Skip to content

Hard gate: downstream closure + reverse-impact revalidation before upstream promotion #90

Closed

Description

@d6g8k5htny-coder

Owner directive — 2026-09-25.

Purpose: prevent an error in a lower/downstream mathematical object from propagating into expensive upstream synthesis.

Effective immediately for the mathematical program:

  1. No theorem/coefficient/asymptotic/synthesis may become controlling unless every required transitive dependency is terminally classified as one of:
    • PROVED_REVIEWED
    • SUPERSEDED_NONBLOCKING
    • REFUTED
    • BLOCKED_ABSENT
  2. Any required BLOCKED_ABSENT dependency forces the dependent object to HOLD.
  3. Any amendment/refutation/domain change/normalization change/source-carrier change to a dependency triggers reverse transitive impact analysis and marks all affected controlling dependents REVALIDATION_REQUIRED until their exact dependency edge is replayed/reviewed.
  4. Green CI, hashes, architectural admission, numerical experiments, same-author review, or navigation success are never mathematical discharge by themselves.
  5. Work proceeds bottom-up on the dependency graph. New upstream work may be explored, but it cannot be promoted while required lower nodes are unresolved.
  6. Existing open campaign items (including CAP/LAW/RATE/PAL/RN review and integration obligations) must be mapped to exact graph nodes and edges; status-only churn does not count as closure.
  7. Implement a machine checker that fails closed on illegal promotion and a reverse-impact report for changed dependencies. Add negative controls proving that a changed lower lemma holds its transitive dependents.
  8. Preserve frozen evidence and predecessor statements; corrections are additive successors.

Acceptance: machine-enforced graph invariant + tests/mutations + dependency closure report. Scientific status must not be promoted merely by implementing this gate.

Owner authorization: Dylan Roy. This is a repository-wide research-integrity control.

Activity

  1. d6g8k5htny-coder commented on Sep 25, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    Implementation clarification from the owner-directed audit, not a status promotion:

    A still-required REFUTED dependency must block its dependents just as a missing/open dependency does. SUPERSEDED_NONBLOCKING is valid only after an explicitly reviewed edge replacement/removal; a label alone does not make a required premise usable. The original issue's terminal classification list is a disposition vocabulary, not an acceptance allowlist.

    Current research-hardening tip 077464e already has claims/graph.json, claims/README.md and FW-UNCONDITIONAL, FW-LM011-PRECONDITION, FW-NO-RECEIPT-PROMOTION, and FW-FLOAT-NOT-CERTIFIED. Reuse these. In particular, preserve frozen-v2.2 versus register-note status columns rather than flattening them. D3-LEMMA-RN-UNIF and its carrier-binding limitations already exist there; default-main code search was not a whole-repository absence test.

    The added reverse-impact component must traverse the UNION of old and new dependency edges, including sub_obligations. Deleting a prerequisite or removing an edge must not erase the dependent from the impact report. Missing nodes, malformed records, and cycles must fail closed. Impact output is a revalidation hold/proposal only, never permission to promote, and the component is not globally deployed until wired to actual research-branch before/after inputs and verified in CI. Vault inspection remains exclusive to the owner-directed auditor under #91; this integration request grants nobody vault access.

  2. d6g8k5htny-coder commented on Sep 25, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    Downstream audit caught a second load-bearing gate defect after PR11 integration, so #90 remains OPEN rather than being prematurely closed.

    PR11 correctly fixed own-node eligibility, but re-reading this issue's clarification exposed two remaining acceptance gaps in the deployed v1: (1) REFUTED was still treated as satisfying a required dependency; your clarification explicitly says a still-required REFUTED premise must block; (2) reverse impact used one graph's current edges, so a deleted dependency edge could erase a dependent from the impact traversal instead of using UNION(old,new).

    Math- draft PR #13 is the additive hotfix. It now distinguishes disposition vocabulary from premise satisfaction: only PROVED_REVIEWED / SUPERSEDED_NONBLOCKING satisfy required edges; required REFUTED and BLOCKED_ABSENT force HOLD. The P15 demand-one counterexample edge is correctly non-required boundary provenance, not a premise of the d>=2 successor. reverse_impact_between(old,new) validates both graphs and traverses union edges; missing/malformed nodes and required cycles fail closed. Test target is31 methods/10 semantic mutants including required-refutation and deleted-edge controls.

    Scientific effect NONE. No live theorem promoted. PR13 remains draft while exact identities/RESULTS are regenerated and actual Python3.11 replay runs. This issue will not close until those controls are green and the hotfix is integrated.

  3. added a commit that references this issue on Sep 25, 2026
  4. d6g8k5htny-coder commented on Sep 25, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    Math-side semantic hotfix is now merged: Math- PR13 → baca69c394ab42130c61771bee74e808703f1ce7. Final cleanup run36168484104 SUCCESS. Source-bound run36168348158 passed32 tests /10 semantic mutants in normal+optimized Python3.11.16 plus177 regressions (209 distinct tests total); artifact10879865394 downloaded/inspected, REPORT passed and sources unchanged.

    The gate now enforces the clarified semantics: own node must be PROVED_REVIEWED; only PROVED_REVIEWED satisfies a still-required premise; required REFUTED/BLOCKED_ABSENT force HOLD; a SUPERSEDED_NONBLOCKING label cannot satisfy a still-required edge unless the edge is actually reviewed/replaced/removed; reverse impact over before/after graphs traverses UNION(old,new) edges; malformed/missing/cyclic required graphs fail closed. P15 demand-one refutation is correctly a non-required scope boundary.

    #90 remains OPEN for one final integration criterion from this issue's clarification: wire these semantics to the actual hardening claims/graph.json / sub_obligations before/after inputs rather than only the Math pilot graph. I assigned that thin adapter to Cursor on #98 comment5836715882, with no status copying and scientific effect NONE. Close #90 only after that adapter + negative controls are green on the real hardening branch.

  5. d6g8k5htny-coder commented on Sep 25, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    Keep this gate issue OPEN after Math- PR13 merge. Fresh read of merged baca69c394ab42130c61771bee74e808703f1ce7 / hard_gate.py blob 17514e490690639450fd35b5c639ecbf0ba977e1 confirms four residual gaps: edge-only changes are not impact seeds; statement/domain/source changes can evade an unchanged manual fingerprint; the changed controlling node itself is omitted; required is not type-checked as Boolean (0 can skip the premise). Detailed review and executable delivery are in Math- PR13 comments 5836643758 and 5836994416.

    A source-pinned loss-only supplement now handles complete node and outgoing-edge records, exact source-binding snapshots, union(old,new) edges, self-hold, malformed/duplicate edges, strict flags, cycles and both graph schemas. Actual 42 tests /9 semantic mutants pass in normal and optimized Python, repeated from downloaded package. Source SHA256 8fa27f5883191274fe37e7a058319d5a069c3d764b4bd0266d912cc6b06f6245; package Drive 1PENg-JmUbaT4UluDCp6F0z_nda0GLrK2 / SHA256 be2af605c65bba2e17adbedc0985cd931f200e5328adcb1af43ff1277467d58e. Integration target: preserve the existing promotion rules, use these exact cases to repair the actual gate, rerun real repository regressions and verify final source identity before closure. These are supplement software tests, not a live Math repository replay, and no mandatory gate deployment or scientific acceptance is claimed here.

    Separate vault #91 progress is 33/80 scoped reviews complete (4 archive +29 navigation),47 open. Its sole-auditor boundary remains unchanged.

  6. added 2 commits that reference this issue on Sep 25, 2026
  7. d6g8k5htny-coder commented on Sep 25, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    I continued #90 after closing the separate Vault 99 forensic review. The four remaining transition holes in merged Math- PR13 are now patched in Math- draft PR #15: complete-node comparison, outgoing-edge-signature change detection, changed-node self-hold, and strict graph/Boolean/duplicate validation. Seven focused tests were added.

    This does NOT close #90 yet. Source pins/RESULTS and the repository's full validation/mutation/regression suite still need regeneration/replay at the final PR source identity. Scientific effect NONE.

  8. added a commit that references this issue on Sep 25, 2026
  9. 10 remaining items

  10. added 4 commits that reference this issue on Sep 25, 2026
  11. d6g8k5htny-coder commented on Sep 25, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    Independent source-first audit (Math- PR45) agrees #90 must remain OPEN. Current Math git_transition_audit.py byte-binds repository-local blob/tree sources but records https:// / external: references as external_unresolved. That is honest, but it means repository-wide promotion enforcement cannot be considered discharged while a controlling external source may be unresolved/unbound. This is the same class of blocker later review comments identified from actual controlling hints.

    Required discharge object remains concrete: structured immutable bindings for every controlling external node, fail-closed unresolved/missing/record-only states at the actual base/head transition, retained source identities + impact/HOLD report, and an end-to-end mutation showing a second-source byte change seeds revalidation. Engineering green tests alone do not close this.

  12. added a commit that references this issue on Sep 25, 2026
  13. d6g8k5htny-coder commented on Sep 25, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    Landing-claim integrity is now deployed in Math main via merge 232a182fee316d2bb428b7351b5c5caae3b8cc2e. It blocks a positive landing disposition when a required dependency is marked UNRESOLVED_EXTERNAL, pins every advertised local source/review by commit+blob, checks current blob identity, and remotely verifies the advertised live queue is OPEN. Run36202860493 is green.

    This strengthens but does not close #90. The repository-wide transition gate still needs structured immutable bindings for controlling external sources at actual base/head transitions, including the second-source mutation control already requested. Landing-manifest correctness and promotion-firewall completeness remain separate objects.

  14. added 5 commits that reference this issue on Sep 26, 2026
  15. d6g8k5htny-coder commented on Sep 26, 2026

    @d6g8k5htny-coder
    OwnerAuthor

    ENGINEERING-GATE CLOSURE — exact integration evidence is now complete. PR98 reviewed head b59359ebb972351802c7bf343363811a8e83150b passed hosted CI run36206280233/job108303405337 with event-compare transition_ok:true, transition_errors:[], unresolved_controlling_sources:[], promotion_permission:false, and full pytest 3414 passed, 2 skipped, 3 subtests passed. Q0 remains visible in controlling_impacted together with its explicit coverage_repairs migration; D1 remains CONDITIONAL/HOLD with unsatisfied premises visible. Navigation and withdrawal-governance workflows also passed. OpenAI final integration review ACCEPT is comment5841860158.

    Cursor then merged exactly that reviewed head into the configured hardening base with expected-head protection: merge ebedb7802024fa557e9071e4c9cec7cddc474b89, parents base 1ae02b9ab759fe5edb31ad04f2a47992a5dee2c5 and reviewed head b59359e...; reviewed-head→merge diff is files:[] and the trees are identical. Adapter bytes are unchanged from the reviewed head.

    Closing #90 at engineering-gate scope only: the loss-only source-bound base→head revalidation mechanism is integrated on the hardening branch. This is NOT theorem acceptance, NOT permission to promote scientific claims, NOT closure of D0–D6, and NOT a merge of the hardening branch to repository main. promotion_permission:false remains structural. Any future gate/source-binding change reopens review through ordinary reverse-impact controls.

  16. added a commit that references this issue on Sep 27, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions