recovers rank 6 and returns UNKNOWN at rank 7, found by exhaustive filtering. Every case has at most one ALLOW with k=2, so adaptive early stopping/multiple eligible cards/work savings are untested. Merged #6646 restores exact 13-file bytes after upload line-ending normalization, no logical or raw-semantic change/rerun. No embedding, model, live task or latency benefit.
• #6523 / merged PR #6642: synthetic IME phase/effect method PASS, 12 traces×4 routes=48 rows and six corruptions rejected. Premature-submit excess raw-enter/symbolic/native/phase-aware 5/1/1/0; two WSLc one-shot processes. Five construction tests pass; two separate index regression tests failed and remain disclosed. No actual IME/browser/application behavior or T1 result.
• #6624 / merged PR #6633: third, distinct allocation retains STOP_PROTOCOL_DEVIATION despite raw 12-row METHOD_PASS_SCOPED. Independent pre-freeze review was missing and execution occurred outside promised isolated copy; acceptance remains invalid, no retry. Fifteen original files preserved without revalidation. Neither merged #6631 accepted fixture nor unpublished local FAIL_AUDIT_REVIEW replication is changed or pooled.
• #5156: merged archive #6654 retains allocation-02 STOP_IMAGE_MISSING_PYTHON_XLIB before main/input, no raw/audit. Merged #6660 retains allocation-03 window expiry, inspections/container/scientific rows 0/0/0. Archives rerun nothing; later key-up results remain separate. #5385 merged #6649 retains preformal relinquishment, zero formal invocations; correction withdraws only erroneous expiry reason, source_commit placeholder remains. #5518 merged #6652 preserves STOP_SOURCE_COMMIT_PROVENANCE, not promotion of historical raw/audit labels.
• #6477 / merged PR #6489: accidental host --help invocation consumed candidate budget, STOP_PRE_FORMAL_ONE_SHOT_CANDIDATE_BUDGET_CONSUMED_OUTSIDE_WSLc; 18 host rows diagnostic, formal 0/0, no retry. #6048 merged #6069 retains STOP_PREFLIGHT_RECEIPT_SCHEMA_MISMATCH/NOT_EVALUATED: wrapper once, model/CUDA/generation/auditor zero. Merge preserves failures, not method success.
• #6604 open Draft #6641 T1 prescreen remains HOLD; source/marker controls 5/5, but a schedule-like marker alone yields REVIEW_REQUIRED_NOT_ELIGIBLE and full occupancy joins remain unvalidated. #59 open Draft #6387 has newer 28/28 WSLc construction, candidate/auditor/retries still 0/0/0; no focus-transfer or game result. #2031 #6639 now merged, artifact/scoring/audit-cap HOLD unchanged; rescue report digest is prior-author verification, not fresh qualified-report hash validation.
• #6530, open temporal-effect identity: New York fold/gap/recurrence versus fixed-UTC fixture preregistered, typed intent/oracle and pinned TZDB required; candidate/auditor 0/0, no calendar result. #5380, unverified reporting-gate extension: refuse fixed-n IID violation limits for adaptive, correlated, missing or discovery-enriched rows; distinct from #5446 optional-stopping evidence. No new allocation/result. #6389 access STOP is caller/sandbox-specific, not a general WSL defect or migration finding.

Reconciliation history — 2026-10-02 09:45 UTC

Source-reported outcomes below are scoped to their retained evidence; this index did not execute experiments or rerun audits. Historical preparation/STOP entries remain chronological, not current formal counts for a successor.

• #6640, open matched-run boundary-spectrum idea / merged PR #6656: T1 FAIL_METHOD on 8,192 synthetic rows/32 seeds. Independent audit PASS with zero errors validates reconstruction, not the hypothesis: matched first-fault MRR lift 0.0729167 misses frozen 0.10 gate despite top-2 retrieval 1.0. Candidate/auditor once, retries zero. Ochiai/raw-frequency comparators were not run; empirical T0/T2 remain HOLD_NO_COMPARABLE_SPECTRUM. No causal or real-interface finding; initial proposal remains history.
• #6604 / open Draft PR #6641: retained MAP01 T1 eligibility review is HOLD_NO_ELIGIBLE_TRACE. v38/v39 retain capture/effect/release/cost but no explicit external target/disturbance schedule. No route×timescale inference or T2 reuse; no new execution. This is separate from merged #6635 synthetic T0.
• #6576 / open Draft PR #6607: formal six-case T0 still 0/0. New pre-formal construction independently recomputes GPD/p99 and TailID selection, 24/24 tests; this does not establish pinned-R numerical parity. Separate host-only censor-cap probe A01 used 2,000 train + 2,000 holdout, candidate/auditor once, PASS_METHOD_SCOPED/PASS_RAW_ONLY. Gate returned NOT_ESTIMABLE_CENSORED_ENDPOINT; all holdout scores NOT_ESTIMABLE_CENSORED_HOLDOUT. Train-only naive p99 is not calibrated population evidence; no physical release bound. Earlier preparation counts and T1 HOLD are retained.
• #6590 / merged PR #6637: preformal geometry screen reconstructs four canvases with zero errors; 96×72 is smallest tested feasible (12/10/12/9 sites), 112×84 selected for headroom (19/16/20/16). Larger input requires fresh model/protocol freeze. Zero images/fits/container runs; method geometry is not measured compression/model benefit. Closed-unmerged #6629 is a static three-test freeze-check regression, not promotion of earlier HOLD.
• #6549 / merged PR #6638: analytical privacy/discovery-frontier finite CPU fixture, 24 rows across eight cohorts and three sensitivities; one candidate/auditor, zero reconstruction errors, six tests. Deterministic noise is not a DP guarantee; suppressed/ambiguous discovery stays UNKNOWN. No real telemetry/privacy certification.
• #6262 / merged PR #6349: retained synthetic GPU certificate reproduction reports PASS_CONTAINER_REPRODUCTION_SCOPED over 4,096 subsets, 12 feasible cells and 23 pair projections; separate CPU audit reconstructs 4,096 rows and rejects four coded controls. One candidate, retries zero, WSL/Podman RTX 3080 pinned runtime. Qualification: fabricated_infeasible_edge only flips a wide-certificate flag, not infeasible-projection robustness; construction exit field is blank despite printed OK. Preservation did not reread the 4.78 MB raw or rerun digests. No real skill, GPU necessity/speed or Docker parity claim.
• #2031: closed-unmerged PR #6378 retains 32 Gemma localization request/response cells. Reported 1/8 arm accuracy is invalid because scorer-required explicit abstain Boolean was missing; auditor invocation-cap deviation and first output-path failure remain. Merged PR #6639 rescues 117 files without rerun; 113/114 manifest matches, REPORT.md mismatch with contemporaneous bytes unavailable. Diagnostic HOLD only, not clean PASS/localization usefulness.
• #5385: canonical merged PR #5396 reports PASS_ACTIVE_LEARNING_SCOPED: four synthetic states, 2,801 words through depth four, four mutations rejected. Reported 204-query budget cannot be independently reconstructed because query transcript is absent; five manual traces are a small baseline. Closed-unmerged #5399 is a duplicate supplementary host lane, not an additional accepted milestone or general sample-efficiency result.
• #5360: closed-unmerged #5401 was superseded by #5408; #5408 was administratively closed after merged #6043 preserved all 23 original blobs plus qualification. Historical PASS_T1_SCOPED host fixture, failed/repeated attempts, missing provenance, weak comparison and deferred formal request remain; no new formal execution or scientific promotion. Issue stays open.
• #5272: closed-unmerged #5331 stopped on main advancement; #5335 stopped on corrupted FREEZE.md UTF-8 bytes. Both pre-formal zero-run STOPs, not tested/falsified hypotheses. #5278 closed delivery #5304 is superseded by canonical merged #5303: seven synthetic workloads/51 jobs, short-burst usable completions 8 versus 5, elastic idle cost 40.016% of fixed-large; raw-only audit zero errors/five controls. Scoped deterministic T0, not production autoscaling or physical capacity.
• #3792 / merged PR #4560: six exact frozen formal-02 source files recovered; no reconstruction of missing raw/logs. Harness STOP/partial rows remain, no rerun or completed result. #5593 / merged #6194: retain T0 denominator/audit failure separately from T1 six-tick roster PASS_METHOD_MULTISTATE_ROSTER_SCOPED (six mutations, 11 tests). Archived stdout is summary, commands have placeholder mounts and exception excerpt is partial; historical counts are not new observations and empirical cohort HOLD remains.
• #5156: #6625 is now merged, still only scoped XSync/server-processing brackets. Merged #6260 preserves fake-Xlib six-row construction (eight tests), not formal X11. Original closed #5298 is preserved via merged #5705; original closed #5415 via merged #6072, with archival qualifications, no rerun/promotion.
• #6319, closed-not-planned: allocation/seed collision STOP before CUDA, candidate/auditor 0/0. Fresh continuation is #6329, not falsification or completion of six-worker memory coexistence. Branch housekeeping comments in #4000/#5133/#5084 do not alter scientific results; no branch cleanup is performed by this index.

Open ideas and current status

• #6619, open action-bound visual residual idea: merged PR #6632 retains a separate synthetic WSLc T0. Raw-delta and action-bound residual both detect 5/5 events by deadline; on valid pans, residual false alarms are 0/2 versus raw delta 2/2. Across all 11 non-events they are 8 versus 10; action-agnostic registration has 3 but suppresses external unreceipted scroll. No uniformly dominant route or live threat/game benefit. Candidate/auditor once, zero errors, five controls; FREEZE_CORRECTION metadata deviation and unenforced-memory limitation remain.

• #6624, closed-completed via merged PR #6631: successor to #6610 representation-contrast critique. Six synthetic rows with equal initial renders; flat routes retain literal/raster state, structured routes retain formula dependencies/addressable elements. Structural follow-up stays UNKNOWN on flat routes; pixel-only changes work on both. Separate WSLc candidate/auditor once, zero errors, 7/7 mutations rejected, METHOD_PASS_SCOPED. Illustrative operation costs are not measured effort. A distinct local replication later returned FAIL_AUDIT_REVIEW because its flat drawing retained native elements; its earlier PASS/checks:0 and review-only infrastructure STOP remain unchanged. That failure qualifies the separate replication, not #6631. Preserve predecessor #6614, no real GUI/user benefit.

• #6615, open adoption-inclusive break-even idea: compare cumulative verified work from common first use, including setup, human work, failure and repair. Merged #6647 is separate 72-row host construction only (11 tests), formal 0/0 and resource HOLD. Merged #6648 retains T0 FAIL_METHOD_GATE and T0B FAIL_PROVENANCE_GATE separately from T0C: 9 synthetic scenarios/58 rows, zero reconstruction errors, learning-fixture break-even prefix 5 (95 vs 100 stipulated seconds), setup-dominates has no crossing by K=5. No measured onboarding/adoption or human benefit; #58 unchanged.

• #6616, open and unverified: permute equal-multiset histories of success/false-success/partial/UNKNOWN while holding final typed receipt, task and raw evidence identical; test appropriate accept/inspect/reconcile/takeover behavior, not subjective trust alone. Proposed source-bound history summary cannot override current evidence or authority. Distinct from #5698 one-shot display, #5991 later unaided audit skill and #6503 vigilance; error-order effects are external prior art. T0 validates stimuli/scorers only; no participant study, allocation/run/PR or human-benefit result.

• #6613, open long-horizon service-debt fairness idea: compare waiting/starvation only among independently eligible GUI requests; scheduling priority never grants authority. #6347 now overlaps repeated-opportunity/per-principal endpoints, so residual variable-service/deadline scope needs differentiation. Local T0 allocation ended STOP_PREREGISTRATION_HASH_MISMATCH: candidate hash and freeze/base provenance disagreed before execution; formal candidate/auditor 0/0, construction only. No fairness result or live multi-user benefit; old allocation/source remains unchanged.

• #6611, open and unverified: compare authorized save/export routes with equal initial verified effects, then reopen copies under a frozen later-reader/version matrix. Score task-critical formulas, links, object/accessibility structure, render, follow-up edit, collateral changes and full reopen/repair cost; retain every failed or UNKNOWN attempt before repair. Original hashes and parse/open success are integrity/transport indicators, not proof of future semantic correctness. Distinct from #6468 current-purpose viability, #6610 same-build future edits and #6227 assurance requalification; digital preservation is external prior art. Merged PR #6620 now records synthetic scorer sensitivity only; no real compatibility or long-term durability result, and no override of the user's required format.

• #6610, closed-completed after synthetic T0 only; real-app hypothesis remains unverified: compare routes with equal verified initial artifact effects, then a separately frozen later edit, including initial overhead, follow-up cost/correctness, preserved unaffected content, exact structure/formulas and rendering. Native editable structure is not always better: include no-follow-up and flatten-preferred controls and declare the follow-up distribution. This is distinct from #6468 present-purpose viability and #5819 repair of unsuccessful outcomes; editable-artifact benchmarks are prior art. Merged PR #6614 reports scoped assay sensitivity, recorded below; no real efficiency/user-agency result.

• #5317, open: predictive action-prefix filtering and viability envelopes. Newer unverified refinements distinguish AUTHORITY_CONTAINED from task-hazard status under external dynamics, and test counterexample-guided splitting of a sound coarse transition abstraction. A false block may be refined only after independent concrete checks; omitted delayed effects are a model-soundness failure. The proposal gives no new authority or live survival/safety evidence.

• #857: open, issue-only research question about compiling a frontier-authored bounded procedure into the smallest reusable typed/guarded macro before considering a learned policy; applicability, current-evidence, authority, release and effect gates must remain explicit, otherwise UNKNOWN/YIELD. Its new minimum-description-length selection refinement is unverified and applies only after hard safety/correctness gates. No #857 experiment, implementation or promotion is authorized/claimed.

• #5760: distinguish the deployable-policy effect (all assigned tasks, including refusal/fallback) from descriptive comparisons among tasks that actually used a route. The issue remains open for prospective work; any human comparison would need separate consent/governance and is not authorized here. Its existing evidence and limitations are summarized below; no prior benchmark is alleged to be biased.

• #5693: proposed fail-latch boundary for malformed response envelopes, so a caller catching a parse error cannot send another effectful request. It succeeds only if the second request is rejected locally while valid/refusal/close controls remain intact. The formal candidate/auditor remain unrun; related #5701 stays Draft with audit-contract gaps.

• #6593, open user-relative, agency-preserving benefit idea / merged PR #6651: finite synthetic accounting PASS_METHOD_SCOPED, 7 cases/13 assigned pairs, zero reconstruction errors, 7 mutations. Planted pooled +10s versus within-pseudoparticipant -10s reversal demonstrates estimand sensitivity, not actual human benefit. Offers/refusals, required/forbidden effects, stable configuration and inadequate-support HOLD remain; agency is distinct from time. No participants/AT/live GUI; human study needs separate consent/privacy/accessibility approval.

• #6600: new, unverified idea to offer opt-in faded agent demonstrations so consenting novices may later perform equivalent GUI tasks unaided. The primary proposed outcome is delayed, independently verified safe/correct completion—not speed alone. No human study, teaching feature, or learning benefit is established; any participant work requires separate consent, accessibility, and privacy safeguards.

• #6231: open, unverified idea to detect silent changes in a reusable GUI skill’s actual effects across app revisions, distinguishing same-schema behavior drift from cosmetic changes. Proposed effect-signature probes remain advisory and isolated; no live skill drift or experiment is established. The updated proposal narrows novelty claims in light of SkillGuard/SkillDroid/EvoSkill-GUI prior art, adds those and reactive failure/recompilation as comparators, and specifies a prequential test against hindsight leakage. No allocation/result exists.

• #5309: open dual-control idea—rank already-admissible task actions for both progress and decision-relevant interface identification, with freshness, harm, cost and typed probe status gating authority. T0–T6 are finite toy fixtures, not a deployed controller result. New comments add unverified persistent-identifiability, action-induced state-displacement and censored-observation controls; no allocation/result for these extensions.

• #6604: new, unverified idea to compare bounded motor-control routes across slow, near-cadence, fast-reversal and abrupt-step disturbances using equal safety/effect gates and route×timescale outcomes. Control-theory sensitivity results are only analogy; no runtime frequency-response or real GUI/DOOM result is claimed.

• #5805: open, unverified multi-principal authorization idea for co-owned effects. Merged synthetic T0 records (#5816/#5818) test finite grant/binding tables only; no real owner discovery, consent, or co-owned effect is evidenced. A current authentic direct DENY combined with an otherwise matching delegation remains untested; a separate question of choosing among jointly permitted outcomes is proposal-only.

• #5442: open end-to-end semantic-receipt idea: intermediate dispatch/verifier success may not prove intended target, precondition, effect and goal post-state; keep UNKNOWN/DISPATCH_ONLY distinct. T8 in merged PR #6606 now reports PASS_TARGET_BINDING_SCOPED on four authored fixtures; it rejected 6/6 corruptions, including a true nested receipt.intent.target_id change. The result covers a deterministic simulator with a stipulated observer—not a real app, GUI, task-effect, safety-rate or product result. Preserve T7 HOLD in open PR #6534 and all earlier evidence unchanged.

• #6590: spatial-block qualification of visual target-position gains, successor to #4752. Predecessor data/confounds stay unchanged. The finite no-model T0 is METHOD_PASS in merged PR #6599 (12,800 rows, all nine gates); its bounds remain synthetic method-only, not the T1/model/GUI hypothesis. Merged PR #6605 adds a separate frozen Docker corruption challenge: clean raw still passes, then changing one (0,0) row’s block_id b00→b01 yields overlap count 1 in b00/b01 folds and corrupted raw HOLD_AUDIT_INTEGRITY with one row-reconstruction error (MUTATION_PASS). Candidate/clean auditor/mutation auditor each ran once, retries 0; pinned network-disabled isolated containers and local Analysis Index checks passed. This demonstrates fail-closed corruption handling, not that a dedicated leakage diagnostic alone rejected it, and not T1, trained-model, GUI or spatial-generalization evidence. #4752/#4814 remain untouched; PR #6605 is merged (2026-10-02); #6590 remains open.

• #6586: after a preregistered no-progress condition, test bounded switching to a source-supported obstacle-side route class. The first attempt remains failed in open Draft PR #6596 (FAIL_OR_HOLD, 15 audit errors; abstract tokens missed the corrected planar ambiguity). A distinct corrected finite T0 is recorded in open PR #6602: PASS_METHOD_SCOPED on 18 rows, zero audit errors; at equal six-unit event budget the class-aware arm chose the open lower route while novelty/backtracking repeated two blocked upper routes, and five ambiguity/same-class/geometry/false-cue/single-class controls held. This is authored-fixture method evidence only, not real navigation or GUI/DOOM benefit. Preserve #6596 failure unchanged.

• #6581, open path-width-conditioned continuous control idea: original retained-corpus STOP_DATA remains unchanged. Separate merged PR #6653 T0b reports PASS_METHOD_SCOPED on six disposable Canvas scenarios; identical trusted pointer events/endpoints save only in wide corridor, with variable-width/corner/exit/no-corridor controls. Candidate/auditor once, zero errors/process retries; construction stops and pre-Python auditor-image launch error preserved. Fixture/oracle feasibility only, not Steering-Law, ordinary GUI, human/agent performance, timing or safety.

• #6580: ask whether static versus time-varying transition uncertainty changes no-authority local continuation under partial observation. Robust-POMDP theory is not GUI evidence; see the scoped T0 encoding, decision, and exhaustive-branch results below; the separate T1 question remains unresolved.

• #6576: gate extreme-tail estimates of input-release delay on endpoint identity, censoring, per-mode support, dependence, and held-out calibration. The read-only T1 audit remains HOLD_NO_ENDPOINT_OR_SAMPLE: no per-key release timestamps, common verified neutral endpoint, or independent tail sample. A separate host-only construction A02 corrected the per-mode minimum-count contract and passed 7/7 audit cases plus 9/9 focused tests; this is not formal T0, generated timing data, tail fit or release bound. Deterministic release/watchdog controls remain authoritative; the formal resource request remains unassigned.

• #6575: test interaction of FULL/crop observation transforms with untrusted rendered instructions while preserving legitimate-task visibility. No Agent Interface vulnerability or model susceptibility result is claimed.

Closed ideas and outcomes — reconciled 2026-10-02

The complete closed-idea inventory is preserved verbatim in Closed ideas and outcomes 2026-10-02. It was split from Home revision 203 to prevent save truncation; every historical STOP/HOLD/FAIL and scoped result remains unchanged. Closure is workflow state, not evidence of success. Recent qualifications remain in the Home reconciliation notes and the linked Issue/PR sources.

Latest outcomes and boundaries

• #6576, separate construction successors: A05 FAIL_PARITY_NUMERICAL_MLE on one of six fresh n=400 fixtures despite index/selection/convergence checks; scale/shape/CI deltas exceed frozen tolerances, so no pinned-R numerical equivalence. A06 then STOP_OUTPUT_PERMISSIONS: one R and one Python candidate exited 1 before fit JSON (host output UID/GID 501 versus container 1000), auditor 0, retries 0; no optimizer finding, chmod/chown or repair. Earlier A02 method result and formal six-case T0 remain separate; no physical-release or safety bound. A06 source.

• #6147, safe distinguishing-sequence idea, reopened: merged PR #6167 preserves STOP_CANDIDATE_RUNTIME_ERROR/NOT_EVALUATED, candidate exit 1, auditor uninvoked. Raw remains unverified; process records 11,020 bytes versus committed metadata 11,022. Closure was premature, not a negative finding. No repair/retry or hypothesis result; a separately frozen successor remains needed.

• #6089, open robust observation/hold horizon / merged PR #6627: separate host finite T0b PASS_METHOD_SCOPED, seven cases, six tests/five corruptions, candidate/auditor once and retries zero. k=0 recovers original horizon; optimistic missing/stale observations fail planted controls; cautious arm releases, bounded-loss arm selects horizon 4 at k=1/2 and yields on invalidation/unsupported bounds. Post-run main advancement timing not captured. No measured channel bound, live GUI/DOOM safety or #59 completion.

• #59 / merged PR #6555: posthoc WSLc raw-auditor replay verifies ten source hashes and 96 retained files/6,472,270 bytes with matching historical values; candidate/auditor/retries 0/1/0. Original v27 injected deterministic live composition remains unchanged. Reproducibility only—not natural threat, survival, reliability, efficiency, latency distribution or MAP01 clear.

• #12, open benchmark/oracle contract: unverified within-domain configuration-interaction extension compares constrained pairwise, one-factor and exhaustive suites under identical task effects/reset. Feasible tuple coverage cannot prove higher-order/temporal faults absent, user coverage or accessibility certification. Existing #5330 internal fault interaction result is distinct; no allocation/run/PR or real fault-discovery result.

• #6576 / open Draft #6607: updated formal package preparation now has 21/21 host tests and six frozen synthetic cases; formal candidate/auditor still 0/0, runtime request unassigned. Forty exceedances is a fixture count, not independence proof; Python TailID/R numerical parity and broad stale-index validation remain unresolved. Earlier 17/17 construction and T1 HOLD stay historical, not superseded scientific results.

• #6604: isolated OrbStack allocation-02 completed via merged PR #6635, PASS_METHOD_SCOPED on 14 authored rows independently reconstructed with zero errors; candidate/auditor/retries 1/1/0. Slow drift direct/local error 16/10 versus fast reversal 45/59, pooled 61/69: finite assay sensitivity only, not an actual controller. Original allocation-01 shared-lane HOLD and separate source-hash STOP remain unchanged; configured limits are not proven enforced.

• #6617, open revision-timed provisional-speech cutover idea / merged PR #6626: WSLc finite event T0 PASS_METHOD_SCOPED, 30 traces/10 cases/3 arms, four corruptions rejected, candidate/auditor/retries 1/1/0. Safe arms admitted no provisional/stale input; one stable scripted case reached effect marker in 3 versus 6 logical ticks. Unknown/irreversible stop dispositions and urgent release remained distinct. No ASR/audio, actual intent/authentication, GUI/effect, wall-clock latency or human safety result; memory enforcement not established.

• #6611, still open / merged PR #6620: METHOD_PASS_SCOPED on 16 synthetic route×reader rows, zero audit errors/all five controls true; formula/render loss, missing dependency and open-failure UNKNOWN distinguished from harmless cosmetics, including compact-route-better control. Candidate/auditor once in separate network-disabled WSLc; complete source/raw/audit retained, memory enforcement unproven. Scorer sensitivity only—not real reader compatibility, durable-format ranking or long-term preservation.

• #6354 / open Draft PR #6621: separate host probe-contract successor picks shared vectors with first bit 1, giving explicit A=0/B=1 probes. Eight tests, three seeds, independent raw-only audit PASS_PROBE_CONTRACT_SCOPED/zero errors; overlap counts 104/102/105 are this packet’s accounting, distinct from earlier contradictory-overlap counts. Original dataset/freeze unchanged. No WSLc validation, LoRA training/quality, GPU or formal candidate result.

• #5279, open cold/warm/resident verifier lifecycle idea / merged PR #6622: eight original source/freeze files byte-identical; recovery restored ten files/43,307 bytes and source suite 6/6 in no-network read-only container. Issue-reported PASS_TINY_VERIFIER_RESIDENT_T0_SCOPED remains separate from HOLD_RAW_DELIVERY_UNVERIFIED: formal raw batches, final audit receipt and corruption records are absent, so independent raw re-audit remains impossible. No lifecycle/formal rerun or regrading.

• #5459, open generation-bound lossy-observation recovery / merged PR #6454: historical WSLc T2 FAIL_METHOD retained, 4,096 grid+28 fault rows, no retry. Report/receipts say 8,224 errors while committed audit has 8,218 (4,120 context-label+4,098 decoded-value); invalid-context explanation is not established by frozen guarded comparison. Provenance discrepancy unresolved; apparent paired wins not validated. All 18 original files unchanged, no large-raw revalidation/rerun or real FEC transport result.

• #6195, open delay×gain stability / merged PR #6272: saturation successor FAIL_AUDIT_EXECUTION, candidate once emitted five trajectories; sole auditor exited on undefined F, no audit output/retry. Two construction tests do not validate trajectories; failure record is diagnostic, not full raw stderr. Host scalar fixture only; no controller/physical-plant or task finding.

• #5156, open owner key-up timestamp idea / merged PR #6625: A18 WSLc/Xvfb candidate and raw audit passed scoped server-processing bracket; six construction mutations rejected. Three request→XSync-return brackets 95,829/106,979/116,762 ns (median 106,979); foreign-intent release rejected, cancellation cleanup and neutral teardown retained. Not physical key-up, app usefulness, safety, MAP01 or task benefit; memory enforcement unproven.

• #6590 / open unmerged PR #6623: allocation-02 geometry replay reproduced retained host enumeration and raw-only audit reconstructed 551 sites, zero errors. Block support 6/6/3/3 misses frozen eight-site minimum, so visual-model T1 held before fitting. Allocation-01 pre-container runner STOP remains; feasibility HOLD does not refute model hypothesis.

• #6389 / open unmerged PR #6618: static auditor successor reads frozen workload fields and rejects booleans in integer counters; seven host synthetic tests pass. Original freeze unchanged; four repository Actions passed but do not execute these package tests. HOLD_RESOURCE_CONTENTION and runtime-parity gates remain; no migration candidate/container, performance or memory result.

• #6461, open broker-boundary challenge successor / open Draft PR #6475: T0b preformal main/source-provenance STOP, candidate/auditor zero. WSLc and post-rebase host construction 6/6 only; memory enforcement unproven. Prior #6179 HOLD unchanged; no formal scientific result.

• #5663, open fixture→live transport idea / merged PR #6491: STOP_BEFORE_CANDIDATE_MAIN_ADVANCED, formal candidate/auditor/retries 0/0/0; assigned window released. Nineteen originals unchanged; five listed logs/three pycache entries absent and earlier construction base distinct from final freeze. Six-test host PASS is historical incomplete-receipt construction, not formal transport evidence; prior failures retained.

• #6509, open interruptible claim-scoped verification / merged PR #6521: WSLc runtime probe ran but frozen unittest command was omitted; one-shot rung consumed as protocol STOP. Candidate/auditor not run, no retry or partial-verdict method result.

• #6081, open error-carry directional compilation / merged PR #6091: STOP_METHOD_INVALID_BASELINE. Candidate/raw auditor agree on 672 authored rows, but maximum-dot-product baseline omitted command norm and is not Euclidean-nearest; five refusal/deadline/release/nonlinear controls unexecuted. Agreement validates reconstruction, not baseline or method benefit. Original failures and 12 files retained; no rerun/real actuator transfer.

• #6222, open crossed effect-oracle repeatability: preserve original PR #6237 reported 288-row PASS_METHOD_SCOPED, five mutations and corrected candidate-raw publication as historical claims. Merged rescue #6658 now qualifies package HOLD_ARTIFACT_HASH_MISMATCH: audit.json, both stdout receipts and RUN.json fail their recorded SHA-256 bindings; matching contemporaneous bytes unavailable and line endings do not explain it. Exact original 18 Git blobs retained, no rerun/reconstruction; T1 HOLD_T1_NO_CROSSABLE_PANEL remains. Not integrity-verified scientific PASS or live benchmark validity.

• #6228 / separate merged PR #6264: FAIL_HARNESS_DEADLINE_SCHEDULE_NOT_ADVANCED retained; omitted clock advance invalidated no-cancel case. Pre-deadline acknowledgement skipped planned timed wait/advance; exact-deadline expiry came from explicit check before wait, despite wait_expired label. Eight preparation tests do not rescue schedule. Host harness failure, not Lease defect or physical-release latency; original STOP remains separate.

• #5895, open strict runner-completion semantics successor to #5630 / merged PR #5982: T4 terminal launch STOP, exit 125. Invalid writable mount syntax was rejected before container creation; CID path also lacked a receipts directory. No target case or independent audit ran and output stayed empty. Historical 9/9 preparation tests do not validate execution. New preservation qualification retains missing published inspect/shell-redirection receipts and arm64-versus-canonical-amd64 scope mismatch; no run, scientific PASS, retry or resource renewal follows from merge. Original files, #5899 and predecessors remain unchanged; later allocations are separate, not repair of this consumed T4.

• #6232, open cooperative clock/cancel successor / merged PR #6234: one candidate trace retained, original auditor STOP on incorrect Git-root path, no retry. New source-review qualification corrects the old narrative: auditor parsed raw before failing at first source retrieval, before source verification/timing gates—not before raw loading. Original STOP/evidence unchanged. Separate merged audit-only PR #6239 for #6236 reports PASS_CLOCK_DELAY_EXTENDS_WITHIN_LEASE_SCOPED, one auditor, candidate/executor 0, zero errors/retries; A-minus-C release delay 400.1414 ms after pinned source/raw reconstruction. C timer-to-release is 0.0562 ms, not 56.2 ms. Deterministic cooperative fake trace with exact executor/Lease objects only—not full _run_arm, transport latency, physical input, MAP01/safety/gameplay or runtime efficacy. This successor does not repair original auditor STOP.

• #6610 / merged PR #6614: finite synthetic T0 METHOD_PASS_SCOPED, six authored cases, 35 independent checks and 5/5 corruptions rejected. Declared operation-cost ranking flips from initially flat-favored to structured-favored under an illustrative follow-up mix; no-follow-up/raster-patch controls favor flat. WSLc memory enforcement is not claimed. Exact candidate/audit JSON is retained locally by SHA-256, not published exactly in Git because the connector base64 path did not preserve bytes; published metadata is not raw-byte availability. Issue closed-completed for this scoped assay, not proof of real editability, user preference, efficiency or product benefit.

• #6604, OrbStack-allocation preformal construction refinement: initial labeled case IDs could reveal semantic-swap truth; that original 8/8 raw/audit and comment remain unchanged. Separate final-source construction uses opaque C01–C07 IDs with names/truth confined to scorer oracle; 11/11 tests and 14/14 raw replay rows, zero errors. Planted slow/fast crossover is still authored-fixture evidence; no-crossover pooled direct/local totals 63/31. Formal candidate/auditor/container remain unconsumed, retries zero, HOLD_SHARED_ORBSTACK_LANE_UNASSIGNED. This corrects construction before formal execution, not the separate host-method STOP_SOURCE_HASH or any real controller result.

• #6228, open v6 clock-delay versus cancellation/lease-release successor / merged PR #6229: STOP_HARNESS_CANCEL_NONCOOPERATIVE, candidate 1, auditor 1, retries 0. The fake backend ignored Lease.wait() returning true, so both treatment A and cancel-first control C released at the 2,000-ms lease deadline rather than cancel; B released around 1,500 ms. Five frozen source identities matched, but actual v6 cancellation/release-extension hypothesis remains unresolved. No-external-effect host fake-executor only, not physical input, live MAP01/runtime finding or formal container allocation; merge preserves the typed STOP.

• #6468, open artifact-purpose viability idea: distinguish weighted partial-check progress from indispensable source/formula/render/export dependencies for the declared use. Original open Draft PR #6487 retains main-SHA-change pre-candidate STOP (formal candidate/auditor/retries 0/0/0); 12/12 construction and five rejected mutations are not a formal result. Separate merged PR #6495 reports SUBSUMED_BY_12: one candidate/independent auditor, six cases/twelve routes, five mutations rejected and no decision disagreement between strict TaskContract and dependency ledger. Equal partial document scores hid a citation failure, but the extra ledger added no distinction over the strict contract in this fixture. New assistive-recipient-purpose extension is unverified: visually matching artifacts may require declared reading order/relationships/text alternatives; tag presence is not a usability oracle, and no synthetic result certifies accessibility or format conformance. No real artifact/GUI/model/efficiency result; original STOP unchanged.

• #6173, open run-scoped answer-surface audit idea / open Draft PR #6185: historical PASS_METHOD_FACTORED_AXES covers 17 symbolic cases, with evaluation-target attempt, acquisition/denial and monitor coverage kept separate. Denied attempts remain visible despite coverage gaps; incomplete/misbound evidence stays UNKNOWN. Retained construction 12/12, candidate/raw-only auditor once each, zero errors, host-only. New preservation qualification discloses shared authored facts, closely parallel reconstructions, lack of process-isolation/real-payload/external-invocation evidence and parser limits. All 15 original files and predecessor verdicts remain unchanged; preservation reruns nothing. Not real contamination detection, benchmark cleanliness, production monitoring completeness or roadmap completion; overlapping T0b/T0-02 and earlier failures remain separate.

• #57 is OPEN/reopened after the 2026-10-02 checklist audit: merged #6601 preserves a bounded selection/save comparison, not a final version-pinned integrated bundle, composition/dependency plan or full requirement-to-test matrix. Required cold acquisition, warm reuse, invalidation and bounded repair comparison phases remain incomplete. Efficiency HOLD is retained, but an allowed HOLD label alone does not complete the milestone. All historical #6601/#6612 successes/failures remain unchanged; no repeat of consumed trials, causal savings or new resource authority is implied.

• #6604, separate allocation DISTURBANCE-TIMESCALE-6604-T0-ORBSTACK-20261002-01: host construction only, 8/8 tests, 14 authored rows and raw replay auditor output PASS_METHOD_SCOPED with zero errors. Planted slow drift direct/local error totals 16/10 versus fast reversal 45/59, pooled 61/69; no-crossover controls retained ordering. These test assay sensitivity, not actual controller or task benefit. Formal candidate/auditor/container remain 0/0/0, retries 0, runtime request pending/HOLD; host output is not formal T0 PASS. This allocation is distinct from DISTURBANCE-TIMESCALE-T0-METHOD-20261002-01 and does not repair or retry its retained STOP_SOURCE_HASH below.

• #6198, open historical held-input receipt successor / merged PR #6271: T2 independently reconciled retained decision IDs/steps, requested durations and shared interval bounds against #6175 (v38 11/11, v39 27/27; displayed 0.001-ms precision, aggregates within 1e-9; interrupted-row identity/timing/classification matched). Raw output remains FAIL_PRIOR_TABLE_RECONCILIATION because all 38 completed-row keyset comparisons lacked the prior table's keys field: missing comparator, not demonstrated mismatch. Overall D remains HOLD, not PASS despite merge. No candidate or T1 auditor rerun; prior #6175 unchanged. Offline host JSON audit only, not physical key occupancy, useful feedback, task effect, live safety, causality or product evidence.

• #57 / merged PR #6612: public batch-reference registration composed with bounded same-owner bridge feedback. Preserve the allocated image-presentation failure (old bridge directory after capture-root change) and three unallocated cases; a separately frozen matched-directory repair saved X80/Y50/W40/H30 through both ordinary and guarded-graph routes, while both undertravel controls stopped before Save and preserved the original file. All eight successor images reached the primary; owned processes ended and neutral releases were verified. Four caller commands per case versus historical five is not a matched latency/token result. Joint usage retains construction/failure/diagnosis/repair/all four cases (96 source records, nine primary image blocks; total 4,887,694 tokens including 4,824,320 cached); nested windows overlap and are not independent savings or billing evidence. Public Python API/bridge composition only—not compiled MCP endpoint or transport validation; production defaults unchanged and efficiency HOLD retained.

• #59 / merged PR #6387: primary WSLc/Xvfb pair PASS_OWNER_FOCUS_RELEASE_SCOPED, candidate/auditor/retry 1/1/0. Repeat control saw four A KeyPress events; focus A→B release verified neutral, B subscribed/fenced/drained and received no matching press before release. Merged #6675 discloses an accidental second pair: allocation-wide 2 candidates/2 auditors, duplicate isolated as protocol deviation, not replication or pooled evidence. One private mechanism sample only; no real desktop/threat/gameplay or MAP01 completion, memory enforcement unverified.

• #6604, open disturbance-timescale idea: allocation DISTURBANCE-TIMESCALE-T0-METHOD-20261002-01 ended before trajectory generation with STOP_SOURCE_HASH / pre_run_hash:T0_SPEC.json. A main-SHA update changed the spec but left its earlier hash in the frozen manifest; six rerun construction tests passed, then the sole candidate invocation exited 2 before output-directory creation. Candidate invocations 1 (stopped), auditor 0, retries 0; no candidate packet/raw trajectory, WSLc/container, GPU/model or GUI. Preserved source/manifest were not edited or retried after STOP. This is no method/controller result; the route×disturbance hypothesis remains unverified.

• #6252, open shared-clock/cancel successor under #59 / open unmerged PR #6254: one construction-only invocation exited 1, before candidate/auditor (0/0, retries 0). It expected cancelled whereas the exact Executor-v10 reports completed when a cancellation-observing mock backend returns normally. The failure row was not persisted, so the exact failing predicate and other gate predicates remain unverified; 7/7 auditor mutation tests and syntax compilation are preparation only. Retain this design/preparation STOP and predecessor #6242/#6250 separately; no runner-clock hypothesis, live control, physical release or scientific result follows.

• #5273, open verifier-registry idea: versioned capability/evidence-role/declared-cost contracts support authority-neutral compatibility preflight. Historical v1–v5 #5363, #5379, #5388, #5400, #5416 are closed unmerged as superseded by v6, not formally validated; branches/evidence remain. V6 #5436 is also closed unmerged after merged archival #6072 preserved 12 exact files / 95,588 bytes. Historical PASS_HOST_CONSTRUCTION_ONLY and eight-case raw audit stay qualified by the 19/21 test-count discrepancy, missing execution receipts, reported post-freeze reruns and STOP_NO_EXACT_RESOURCE_LEASE (zero formal invocations). Costs are synthetic estimates; no operational verifier, scheduler, runtime, measured-latency or authority claim. Archive execution was zero; registry work remains unresolved.

• #6230, open oracle-boundary idea: distinguish semantic, observation, execution and joint interface headroom without crediting answer leakage or safety bypass. Open unmerged PR #6279 retains overall HOLD_AUDITOR_INTENT_EQUIVALENCE: its historical raw PASS and seven rejected corruptions did not prove execution-only intent matched the paired actual arm; local model/executed equality was insufficient. Separate merged PR #6282 retains a fresh host-only PASS_METHOD_SCOPED, 28 authored rows, eight mutations rejected including paired intent alteration, one candidate/independent auditor and no retry. Ambiguous cases remain UNDEFINED and leaking oracle arms uncredited. Preservation reruns nothing and does not repair #6279 HOLD; no empirical model headroom, deployability, real GUI/effect or #59 benefit.

• #6219, open: account for authenticated mid-task additions/revisions/retractions after persistent effects, without treating retraction as undo authority. Historical host-only finite T0 reports PASS_METHOD_SCOPED over ten scenarios; one candidate/independent audit, four corruptions detected, all releases empty. Confirmed effects survive retraction, unknown delivery stays unknown, irreversible conflicts are explicitly impossible, and forged amendments retain the old contract. No real authentication/delivery, GUI, compensation safety or task benefit. Updated proposal adds a reversibility-qualified earliest-conflict rollback comparator (external streaming-revision prior art); scripted revisions do not establish preference formation, preview influence or human value. Preserve #3442 fencing result/STOP separately.

• #5325, open capability attenuation/confinement idea / merged archival PR #6036: preserves eight original files / 36,073 bytes from PR #5377, now administratively closed unmerged. Historical PASS_CONSTRUCTION_ONLY covers 55 hand-authored host rows and four rejected corruptions; formal runner/auditor/container remain 0/0/0, CPU request withdrawn. Intake/source-parent mismatch remains qualified, not repaired. No archived execution/rerun or security/backend/authority efficacy claim; later owner toy outcomes are distinct.

• #5346, open scoped expiring-trace coordination idea / merged archival PR #6016: preserves all 11 original files / 86,873 bytes from PR #5365, now administratively closed unmerged. Pre-formal chronology/model-audit STOP remains controlling: out-of-order lease/effect events were not audited by aggregate checks. Historical 15/15 host tests and PASS_READONLY are retained, not promoted or rerun; formal invocations remain zero. Stale prior-execution plan hash and absent original process/audit-output receipts remain disclosed. Separate #5382 T1 STOP/T2 scoped finite-model evidence does not supersede this T0 STOP; no real coordination/safety/performance result.

• #6561 / merged PR #6609: separate additive T0b serializer/raw-auditor successor reports PASS_METHOD_SCOPED, six finite cases, 15/15 corruption controls rejected and zero audit errors. Construction (8 test groups), candidate and independent auditor each ran once and exited 0 in native WSLc 3.0.1.0 with cached pinned Python 3.12.14 and no network. Swap/cgroup warning prevents an effective memory-enforcement claim. Preserve #6568 construction, #6591 T1 counterexample/STOP and #6603 host-only T2 separately; this is not live planner, GUI/gameplay or #59 benefit, and #6561 remains open.

• #6354, open successor to #6321: explicit-role gated online LoRA versus unconditioned same-vector control remains a proposal, not a GPU result. Independent preformal source/data review of merged construction PR #6376 found the declared role-conflict probe has A/B labels 0/0 in all three seeds; the tests/auditor do not require contradictory probe labels. Genuine same-vector conflicts remain in the full deck (54/52/52 per seed), so this does not erase that contrast, but the explicit probe cannot support it. No candidate, CUDA, optimizer or formal auditor ran in this review; preserve allocation-01 and predecessor STOPs unchanged.

• #6143, open / open Draft PR #6153: guard-disclosure proposal-risk adaptation remains unverified in real agents. Historical exact finite T0 planted an aggressive policy shift: guard-on harm 33/512→117/512 (difference 21/128), while conditional harmful-action TPR stayed 3/4; null/protective controls were not flagged. Eight construction tests; one candidate and one separate raw-only auditor, retries 0, method-scoped arithmetic PASS. The post-run zero-active-container/exclusive-resource gate is UNVERIFIED, not cleared by preservation/integration: no unqualified governed formal PASS, rerun, real adaptation or safety claim. The updated idea narrows novelty to truthful advance guard disclosure with the same tools/hard guard, rather than generic tool-enabled safety behavior (external prior art).

• #5317 / merged archival PR #6027: preserves six original files / 34,854 bytes from host-only PR #5336, now administratively closed unmerged. Historical 35-cell raw reported horizon-2 blocking a modeled unsafe prefix and viability filtering a stranded state; 7/7 tests and PASS_READONLY are retained claims, not rerun by the archive. Keep endpoint-only checking, bounded existential recovery, literal-matrix audit and missing execution/provenance receipts visible. The withdrawn container request and STOP/HOLD remain, with original-allocation container invocations zero. Distinct T3 evidence merged through PR #5505 is separate; owner #5317 remains open. Closure completes source delivery only.

• #5518, open: idea to check history-conditioned allowed adapter outputs and shortest forbidden prefixes, including explicit UNKNOWN/quiescence and hidden retry. Historical synthetic logical-tick T7 accepted 3 traces/rejected 3, rejected 3 mutations, and its unchanged replay reached main via PR #5651; ticks and observer semantics were stipulated. A separate attempt retains STOP_SOURCE_COMMIT_PROVENANCE and duplicate/insufficient-novelty disposition (closed unmerged #5538). Merged PR #6608 retains a separate pre-candidate WSLc STOP: its frozen --read-only flag was rejected before container creation, candidate/auditor/retries 0/0/0. Host 5/5 tests and 21-check pipe audit are construction only. No real adapter/GUI conformance or timing result; all earlier outcomes remain separate.

• #5263, open grounding question: merged archival PR #5750 preserves 27 original files / 550,950 bytes, including 14 PNG paths, by exact Git blob identity. Original PR #5300 is now closed unmerged for administrative delivery, with zero inference and HOLD_NO_FORMAL_RUN_RESOURCE_AND_FREEZE unchanged. Comparator correction, Docker coordination deviation and later model-load STOP remain attached; preservation establishes no grounding/latency/backend qualification.

• #59 / merged PR #6212: retains a historical v6 construction replay; five scripts ran once each on Windows Python 3.12.10 and native Ubuntu WSL Python 3.12.3 with matching 3/3, 4/4, 3/3, 3/3 results and 6/6 binding mutations rejected. WSL was not container-isolated; the receipt adds no live/formal v6 run or game, feedback, survival, navigation or model-efficacy result. Merge is historical-record preservation.

• #6576 / open Draft PR #6607: host-only deterministic construction records an eventual shift alarm as protection-inadequate when a safety-deadline miss precedes it; unknown modes are NOT_ESTIMABLE_NEW_MODE. Focused tests passed 17/17, and the retained eligibility-case raw-only audit passed 7/7 with six rejection controls. Earlier direct-script import STOP, positive-infinity acceptance FAIL and comparator candidate-count failure remain retained. Formal T0 remains pending, and the Python TailID port needs numerical comparison against pinned R/ismev. Repository-wide analysis-index validation failed on reported unrelated stale entries. T1 HOLD_NO_ENDPOINT_OR_SAMPLE remains; these constructions provide no release-delay sample, tail estimate, statistical detector characteristics or safety guarantee.

• #6586 / open PR #6602, preserving failed predecessor Draft PR #6596: the first attempt’s independent audit failed (FAIL_OR_HOLD, 15 missing-action-after-observation errors) and abstract route tokens missed the corrected planar-class question. The separate corrected T0 in #6602 reports PASS_METHOD_SCOPED on 18 authored rows, zero audit errors; with six-unit event budget the class-aware arm selected the lower route while novelty/backtracking repeated blocked upper routes, with five ambiguity/same-class/changed-geometry/false-cue/single-class controls. Construction 14/14; candidate and auditor each ran once in separate network-disabled WSLc containers. No real GUI/DOOM navigation benefit; #6596’s failure remains unchanged.

• #6580: WSLc T0 reports PASS_METHOD_SCOPED for finite lifetime encoding only: construction 5/5; independent audit reconstructed 12/12 rows and 56 histories, six negative controls agreed, seven corruption controls rejected, zero audit errors. This encodes FULL/ZERO/EVENT distinctions; it does not compare continuation policies or establish GUI/model lifetime, responsiveness, safety, performance, or task benefit. A separate read-only T1 corpus check is HOLD_MODEL_LIFETIME_UNIDENTIFIED. WSLc warned that swap/cgroup enforcement is unavailable; memory enforcement is not claimed. No retries; merged evidence package is PR #6588.

• #5309 / open Issue: comments T0–T5 retain only hand-authored finite toy results for task-vs-dual-action ranking, expected-harm/freshness/probe-status gates and budget sensitivity. Latest T6 is PASS_T6_INDEPENDENT_FINITE_ORACLE on 432/432 cases, zero authority grants; among 54 FAILED-probe cases per arm, TASK fallback produced 27 wrong-target outcomes while YIELD produced 0. Candidate and independent oracle each ran once in digest-pinned network-disabled Docker; these are synthetic policy outcomes, not calibrated utilities or live GUI effects. The issue remains open. Persistent distinguishability, action-induced state displacement and refused-action observation are newer unverified method controls, not additional findings or allocations.

• #5442 / merged PR #6606, distinct from T7 HOLD in PR #6534: T8 ran candidate and independent raw-only auditor once each in separate pinned-image, network-disabled OrbStack containers (retries 0). Audit PASS_TARGET_BINDING_SCOPED, 4 rows, zero errors; all 6 corruption controls rejected, including actual nested intent-target and source-target mutations. Wrong-target, stale-prestate and no-op fixtures yielded UNKNOWN and denied commit; valid fixture confirmed and admitted. Synthetic deterministic object store/observer only—the object state was not independently sensed from a real application. No live GUI/effect, authority, broad safety, latency/cost, or product claim. T7’s historical HOLD remains unchanged.

• #2441 remains open; merged PR #2445 and PR #2511 preserve scoped fail-closed adapter evidence. In the retained 800-row synthetic table (1,600 adapter outputs), baseline returned success on 24 cases with ambiguous/unsupported evidence, while the corrected mapping had zero adverse-evidence or partial-delivery successes; all 10 evidence mutations were rejected. The dedicated pinned contract workflow passed 8/8; no live dispatch, model, GUI, or network experiment occurred. Unknown future delivery literals remain a follow-up risk; these are adapter-contract results, not transport failure rates.

• #5805 / merged PR #5818: finite synthetic admission-table T0 (68 rows) is preserved without candidate rerun. The 25 direct-grant cases in #5816 and #5818's recipient/epoch binding table are complementary, not pooled; a current authentic direct DENY paired with otherwise matching delegation is still missing, so no complete joint-policy claim. No real co-owned consent/effect or safety result; Issue #5805 stays open, and post-authorization preference choice remains unverified.

• #6590 / merged PR #6599: finite no-model T0 METHOD_PASS, 12,800/12,800 rows independently reconstructed, zero audit errors, all nine gates passed; candidate/auditor each ran once, retries 0. Uniform block rate was 0.80 throughout; planted hotspot b11 was 0.20 versus 0.80 elsewhere, and the position-random diagnostic was 0.64375 (0.44375 gap to worst block). This validates only the synthetic block/split/leakage/empty-support method, not Issue #6590's model-level T1 hypothesis or any real GUI/spatial generalization. Host-only macOS/CPython 3.14.5 was explicitly prescribed by the freeze; no GPU/container. Historical #4752/#4814 outcomes remain untouched.

• #6590 / predecessor #4752: read-only reconstruction joined 2,400 old prediction rows and 1,200 control/treatment images with zero hash mismatches; A remains 0/200 treatment ACCEPT and B 200/200 (control 0/400 overall; 0/600 false ACCEPTs). This is not new inference, byte-level SHA audit, or spatial-dependence evidence. The read-only geometry check found A farther from training support and at the bottom edge while B is interior, so distance and boundary confound locality. At this reconstruction stage, #6590's finite-field candidate/leakage audit was unrun; its later separate synthetic T0 is recorded above. #4752's scoped PASS is unchanged.

• #6580 / T0b decision-level successor: a separate WSLc candidate/auditor run reports PASS_METHOD_SCOPED for the frozen toy decision gate (48/48 rows and 116 histories reconstructed; six negative controls matched; zero unsafe dispatches; seven corruption controls rejected). Six rows continued only with explicit lifetime evidence while the lifetime-unspecified comparator yielded; 36 point-comparator rows were counterfactually unsafe-possible, not dispatched effects. The earlier T0 encoding PASS and T1 HOLD_MODEL_LIFETIME_UNIDENTIFIED remain unchanged. No GUI/model, probability, task-performance, or safety claim; WSLc memory enforcement is not claimed and no stage was retried. The result is retained in merged PR #6592. Its nature-first/public rows use one representative theta rather than enumerating both public branches; this historical coverage gap was addressed by T0c, recorded separately below.

• #6580 / merged PR #6597: T0c PASS_METHOD_SCOPED exhaustively audited 66/66 finite rows (42 nature-first public branches + 24 reactive), six controls, zero errors; construction 5/5 and candidate/auditor each ran once. This closes only T0b's finite branch-coverage gap. WSLc memory enforcement is not established; the separate retained-evidence T1 remains HOLD_MODEL_LIFETIME_UNIDENTIFIED, with no real-interface, safety, or T2 claim. PR #6597 is merged; #6580 remains open and T1 remains unresolved.

• #5760 / merged PR #5772 and PR #5769: PR #5772 retains a 16-row synthetic PASS_METHOD_SCOPED with six corruption controls rejected. Preserve the separate historical T0 audit v1 FAIL_T0_CONTRACT and v2 STOP_AUDIT_V2_CONFIGURATION_MISMATCH; the later v3 PASS_AUDIT_V3_REPAIR_SCOPED was audit-only (candidate/oracle 0/0), not a rerun or rewrite. No empirical random assignment, causal route effect, or evidence of bias in existing comparisons; #5760 remains open for any separate prospective study.

• #6422 / merged PR #6465: A02 host-CPU finite synthetic no-effect boundary result, 8/8 rows independently reconstructed and 5/5 corruption controls rejected; effect authorization false in all rows. Conditional deferral held three A01-style eligibility cases and other invalid/mismatched evidence cases while leaving the valid authorized/source-stated case eligible for a fresh approval request. A preserved report has a free-text “six” typo despite count/reconstruction of eight. This is not full #6422 T0 and does not authorize T1; no model, GUI, effect, WSLc, Docker, or GPU result.

• #6561 / merged PR #6591 and open Draft PR #6603: preserve the T1 target-auditor counterexample—four preregistered corruptions were accepted, falsifying that frozen auditor-integrity claim—and its STOP_POSTRUN_AUDIT_NOT_RETAINED (no retry). Separate host-CPU T2 in #6603 retained control accepted 1/1, five mutations rejected 5/5 and original audit zero errors, but that original audit did not independently recompute mutation decisions. A separate posthoc frozen-rule reconstruction (no candidate/auditor imports or candidate rerun) matched all six packets and decision/reason vectors; its own checker exited FAIL/STOP on an over-specific environment-label assertion, not conflicting execution receipts. Keep this posthoc evidence distinct and do not call the original audit an independent re-execution. The old T1 false accepts remain unchanged. T2 is synthetic packet/auditor-boundary evidence only; no WSLc/container/GPU, live planner effect, GUI/gameplay, or #59 threat-control result. PR #6591 is now merged; #6603 remains an open Draft. #6591 qualification notes audit_probe.py defines main without invoking it, so its documented direct command would not run the probe; the lost-output explanation is incomplete and claimed RESULT.json is absent. No rerun or source repair; #6561/#59 remain open.

• #6575 / merged PR #6582: allocation 01 stopped as STOP_HARNESS_FIXTURE_MISMATCH. Construction passed 10/10; candidate emitted 24 rows; independent audit found six sham-crop pixel-provenance mismatches among 36 image items. No retry or model call; susceptibility remains unevaluated.

• #6556 / merged PR #6574, followed by merged audit-only successor PR #6577: the original finite synthetic METHOD_PASS_SCOPED remains 50 rows, zero audit errors, 50 obligations preserved. #6577 exposed that the original auditor accepted arbitrary GARBAGE as refusal; its enum/oracle hardening passed 6/6 tests, rejected five decision corruptions, and the unchanged 50-row raw still passes. This validates only the bounded auditor contract and does not alter the original result. No live runtime, OS, watcher, GUI, model, or human result.

• #6533 / merged PR #6573: finite PASS_METHOD_SCOPED, 44/44 reconstructed. Qualified-frame matched full-state on 11 traces and reduced aggregate checker-accounted bytes 79.1%, but regressed 2.13× on the 8-cell case. Not GUI completeness, race freedom, safety, or end-to-end performance.

• #6532 / merged PR #6584: allocation 01 is a protocol STOP, not a scientific result. A host help probe happened once before source freeze and outside the required container; formal container candidate/auditor/retries/launches stayed 0/0/0/0. No rerun or clock-truth claim.

• #4580 / merged PR #6583: evidence-restoration audit PASS_EVIDENCE_RESTORATION_SCOPED for 10 original seed archives (9,279,674 bytes), zero errors. No candidate/training/loader/timing rerun; original synthetic robustness result stays narrowly scoped.

• #6561 / open Draft PR #6568: host-only construction suite reports 11/11, but static review found unvalidated guard outcome semantics and a candidate-controlled binding anchor that its auditor does not independently bind. Treat the construction count as such, not as validation of those provenance/safety claims. WSLc/formal run remains 0/0/0; no authority or planner-benefit claim.

• #6206 / open Draft PR #6215 and merged clean evidence successor PR #6598: preserve the finite synthetic PASS_METHOD_SCOPED result (14/14 checks, four corruptions rejected): 33-trace shortest-witness adjacent-pair coverage caught REV→ACT missed by an order-blind 48-trace suite, but an equally sized random draw also caught it, so superiority over random is not shown; a 135-trace length-3 suite caught OBS→REV→ACT missed by the pair suite. No runtime/GUI reliability claim. #6215 remains open Draft and not mergeable as-is; Issue #6206 says do not merge it as-is. #6598 was merged as a separate 13-file additive evidence rescue against current main (12-file package plus generated analysis README), with refreshed index, 4/4 construction tests, index check, and all 12 manifest entries matching Git blob SHA-256. It did not rerun the original candidate/auditor. Auditor denominator tautology remains a caveat; external enumeration corroborated 59/59 presence tuples but is not encoded in that auditor. Preserve the original result and caveat unchanged.

• #857 / merged parent evidence PR #895 for Issue #890: a scoped replay of retained #57 Chromium evidence compiled the field-entry procedure into a typed macro {token, field_handle, submit_handle}. It matched retained layout-A targets (2/2); when the old handle was missing after the A→B change, it yielded before pointer input (0 pointer events), then admitted current B handles after the retained repair without recompilation. PASS_CHROMIUM_GUARDED_MACRO_COMPILATION_SCOPED is limited to this retained workflow; no fresh live task, general compilation, or new economics claim. #857's own MDL refinement remains proposal-only.

• #4817 / predecessor #4814: the preserved branch remains HOLD_ARCHIVE_INTEGRITY—10 manifest hashes mismatch, one JSON file is invalid, and raw outputs needed to replay the audit are absent; no training/audit rerun, PR, merge, or branch deletion. A separate exact no-noise analytic check found identical sets of 26 local 3×3 patterns for the specified positive square and negative distractor, proving non-identifiability only for that noiseless construction; it is not evidence about #4814's noisy formal result, which remains unchanged.

• #5330 / merged PR #6594 (merged): archive 12 supplemental files byte-identically, replay the corrected raw-only audit and reject 4/4 mutations. Candidate not rerun; this does not replace or reconcile the canonical #5330 history, and remains a finite synthetic artifact. Supplemental parent PR #5358 is closed unmerged as an archival snapshot because its study path was independently populated on main; closure is superseded delivery, not a new scientific outcome.

• #1628 / merged PR #4227: preserve STOP_PREFORMAL_SOURCE_IDENTITY because six required exact A2 source files were absent from the reviewed history; formal runs/results 0/0. No code reconstructed from narrative. #1628 remains open; merge is evidence preservation, not a scientific result.

• #2988 / merged PR #5812: synthetic construction PASS distinguishes useful directed detours from coverage-only/wrong-direction/unknown states, with independent mutation checks; two failed fixture attempts remain preserved. No MAP01/game topology extraction, evaluator-leakage prevention, navigation benefit, or formal #2988 result. This scorer construction is related background for #6586, not its path-class experiment.

• #58: current Agent Interface DOOM candidate remains NOT RUN. Two WSLg/Xvfb starts were denied before game initialization; a separate headless alternative attempted a forbidden audio socket and exited by SIGSEGV before any frame/action. No gameplay, matched baseline, or performance result; the launch gate remains open.

• Core Inkscape idea #57 / merged PR #6578, PR #6587, and merged PR #6601: prior canvas-prepared arms showed no count/efficiency advantage. In the newer full select/move/check/Save comparison, compiled normal-route selection first failed (1/2); a separate interior-point/reference-context successor then had both normal routes save X80/Y50/W40/H30 and both undertravel controls stop before Save, preserving the original file. All eight live attempts and evidence are retained. Point and context changed together, so the earlier failure cause remains unknown; selection-cue preflight is not held-out qualification. PR #6601 merged the bounded evidence and local fallback checks (401 protocol +192 harness tests; deterministic replay 2/2); the remote replay remained QUEUED and formal check SKIPPED—not a remote PASS. No production/default/guard change or efficiency result. Issue #57 is now reopened; preserve the earlier administrative closure and all trial results as history, not completion evidence. Five checklist items and broad integration remain unresolved.

• #6389 / open Draft PR #6434: migration start-gate STOP, not a native-vs-WSLc measurement. Candidate arms are 0/0; Python versions differ (native 3.12.3 vs pinned WSLc 3.12.14), and seven active workers made timing attribution unavailable. Synthetic construction tests passed 10/10 in each runtime, but preflights occurred before the latest coordination check; no further run until quiet-host/runtime-parity gates clear. No migration approval or performance/memory result.

• #6210 / open Draft PR #6589: finite T0 was frozen but STOPped before execution because WSLc was unavailable and no authorized runtime substitute existed. Construction passed 5/5 only; candidate/auditor/container 0/0/0, no formal result.

• #6562 / open Draft PR #6569: preserve the one-shot source-hash-before-invocation STOP unchanged. A separate read-only source audit found the runtime exposes point capture timestamps, not a capture-interval contract; the synthetic inclusive-endpoint result does not establish live endpoint semantics. No retry or live-control claim.

• #6410: reopened because the closed T0 was an auditor/method failure, not evidence for or against lifecycle reuse. Audit-only Allocation 02 ended at STOP_AUDIT_ENVIRONMENT_PATH: its sole WSLc auditor invocation used an erroneous extra research/ path prefix and failed before reconstruction; candidate=0, auditor=1, retries=0, no audit report or lifecycle result. Preserve this terminal STOP and do not rerun the candidate. Allocation 03 is a distinct raw-only audit, refrozen to current main; native WSLc is available and an inventory snapshot was empty, but this is not an owner release or coordinator assignment. Candidate/auditor remain 0/0 with no invocation; the resource request is still pending.

• #6181: reopened after closure of finite successor #6184 proved only the stipulated 16-subset sensor-cover method, not real-GUI sensor independence, measured cost, safety, or provenance. The transfer idea remains unresolved.

• #5927: merged PR #6595 records a fresh successor T0 PASS_METHOD_SCOPED: in authored finite controls, intent plus dispatch acknowledgement required one independent persistence receipt to distinguish persisted/modal-blocked worlds, while an already-fresh local typed receipt required zero additional exchanges. Construction 7/7; candidate and separate raw-only auditor ran once each, audit had zero errors, retries 0. This is not a live receipt, GUI/effect, general model-call lower bound, safety, latency, or cost result. Earlier allocations and the four predecessor manifest/raw mismatches remain unchanged and unregraded.

• #5276: a proposed deterministic-selection baseline now states its assumptions explicitly: greedy submodularity applies only to optional set-union coverage after mandatory checks are fixed; complementary checker pairs/common-source failures break the guarantee. No benchmark result is recorded.

• #6532: after the preserved allocation-01 protocol STOP, allocation-02 has a 6/6 host construction check but remains only prepared; formal candidate/auditor/container 0/0/0 pending a separately assigned slot. No timing-receipt result.

• #3352: execution-evidence gate now has an additive 2026-10-02 WSLc clarification: eligible single-container local experiments may use WSLc while retaining target-boundary execution, raw identity/evidence, independent audit, scoped H/T/D/C/U and PR review. A fresh host smoke passed pinned-image no-network execution, read-only Windows bind input, write rejection (EROFS), and cleanup; this establishes availability only. It does not prove effective memory limits, Docker parity/performance, or migration completion: preserve the scoped #6309/#6355 memory-cap failure/audit caveats, and keep #3352 open pending representative workload comparison and review. No previous STOP/result is changed.

• #6507: closed administratively as a duplicate of #3352, not as a completed migration or scientific PASS. Main already prefers eligible WSLc CPU/single-container runs; #3352 owns the remaining runtime-parity decision, while native Ubuntu/WSL2 remains tracked separately under #6389. Preserve the proposal and history; it adds no separate experiment or integration decision.

• #5215 / Draft PR #5228: audit-only successor preserves #5216 as HOLD_LEGACY_PLAN_CONFLICT; host and pinned-container audit suites passed 8/8 and rejected 18 mutations, but resource-governance HOLD remains because container use preceded required assignment. Draft remains unmerged; no live kernel/OS/GUI claim.

Reading the status

• Open does not mean untested; closed does not mean passed.
• A scoped PASS applies only to its declared fixture, method, and assumptions.
• STOP/HOLD/FAIL remain visible and unchanged; evidence-repair or audit successors do not create a new scientific candidate result.
• T0 or construction results do not establish model, GUI, human, production, or end-to-end benefits unless that specific scope was tested.

Sources

• Current curated outcomes ledger
• Full selected Issue inventory
• Research evidence ledger
• Current goal and roadmap define direction, not evidence