linear_github_mapping.json

{
  "schema": "coordpy.linear_github_mapping.v1",
  "team": "CoordPy",
  "project": "CoordPy",
  "linear_workspace_url": "https://linear.app/coordpy/",
  "github_repo": "adotdong29/CoordPy",
  "comment": "Canonical mapping of milestone → (commits, docs, Linear issues). Update this file at end-of-milestone; run scripts/sync_linear_github_v1.py to validate and emit Linear-ready summaries. Adding a milestone entry: append to milestones array, list commits + result/runbook docs + Linear issue IDs.",
  "milestones": [
    {
      "id": "W89",
      "title": "HumanEval-70B-Reflexion-K=5 same-budget multi-agent superiority retirement",
      "outcome": "RETIRED — first confirmed same-budget multi-agent superiority retirement (+5.56 pp B over A1, 2-of-3 per-seed majority)",
      "commits": [],
      "docs": [
        "docs/RESULTS_W89_HUMANEVAL_REFLEXION_V2.md"
      ],
      "linear_issues": [],
      "carry_forwards_retired": [
        "W86-L-HUMANEVAL-V1-A1-SAME-BUDGET-NOT-BEATEN"
      ],
      "carry_forwards_added": [
        "W89-L-HUMANEVAL-REFLEXION-V2-HUMANEVAL-K5-SCALE-CAP"
      ]
    },
    {
      "id": "W93",
      "title": "Preflight-first empirical superiority wave (5-gate harness; 3 candidate kills; no expensive bench)",
      "outcome": "DISCIPLINE — preflight infrastructure landed; 3 candidates killed in cheap evidence; no NIM spend",
      "commits": [
        "b79cd82"
      ],
      "docs": [
        "docs/RUNBOOK_W93.md",
        "docs/RESULTS_W93_PREFLIGHT_DISCIPLINE_V1.md",
        "docs/W93_FAILURE_DIAGNOSIS.md"
      ],
      "linear_issues": [
        "COO-13"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": []
    },
    {
      "id": "W94",
      "title": "Preflight-earned K=10 pilot killed; cross-modal pivot to MathVista documented",
      "outcome": "KILLED — K=10 pilot ceiling-saturated A1 to 100% on 15-problem slice; 3-of-6 pre-committed gates failed; HumanEval-Visual K=5 retired as cross-modal battlefield; MathVista selected as W95 line",
      "commits": [
        "9355b2f"
      ],
      "docs": [
        "docs/RUNBOOK_W94.md",
        "docs/RESULTS_W94_K10_PILOT_V1.md",
        "docs/W94_CROSS_MODAL_BATTLEFIELD_SCOUTING.md"
      ],
      "linear_issues": [
        "COO-7",
        "COO-8",
        "COO-10"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W94-L-K10-PILOT-CEILING-SATURATION-CAP"
      ]
    },
    {
      "id": "W95",
      "title": "MathVista cross-modal: preflight + Phase 2 pilot (9/9 gates PASS; +10pp) + Phase 3 retirement bench (5/6 W88 bars PASS; margin bar misses by 1.33 pp at +3.67 pp; NOT retirement)",
      "outcome": "PHASE 3 NARROWLY FAILS RETIREMENT — A0=30.33% / A1=67.67% / B=71.33%; B-A1=+3.67pp (vs +5pp bar); B-A0=+41pp; per-seed B-A1=[+6,+10,-5]; B>A1 on 2/3 seeds; B≥A1 on 267/300 problems (89%); 13/14 audit re-derivations PASS offline. Single-seed +10pp pilot narrowed to +3.67pp at multi-seed scale; preflight-first discipline validated for the third time.",
      "commits": [
        "c14a8a1",
        "ab022ce",
        "1ebcdb5",
        "90c7aca",
        "70321d4",
        "17ff0dc"
      ],
      "docs": [
        "docs/RUNBOOK_W95.md",
        "docs/RESULTS_W95_MATHVISTA_PREFLIGHT_V1.md",
        "docs/RESULTS_W95_MATHVISTA_PILOT_V1.md",
        "docs/RESULTS_W95_MATHVISTA_PHASE3_V1.md",
        "docs/LINEAR_GITHUB_SYNC.md",
        "linear_github_mapping.json"
      ],
      "linear_issues": [
        "COO-11",
        "COO-16"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W95-L-MATHVISTA-PILOT-SINGLE-SEED-CAP",
        "W95-L-MATHVISTA-RETIREMENT-MARGIN-CAP"
      ]
    },
    {
      "id": "W96-A",
      "title": "MathVista at 90B-Vision: preflight + Phase 2 cheap pilot (9/9 gates PASS; +10pp, byte-equiv to 11B) + Phase 3 retirement bench (3/6 W88 bars FAIL; B-A1 = -5.00pp; scaling 11B->90B HURTS the team; NOT retirement)",
      "outcome": "PHASE 3 DECISIVE NEGATIVE — A0=28.00% / A1=71.33% / B=66.33%; B-A1=-5.00pp (vs +5pp bar); B-A0=+38.33pp; per-seed B-A1=[+4,-7,-12]; B>A1 on 1/3 seeds; B>=A1 on 255/300 problems (85%); 11/14 audit re-derivations PASS offline; only 3 retirement-margin bars FAIL by sign/margin. Cross-scale shift on B-A1 = -8.67pp (11B +3.67pp -> 90B -5.00pp); H2-saturation empirically confirmed at multi-seed retirement scale; scaling the VLM weight class does NOT retire the cross-modal carry-forward.",
      "commits": [
        "8b5db54",
        "1a2522d",
        "b5778af",
        "41f64e2",
        "07ea9ed"
      ],
      "docs": [
        "docs/RUNBOOK_W96A.md",
        "docs/RESULTS_W96A_MATHVISTA_90B_PILOT_V1.md",
        "docs/RESULTS_W96A_MATHVISTA_90B_PHASE3_V1.md"
      ],
      "linear_issues": [
        "COO-17"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W96-L-MATHVISTA-90B-PILOT-SINGLE-SEED-CAP",
        "W96-L-MATHVISTA-90B-RETIREMENT-MARGIN-CAP"
      ]
    },
    {
      "id": "W96-C",
      "title": "MathVista C1 (VLM-Verifier-Final-Turn) cross-scale Phase 2: 11B FAIL (+0pp; verifier 0/11) + 90B PASS (+13.33pp; verifier 1/7 = mechanism NOT load-bearing); NOT Phase 3 entitled",
      "outcome": "CROSS-SCALE AMBIGUOUS — 11B Phase 2: A0=36.67% / A1=63.33% / B_v2=63.33%; B_v2-A1=+0.00pp; verifier rescue rate 0/11=0.0%; 2/9 gates FAIL; V2 LOST 13.34pp on B arm vs V1. 90B Phase 2: A0=36.67% / A1=66.67% / B_v2=80.00%; B_v2-A1=+13.33pp; verifier rescue rate 1/7=14.3%; 9/9 gates PASS but verifier mechanism not load-bearing (1 of 5 B-only rescues; rest from text-only chain sampling variance vs A1 K=5). Cross-scale collapse pattern (one-scale-good-one-scale-bad) triggers runbook's 'warning not green light' rule; NOT entitled to Phase 3 in any scenario. Adds two W96-L carry-forwards. Next move per Linear: COO-20 W96-D battlefield pivot.",
      "commits": [
        "63ccd29",
        "080a716",
        "0b20f19"
      ],
      "docs": [
        "docs/RUNBOOK_W96C.md",
        "docs/RESULTS_W96C_ARSENAL_MINING_V1.md",
        "docs/RESULTS_W96C_MATHVISTA_C1_11B_PHASE2_V1.md",
        "docs/RESULTS_W96C_MATHVISTA_C1_CROSS_SCALE_PHASE2_V1.md"
      ],
      "linear_issues": [
        "COO-19"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W96-L-MATHVISTA-V2-C1-VERIFIER-FINAL-11B-PHASE2-CAP",
        "W96-L-MATHVISTA-V2-C1-VERIFIER-FINAL-90B-PHASE2-SINGLE-SEED-NON-MECHANISM-DRIVEN-CAP"
      ]
    },
    {
      "id": "W97",
      "title": "RealWorldQA D2-B0 Phase 2 cheap pilot at 11B: FAIL (3 of 9 gates) — structurally informative; +6.67pp shortfall driven by vision-bound yes/no rescue cluster; cross-scale 90B not entitled",
      "outcome": "PHASE 2 FAIL — A0=36.67% / A1@K=5=90.00% / B=83.33%; B-A1=-6.67 pp (gate 4 FAIL); A1 saturated at 90% (gate 2 FAIL); B does not strictly beat A1 (gate 3 FAIL). Gates 1, 5-9 PASS, including B-A0=+46.67pp (image extraction is real signal) and B>=A1 on 25/30 problems (per-problem majority). Per-problem disagreement structure: 22/30 both pass; 5/30 unique-A1-rescues on vision-bound yes/no questions (existence detection, color/state, orientation, depth ordering); 3/30 unique-B-rescues on multi-choice spatial questions; 0/30 neither pass. Per the W96-C cross-scale rule, 90B Phase 2 is NOT auto-launched (estimated A1@K=5 >= 90% on this saturation-prone slice, almost certainly failing gate 2 again). Adds carry-forward W97-L-REALWORLDQA-D2-B0-PHASE2-11B-CAP. Discipline validation #7: W93/W94/W95/W96-A/W96-C/W96-D/W97.",
      "commits": [
        "4693467"
      ],
      "docs": [
        "docs/RUNBOOK_W97.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W97_V1.md",
        "docs/RESULTS_W97_ARSENAL_MINING_V1.md",
        "docs/RESULTS_W97_REALWORLDQA_D2_B0_PHASE2_V1.md"
      ],
      "linear_issues": [
        "COO-21"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W97-L-REALWORLDQA-D2-B0-PHASE2-11B-CAP"
      ]
    },
    {
      "id": "W96-D",
      "title": "Cross-modal battlefield pivot — ChartQA preflight FAIL at both 11B and 90B (P3 saturation); RealWorldQA preflight PASS at both 11B and 90B (all 9 composite gates); D2 preflight-earned for Phase 2; no NIM spend",
      "outcome": "BATTLEFIELD PIVOT D1->D2 — ChartQA (lmms-lab/ChartQA, 2500 problems; parquet SHA 165263505f2998aba65d819b44be832edecd92d676fee2c030645f784cd55d06) FAILS P3 saturation at both Llama-3.2-11B-Vision (A1@K=5 91.69%; residual 8.31pp) and Llama-3.2-90B-Vision (A1@K=5 92.75%; residual 7.25pp) — both far below the W95 +20pp floor; D1 killed cheaply. RealWorldQA (lmms-lab/RealWorldQA, 765 problems; 2 shards SHA-anchored) PASSES all 4 composite probes + W93 G1-G5 at both scales — 11B A1@K=5 73.44% (residual 26.56pp; clean); 90B A1@K=5 79.49% (residual 20.51pp; narrow 0.51pp above floor). D2-B0 (W95-B0 scene-port) is preflight-earned for a NIM smoke test + 1-seed x 30-problem Phase 2 cheap pilot at one or both scales; Phase 2 launch out of scope for this milestone. Adds 1 carry-forward (CHARTQA-PREFLIGHT-D1-B0-P3-SATURATION-CAP); no retirements. Discipline: W93/W94/W95/W96-A/W96-C/W96-D preflight-first + cross-scale rule caught a battlefield-level saturation cap cheaply at $0 NIM spend.",
      "commits": [
        "d14e46d"
      ],
      "docs": [
        "docs/RUNBOOK_W96D.md",
        "docs/RESULTS_W96D_ARSENAL_MINING_V1.md",
        "docs/RESULTS_W96D_CHARTQA_PREFLIGHT_V1.md",
        "docs/RESULTS_W96D_REALWORLDQA_PREFLIGHT_V1.md"
      ],
      "linear_issues": [
        "COO-20"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W96-L-CHARTQA-PREFLIGHT-D1-B0-P3-SATURATION-CAP"
      ]
    },
    {
      "id": "W98",
      "title": "RealWorldQA multi-candidate slate (B1 + B2 preflight earn BOTH at both scales; B1 cheap NIM pilot at 11B FAIL 2/9 gates) - structurally informative; B1 recovers 4/5 W97 unique-A1-rescues on yes/no perception but regresses 5 D2-B0 multi-choice wins; W95-B0 family capped at B-A1 ≈ -6.67pp through two distinct mechanisms; B2 deferred to W99",
      "outcome": "PHASE 2 FAIL VIA DIFFERENT MECHANISM — A0=36.67% / A1@K=5=86.67% / B1=80.00%; B1-A1=-6.67pp (gate 4 FAIL, IDENTICAL margin to W97 D2-B0); B1 does NOT strictly beat A1 (gate 3 FAIL). Gates 1, 2, 5-9 PASS — notably gate 2 (A1<90% this run; W97's saturation was sampling noise) and gate 6 (B>=A1 on 27/30 = IMPROVED per-problem majority vs W97 D2-B0's 25/30). Per-problem diff vs W97 D2-B0 on same slice: 4 B1-new-rescues on yes/no perception (000135, 000403, 000555, 000718) - EXACTLY the W97 failure cluster the typed schema + question-typed solver was designed to attack (4/5 recovery rate matches AddrP1 + AddrP2 predictions) - BUT 5 B1-regressions on multi-choice/numeric (000013, 000155, 000204, 000225, 000713) via a NEW failure mode: typed solver becomes more confident in reader's (often wrong) direct_answer_hint and stops K=4 reflexion-cycling. Net 4-5 = -1 problem = same architectural cap. W95-B0-derived extract-then-text-reason architecture family empirically capped at B-A1 ≈ -6.67pp at 11B through TWO distinct mechanisms (D2-B0 free text + D2-B1 typed schema). Cross-scale 90B NOT entitled per W96-C rule. B2 (direct-vision final-turn answerer, structurally distinct, no reader-hint dependency) deferred to W99 per runbook. Preflight + addressability probes (AddrP1..AddrP7) earned BOTH candidates at both 11B and 90B at $0 NIM (single pilot per runbook decision logic; tie-break = lower expected NIM cost). Discipline validation #8: W93/W94/W95/W96-A/W96-C/W96-D/W97/W98.",
      "commits": [
        "4e77107"
      ],
      "docs": [
        "docs/RUNBOOK_W98.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W98_V1.md",
        "docs/RESULTS_W98_ARSENAL_MINING_V1.md",
        "docs/RESULTS_W98_PREFLIGHT_V1.md",
        "docs/RESULTS_W98_REALWORLDQA_B1_PHASE2_V1.md"
      ],
      "linear_issues": [
        "COO-22"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W98-L-REALWORLDQA-B1-TYPED-SCHEMA-PHASE2-11B-CAP"
      ]
    },
    {
      "id": "W99",
      "title": "RealWorldQA candidate tournament (B2 + B4 + B5) at 11B: B2 + B5 PASS structural Phase 2 (100%/+6.67pp; Option A slice-saturation); B4 (typed schema sans hint) FAILS (76.67%/-16.67pp); typed-extract-then-text-reason sub-family of W95-B0 capped through THREE mechanisms; only B2 (image at decision boundary) clears Phase 2",
      "outcome": "2 of 3 PASS STRUCTURAL Phase 2 — B2 (frontier) + B5 (switch baseline ceiling); B4 (close-cousin repair) FAILS empirically refuting hint-removal-only hypothesis. NIM-free preflight + addressability probes earned all three at both 11B+90B; NIM-free B5 oracle prediction (+10pp) and B2 realistic prediction (+6.67pp) matched empirical to within sampling variance. B2 final-VLM rescued 3/3 invocations including the residual viewer-pov problem (000615) NEITHER W97 D2-B0 NOR W98 B1 could solve. Discipline validation #9 (W93-W99 preflight-first + cross-scale + multi-candidate tournament). Stable boundary preserved.",
      "commits": [
        "5526b4e"
      ],
      "docs": [
        "docs/RUNBOOK_W99.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W99_V1.md",
        "docs/RESULTS_W99_ARSENAL_MINING_V1.md",
        "docs/RESULTS_W99_PREFLIGHT_V1.md",
        "docs/RESULTS_W99_REALWORLDQA_B5_PHASE2_V1.md",
        "docs/RESULTS_W99_REALWORLDQA_B2_PHASE2_V1.md",
        "docs/RESULTS_W99_REALWORLDQA_B4_PHASE2_V1.md",
        "docs/RESULTS_W99_MILESTONE_SUMMARY_V1.md"
      ],
      "linear_issues": [
        "COO-23"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W99-L-REALWORLDQA-B5-SWITCH-BASELINE-PASS-11B-SLICE-SATURATION-CAP",
        "W99-L-REALWORLDQA-B5-SWITCH-IS-BASELINE-NOT-FRONTIER-CAP",
        "W99-L-REALWORLDQA-B2-DIRECT-VISION-FINAL-TURN-PHASE2-11B-STRUCTURAL-PASS-SLICE-SATURATION-CAP",
        "W99-L-REALWORLDQA-W95-B0-FAMILY-NOT-STRUCTURALLY-CAPPED-AT-IMAGE-AT-DECISION-BOUNDARY-CAP",
        "W99-L-REALWORLDQA-B4-TYPED-SCHEMA-WITHOUT-HINT-PHASE2-11B-FAIL-CAP",
        "W99-L-REALWORLDQA-TYPED-EXTRACT-THEN-REASON-SUBFAMILY-EMPIRICALLY-CAPPED-AT-11B-CAP"
      ]
    },
    {
      "id": "W100",
      "title": "RealWorldQA cross-scale 90B Phase 2 confirmation: B2 (frontier lead) FAILS by -3.33pp with mechanism-load-bearingness sub-gate MLB-2 also FAILING (clean cross-scale collapse matching W96-C C1); B5 (baseline-only ceiling reference) narrowly MISSES gate 4 by +1.67pp short of bar; pre-committed Part H code-pivot contingency triggers COO-9 (second code benchmark) promotion to lead path; cross-modal RealWorldQA arc FROZEN AT 11B; Phase 3 NOT launched",
      "outcome": "BOTH 90B Phase 2 pilots FAIL. B2 90B Phase 2: A0=46.67/A1@K=5=76.67/B2=73.33 / B-A1=-3.33pp / gates 3+4 FAIL / MLB-2 FAILS (final-VLM rescue rate 1/9=11.11% vs 11B 3/3=100%); MLB-1 PASS (invocation 30%); cross-scale shift -10pp; 22/30 both-pass with W99 11B, 0 new wins, 8 new losses; clean cross-scale collapse pattern matching W96-C C1 verifier. B5 90B Phase 2: A0=46.67/A1@K=5=80/B5=83.33 / B-A1=+3.33pp / 8/9 gates PASS (gate 4 alone fails); per-route vlm_team_b0(D2-B0) 14/18=77.8% (down from 18/18 at 11B; W95-B0 cross-scale fragility on multi-choice), a1_vlm_k5 11/12=91.7%; cross-scale shift -3.33pp; B5 stays classified baseline-only. NIM-free preflight reused W99 90B verdict cid 0bacd989... at /bin/zsh NIM; AddrW100-B2-P5 + AddrW100-B5-P4 cross-scale-stability probes both PASSed before any 90B NIM call. COO-9 PROMOTED to lead path; W101 charter = COO-9 runbook + corpus selection from {MBPP+, HumanEval+, APPS, LiveCodeBench, SWE-bench-lite}. Discipline validation #10 (W93-W100 preflight-first + cross-scale + new MLB sub-gates + multi-candidate-tournament-then-confirm). Stable boundary preserved. ZERO new coordpy.* modules introduced.",
      "commits": [
        "2383e6e"
      ],
      "docs": [
        "docs/RUNBOOK_W100.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W100_V1.md",
        "docs/RESULTS_W100_REALWORLDQA_B2_PHASE2_90B_V1.md",
        "docs/RESULTS_W100_REALWORLDQA_B5_PHASE2_90B_V1.md",
        "docs/RESULTS_W100_MILESTONE_SUMMARY_V1.md"
      ],
      "linear_issues": [
        "COO-24"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W100-L-REALWORLDQA-B2-DIRECT-VISION-FINAL-TURN-PHASE2-90B-CAP",
        "W100-L-REALWORLDQA-B2-CROSS-SCALE-COLLAPSE-MECHANISM-NON-LOAD-BEARING-AT-90B-CAP",
        "W100-L-REALWORLDQA-B5-SWITCH-BASELINE-90B-NARROW-MISS-CAP",
        "W100-L-REALWORLDQA-W95-B0-D2-B0-MULTI-CHOICE-EXTRACTION-DEGRADES-CROSS-SCALE-CAP"
      ]
    },
    {
      "id": "W101",
      "title": "Second-code-benchmark battlefield tournament + MBPP+ lead selection + cheap NIM-free preflight: 5-candidate x 8-criterion battlefield ranking (MBPP+ LEAD + HumanEval+ BACKUP); arsenal mining of W88+W91 sidecars via 2640 offline subprocess re-executions; 4 new explicit-import-only coordpy.* modules (loader+executor+reflexion-bench+preflight) + 3 driver scripts + 25 unit tests (all PASS); NIM-free preflight verdict 6/8 PASS with 2 DEFERRED on operator MBPP+ fetch; cheap NIM pilot NOT YET earned (conditional on operator MBPP+ download + SHA pin + preflight re-run clean)",
      "outcome": "INFRASTRUCTURE + PREFLIGHT — COO-9 charter executed end-to-end with NO NIM call. Battlefield selection (locked BEFORE any code built): MBPP+ chosen as LEAD (EvalPlus's hardened MBPP; ~35x more hidden tests per problem; surgical attack on W91-L-MBPP-REFLEXION-V2-5SEED-PARTIAL-CAP); HumanEval+ BACKUP (built only if MBPP+ preflight FAILs); APPS/LiveCodeBench/SWE-bench-lite explicitly out of scope. Arsenal mining: 2640 W88+W91 candidate responses re-executed offline; per-seed numbers match published W89+W91 byte-for-byte; reflexion-rescue surface 2.5x richer on HumanEval (9.76%) than on base MBPP (3.97%) = empirical symptom of ceiling-saturation cap MBPP+ relieves; MBPP shared-fails cluster (14.0%) 2.5x larger than HumanEval (5.6%). Preflight 6/8 PASS: P3 predicts MBPP+ A1@K=5=69.97% (saturation margin 20.03pp); P4 decomposition argument 1727 chars; AddrW101-P1 W89 rescue 9.76% >= 5% threshold; AddrW101-P2 both partitions well-formed; AddrW101-P3 cross-bench margin 20.03pp >= 10pp floor; AddrW101-P4 no anti-pattern tokens. 2 DEFERRED: P1 corpus integrity + P2 executor self-test (both gated on operator MBPP+ fetch + SHA pin). Cheap pilot NOT YET earned. Pre-committed Phase 2 gates: W95 9-gate + MLB-1 (reflexion-cycle invocation >=33%) + MLB-2 (reflexion rescue rate >=33%) per W100 lesson. NO carry-forwards retired (W89 70B HumanEval K=5 remains the only confirmed multi-seed same-budget multi-agent superiority retirement); NO carry-forwards added (infrastructure + preflight only; no empirical NIM result yet). Discipline validation #11 (W93-W101). Stable boundary preserved. 4 new modules explicit-import only; 25 unit tests all PASS.",
      "commits": [
        "e811893"
      ],
      "docs": [
        "docs/RUNBOOK_W101.md",
        "docs/RESULTS_W101_BATTLEFIELD_SELECTION_V1.md",
        "docs/RESULTS_W101_ARSENAL_MINING_V1.md",
        "docs/RESULTS_W101_PREFLIGHT_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W101_V1.md",
        "docs/RESULTS_W101_MILESTONE_SUMMARY_V1.md"
      ],
      "linear_issues": [
        "COO-25"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": []
    },
    {
      "id": "W102",
      "title": "MBPP+ V2 schema-fix lead lane + HumanEval+ backup lane + COO-14 helper lane + MBPP+ V2 cheap pilot at 70B FAIL (B-A1 = -6.67pp; MLB-1 + MLB-2 both FAIL; carry-forward W102-L-MBPP-PLUS-V2-REFLEXION-PHASE2-70B-CAP added; W103 pivots to HumanEval+ via the W102-built backup-lane infrastructure per pre-committed runbook branch-3 decision logic)",
      "outcome": "PHASE 2 FAIL — A0=73.33% / A1@K=5=83.33% / B=76.67% / B-A1=-6.67pp / 6 of 9 Phase 2 gates PASS / MLB-1 30% FAIL / MLB-2 22.22% FAIL on the W102-locked seed-101_001 / 30-problem slice at meta/llama-3.3-70b-instruct. Critical W102 finding (lead lane): during the W101 deferred operator step, W102 discovered the W101 V1 loader's schema assumption is wrong (parallel plus_input/plus_output arrays do NOT exist in the actual EvalPlus release; the extra-test surface lives in a single `test` Python program). V1 would have silently degenerated the cheap pilot to a base-MBPP run. W102 fixes this BEFORE any NIM spend by building V2 infrastructure (4 new explicit-import-only coordpy.* modules: mbpp_plus_loader_v2 + executor_v2 + reflexion_bench_v2 + preflight_v2 with NEW P5 silent-degeneration guard + P6 V1-vs-V2 sanity); V2 preflight 10 of 10 PASS (verdict cid aab6d9ab...). HumanEval+ backup lane (4 new explicit-import-only modules): humaneval_plus_loader_v1 + executor_v1 + reflexion_bench_v1 + preflight_v1; HumanEval+ preflight 7 of 7 PASS (verdict cid 4f57a2cf...); cheap pilot NOT launched in W102 — reserved as the W103 immediate pivot. COO-14 helper lane (1 new explicit-import-only module): code_slice_selector_v1 implements the 4-item COO-14 DoD verbatim (rank_candidate_benches + propose_cheap_pilot_slice + format_*_markdown + anti-pattern guard); 15 unit tests PASS. W102 arsenal-mining extension (scripts/run_w102_arsenal_mining.py) re-executes 2640 W88+W91 candidate responses offline against HumanEval+ + MBPP+ V2 surfaces; key empirical finding: MBPP+ V2 cluster surface on the SAME W91 70B responses lifts B-A1 from +1.33pp (base-MBPP cap) to +5.28pp — BUT the empirical W102 cheap pilot at seed 101_001 with FRESH K=5 sampling landed at -6.67pp (11.95pp swing below the arsenal-mining prior). This is added as a NEW W102 anti-pattern: cross-bench arsenal-mining priors are an UPPER BOUND, not cheap-pilot earning evidence. Per-problem cluster on the cheap-pilot slice: 3 a1_only_wins (B regression on problems where i.i.d. K=5 found a passing candidate but reflexion chain didn't — same W97/W98/W99 RealWorldQA-B0 structural pattern), 1 b_only_win (B rescue at attempt index 1), 22 shared_wins, 4 shared_fails. B's first-pass-attempt distribution: 21/23 PASSes came at attempt 0 (reflexion never invoked); 1/23 at attempt 1; 1/23 at attempt 3. MLB-2 rescue rate 22.22% on MBPP+ V2 is well below W89 HumanEval rescue rate of 47% — reflexion-mechanism load-bearingness on MBPP-family at 70B is empirically weaker than on HumanEval-family. Decision applied per pre-committed runbook branch 3: 2 carry-forwards added (W102-L-MBPP-PLUS-V2-REFLEXION-PHASE2-70B-CAP + W102-L-MBPP-PLUS-V2-MECHANISM-LOAD-BEARINGNESS-WEAK-AT-70B-CAP); W103 = HumanEval+ cheap pilot via W102-built backup; COO-9 REMAINS lead path; NO cross-scale MBPP+ V2; NO Phase 3. W101 V1 loader/executor/reflexion-bench DEMOTED to historical artifact + anti-pattern (silent-degeneration failure mode); V1 stays in-repo for the W101 audit trail. Discipline validation #12: W93/W94/W95/W96-A/W96-C/W96-D/W97/W98/W99/W100/W101/W102 with W102 extending the discipline with (a) silent-degeneration-via-schema-assumption anti-pattern guard (P5+P6 probes), (b) cross-bench-arsenal-mining-priors-as-cheap-pilot-earning-evidence anti-pattern guard, (c) mechanism-load-bearingness-varies-by-benchmark-family-at-70B structural lesson. Stable boundary preserved: coordpy.__version__=0.5.20; SDK_VERSION=coordpy.sdk.v3.43; no PyPI publish; coordpy/__init__.py untouched. 9 new explicit-import-only coordpy.* modules + 5 new driver scripts + 3 new unit-test files (67 tests total; all PASS).",
      "commits": [
        "55f40cf",
        "7419a6b"
      ],
      "docs": [
        "docs/RUNBOOK_W102.md",
        "docs/RESULTS_W102_MBPP_PLUS_LOADER_V2_FIX_V1.md",
        "docs/RESULTS_W102_HUMANEVAL_PLUS_PREFLIGHT_V1.md",
        "docs/RESULTS_W102_CODE_SLICE_SELECTOR_V1.md",
        "docs/RESULTS_W102_ARSENAL_MINING_V1.md",
        "docs/RESULTS_W102_MBPP_PLUS_V2_PHASE2_70B_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W102_V1.md",
        "docs/RESULTS_W102_MILESTONE_SUMMARY_V1.md"
      ],
      "linear_issues": [
        "COO-26"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W102-L-MBPP-PLUS-V2-REFLEXION-PHASE2-70B-CAP",
        "W102-L-MBPP-PLUS-V2-MECHANISM-LOAD-BEARINGNESS-WEAK-AT-70B-CAP"
      ]
    },
    {
      "id": "W103",
      "title": "HumanEval+ lead pilot at 70B PASS_MECHANISM_DRIVEN (B-A1 = +20.00pp; 9/9 Phase 2 gates + MLB-1 56.67% + MLB-2 47.06% all PASS on the helper-anchored slice; first NIM-spending pilot to consume coordpy.code_slice_selector_v1 as a load-bearing input; W104 cross-scale confirmation pre-committed; COO-9 remains lead path; W89 sequential-reflexion mechanism extends to a SECOND published code benchmark family (HumanEval+ EvalPlus-hardened) at the cheap-pilot scale; carry-forward W103-L-HUMANEVAL-PLUS-REFLEXION-PHASE2-70B-PASS registered; ZERO new coordpy.* modules)",
      "outcome": "PHASE 2 PASS_MECHANISM_DRIVEN — A0=50.00% / A1@K=5=50.00% / B (sequential reflexion K=5)=70.00% / B-A1=+20.00pp (4x larger than the W89 retirement's +5.56pp) / B-A0=+20.00pp / per-problem majority B ≥ A1 on 29 of 30 (only 1 a1_only regression) / 7 b_only_wins (mechanism rescues) / 14 shared_wins / 8 shared_fails / MLB-1=56.67% (17/30) PASS / MLB-2=47.06% (8/17) PASS (matches W89 base-HumanEval retirement template's 47% rescue rate byte-for-byte) / 9 of 9 Phase 2 gates PASS. Pilot 124 min wall (7424.3 s; heavy 429 throttling + early-launch socket hang; killed + relaunched at 12 min; hardened retry budget kept the run grinding without exhausting). Bench Merkle root 68f4a9669f1bd03e6b3cb393a436e4f04aca034a0bad9c4b5ea8a002faabfd6d. Slice CID c35155956ece605c0169b0cf35a6b69267bee04f5f68cf5a5de466dcc01dd8d2 (helper-priority order; 28 unique HumanEval+ task_ids + 2 base-HumanEval top-ups; cluster mix 7 humaneval_plus:b_only_wins + 10 humaneval_plus:shared_fails + 2 humaneval_plus:a1_only_wins + 9 humaneval_plus:shared_wins + 2 humaneval(top-up):shared_wins = 63.3% historically-hard). Notable empirical finding: A1@K=5=50.00% is EXACTLY EQUAL to A0=50.00% on this helper-anchored slice (K=5 i.i.d. sampling produced zero improvement over single-shot); sequential reflexion (B) recovered 8 of the 17 problems where A0 failed via cumulative-stderr conditioning. This is a stronger mechanism-load-bearing signal than the W89 retirement showed (where A1 ≈ 85.56% so the i.i.d. surface was structurally smaller); reflexion is doing real work that i.i.d. sampling cannot substitute for on this slice. Decision applied per pre-committed Branch A: W103-L-HUMANEVAL-PLUS-REFLEXION-PHASE2-70B-PASS registered (single-seed cheap-pilot PASS; NOT retirement); COO-9 REMAINS lead path; W104 = HumanEval+ cross-scale confirmation at a SECOND model class (W104 RUNBOOK locks the exact target); W105+ = Phase 3 retirement bench (3 seeds × 100 problems × K=5) IF W104 cross-scale PASSes. W102 / W103 cross-bench contrast: reflexion-mechanism load-bearingness is benchmark-family-dependent at 70B — MBPP-family rescue rate 22.22% (W102 FAIL) vs HumanEval-family rescue rate 47.06% (W103 PASS, identical to W89 47%). This is now an empirical fact, not a conjecture. Hardening lane: 9 unit tests codifying W102 + W103 lessons (silent-degeneration anti-pattern + arsenal-mining-prior anti-pattern + sidecar-flush hardening + helper-anchored-slice integrity + anti-pattern token absence + production slice CID determinism pin); driver now flushes the sidecar after every write (W102 buffered all 330 entries until pilot exit, blinding mid-run operator audits; W103 codifies the fix via test_w103_pilot_driver_flushes_sidecar_after_each_write). Planning lane: W104 pre-committed by outcome BEFORE pilot launched (Branch A applied; Branch B `PASS_NON_MECHANISM_DRIVEN` and Branch C FAIL decision logic remain in docs/RUNBOOK_W103.md as durable artifacts of pre-commit-then-execute discipline). Helper-consumption attestation (W102 COO-14 downstream-consumption deliverable): coordpy.code_slice_selector_v1 transitions from \"shipped helper with worked example\" to \"shipped helper as load-bearing pilot input\". Stable boundary preserved: coordpy.__version__=0.5.20; SDK_VERSION=coordpy.sdk.v3.43; no PyPI publish; coordpy/__init__.py untouched. ZERO new coordpy.* modules in W103. 2 new driver scripts (run_w103_humaneval_plus_pilot.py + _w103_emit_pilot_result_doc.py); 1 modified driver (+5 lines for sidecar flush); 1 modified test file (+21 lines for sidecar-flush regression); 6 new docs (RUNBOOK + 4 RESULTS + FRONTIER_AUDIT). 75 tests across W101+W102+W103 code line; all PASS. Discipline validation #13: W93/W94/W95/W96-A/W96-C/W96-D/W97/W98/W99/W100/W101/W102/W103.",
      "commits": [
        "b7a4c3d",
        "1702e28"
      ],
      "docs": [
        "docs/RUNBOOK_W103.md",
        "docs/RESULTS_W103_HELPER_CONSUMPTION_V1.md",
        "docs/RESULTS_W103_HUMANEVAL_PLUS_PREFLIGHT_RECONFIRM_V1.md",
        "docs/RESULTS_W103_HUMANEVAL_PLUS_PHASE2_70B_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W103_V1.md",
        "docs/RESULTS_W103_MILESTONE_SUMMARY_V1.md"
      ],
      "linear_issues": [
        "COO-27"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W103-L-HUMANEVAL-PLUS-REFLEXION-PHASE2-70B-PASS"
      ]
    },
    {
      "id": "W104",
      "title": "HumanEval+ cross-scale Phase 2 cheap pilot at the pre-locked-backup target (Llama-3.1-70B-Instruct, cross-generation) PASS_MECHANISM_DRIVEN + cross-scale-discipline hardening + W105 entitlement pre-built",
      "outcome": "PHASE 2 PASS_MECHANISM_DRIVEN AT CROSS-GENERATION FORM — A0=46.67% / A1@K=5=53.33% / B=63.33% / B-A1=+10.00pp / B-A0=+16.66pp / 9 of 9 Phase 2 gates PASS / MLB-1=56.67% (17/30) PASS / MLB-2=35.29% (6/17) PASS. Pre-locked primary `meta/llama-3.1-405b-instruct` returned HTTP 404 on NIM (not hosted); pre-locked backup `meta/llama-3.1-70b-instruct` was applied deterministically per the W104 RUNBOOK § Target-model selection rule criterion 2(b). Cross-scale form actually achieved: cross-GENERATION (Llama 3.1 vs Llama 3.3 at 70B), NOT cross-scale-UP (70B → 405B). Pilot ran 1 seed × 30 problems × K=5 = 330 NIM calls on the BYTE-EQUAL W103 helper-anchored slice (slice CID c35155956ece605c0169b0cf35a6b69267bee04f5f68cf5a5de466dcc01dd8d2 verified at run start; corpus SHA 908377f1daf28dcb36846db73a5662b2e05a9907407c2696c89ad9d3b0b04492 verified at run start; pilot refuses to run on either mismatch). 7506.9 s wall (~125 min). Cross-scale comparator emitted (after a brief permission-outage retry): 8 stayed + 11 improved + 11 regressed + 0 flipped on the byte-equal slice; cross-generation shift on B-A1 = -10.00 pp (W103 +20 → W104 +10; mechanism keeps half its W103 margin); cross-generation shift on MLB-2 = -11.77 pp (47.06 → 35.29, still load-bearing above the 33% floor). Hardening lane: 4 durable guardrails landed (cross-scale comparator V1 schema/provenance refuse-to-run; byte-equal W103 slice reuse with run-start CID verification; reachability smoke probe BEFORE NIM spend with pre-locked backup; resume-from-sidecar capability for socket hangs / 429 storms). 14 PASSing unit tests codify all four guardrails. Planning lane: W105 Phase 3 retirement bench slice pack pre-built (pack CID 8be55f3bf1650df397cb875543c69a48473483de8089dc3c40be45cc635a1314; 100 problems with W103 30-problem inner kernel preserved + 45 mid-shell helper extension + 25 corpus-fill; 3 seeds 105_001 / 105_002 / 105_003; cross-scale axis Llama-3.3-70B + Llama-3.1-70B; 6 600 NIM calls total). Branch C fallback dispatch table (machine-readable JSON) pre-committed: LiveCodeBench preflight on mechanism-distribution shift; APPS preflight on G2 saturation; HumanEval+ multi-seed at cross-scale target on per-seed-sampling variance; cross-scale-collapse audit + 70B Phase 3 on W96-A regression pattern; SWE-bench-lite stays unconditionally out of scope. Helper consumption: coordpy.code_slice_selector_v1 consumed as a real downstream input for the second time (first was W103; COO-14 deliverable extended from `load-bearing pilot input` to `load-bearing Phase 3 slice-pack input`). Decision applied: Branch A. Carry-forwards added: W104-L-HUMANEVAL-PLUS-REFLEXION-PHASE2-CROSS-GENERATION-70B-LLAMA31-PASS + W104-L-HUMANEVAL-PLUS-CROSS-SCALE-UP-PRIMARY-TARGET-405B-UNREACHABLE-ON-NIM-CAP. Carry-forwards retired: NONE; W89 70B-HumanEval K=5 remains the only confirmed multi-seed same-budget multi-agent superiority retirement. `COO-9` REMAINS the lead path. Discipline validation #14: W93/W94/W95/W96-A/W96-C/W96-D/W97/W98/W99/W100/W101/W102/W103/W104. Stable boundary preserved: coordpy.__version__=0.5.20; SDK_VERSION=coordpy.sdk.v3.43; no PyPI publish; coordpy/__init__.py untouched. Exactly ONE new coordpy.* module added explicit-import only: coordpy.cross_scale_comparator_v1.",
      "commits": [
        "35e7c3d"
      ],
      "docs": [
        "docs/RUNBOOK_W104.md",
        "docs/RESULTS_W104_HUMANEVAL_PLUS_PHASE2_405B_V1.md",
        "docs/RESULTS_W104_CROSS_SCALE_COMPARATOR_V1.md",
        "docs/RESULTS_W104_HELPER_W105_PLANNING_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W104_V1.md",
        "docs/RESULTS_W104_MILESTONE_SUMMARY_V1.md"
      ],
      "linear_issues": [
        "COO-28"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W104-L-HUMANEVAL-PLUS-REFLEXION-PHASE2-CROSS-GENERATION-70B-LLAMA31-PASS",
        "W104-L-HUMANEVAL-PLUS-CROSS-SCALE-UP-PRIMARY-TARGET-405B-UNREACHABLE-ON-NIM-CAP"
      ]
    },
    {
      "id": "W105",
      "title": "HumanEval+ Phase 3 retirement bench (the earned 6,600-call run) + run-hardening + W106 contingency — SPLIT outcome: meta/llama-3.3-70b-instruct RETIRED (second confirmed multi-seed same-budget multi-agent superiority retirement after W89); meta/llama-3.1-70b-instruct FAIL_MARGIN; cross-class retirement NOT entitled",
      "outcome": "SPLIT — meta/llama-3.3-70b-instruct RETIRED (6/6 W88/W89/W95 retirement bars; 3 seeds x 100 problems x K=5; mean B-A1=+7.00pp; per-cell +5.00/+9.00/+7.00; per-seed majority 3/3; per-problem majority 295/300; A1 84/82/82% all <90%; audit chain 3/3; executor clean 100%; MLB-2=55.62% load-bearing) = the SECOND confirmed multi-seed same-budget multi-agent superiority retirement after W89, on a different benchmark family (EvalPlus-hardened HumanEval+) at +7.00pp vs W89's base-HumanEval +5.56pp. meta/llama-3.1-70b-instruct FAIL_MARGIN (5/6 bars; mean B-A1=+2.33pp; per-cell +5.00/+1.00/+1.00; only the margin bar fails; per-seed majority 3/3; MLB-2=50.54% still load-bearing). The W104 cross-generation cheap-pilot +10.00pp on the 30-problem rescue-concentrated slice did NOT survive scale-up to the broad 100-problem Phase 3 slice for Llama-3.1 (A1 rose to 86.33%, compressing reflexion headroom) — confirms the W102 cheap-pilot-margin-is-an-upper-bound anti-pattern + the W96-A/W96-C/W100 cross-scale-collapse pattern. CROSS-CLASS RETIREMENT NOT ENTITLED: the W105 RUNBOOK rule requires BOTH classes to clear all 6 bars; only Llama-3.3-70B did (cross-class B-A1 diff 4.67pp is within the +-5pp envelope but condition (i) both-RETIRED fails); bounded claim is single-class on Llama-3.3-70B. Executed the pre-built slice pack (pack CID 8be55f3bf1650df397cb875543c69a48473483de8089dc3c40be45cc635a1314) BYTE-FOR-BYTE unchanged; 6600 NIM calls; canary smoke (66 calls) PASSed both classes before launch; 405B re-probed HTTP 404 (still unreachable; W104 cap stands; core two-class matrix unaffected); per-seed-aligned cross-class comparator clean (242 stayed / 27 improved / 28 regressed / 3 flipped across 300 cells). Three lanes: lead (Phase 3 execution) + hardening (phase3_retirement_evaluator_v1 + cross_class_comparator_v1, both explicit-import-only; 18 unit tests; 79 PASS across W101-W105) + planning (W106 pre-committed under all 5 verdict shapes; empirical SPLIT maps to Verdict C sub-case C1). COO-9 REMAINS lead path. 15th consecutive preflight-discipline validation (W93-W105). Stable boundary preserved: coordpy.__version__=0.5.20; SDK_VERSION=coordpy.sdk.v3.43; no PyPI publish; coordpy/__init__.py untouched; 2 new explicit-import-only coordpy.* modules.",
      "commits": [
        "d5f9133"
      ],
      "docs": [
        "docs/RUNBOOK_W105.md",
        "docs/RESULTS_W105_HUMANEVAL_PLUS_PHASE3_LLAMA33_V1.md",
        "docs/RESULTS_W105_HUMANEVAL_PLUS_PHASE3_LLAMA31_V1.md",
        "docs/RESULTS_W105_CROSS_CLASS_COMPARATOR_V1.md",
        "docs/RESULTS_W105_W106_PLANNING_V1.md",
        "docs/RESULTS_W105_MILESTONE_SUMMARY_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W105_V1.md"
      ],
      "linear_issues": [
        "COO-29"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W105-T-HUMANEVAL-PLUS-RETIREMENT-LLAMA33-70B",
        "W105-L-HUMANEVAL-PLUS-RETIREMENT-LLAMA31-70B-MARGIN-CAP",
        "W105-L-HUMANEVAL-PLUS-CROSS-CLASS-RETIREMENT-NOT-ENTITLED-CAP"
      ]
    },
    {
      "id": "W106",
      "title": "Bounded second-retirement registration + Llama-3.1 margin-cap dispatch NO-GO + graphify truth-sync",
      "outcome": "REGISTRATION + DISPATCH milestone (NOT a new benchmark tournament; NO expensive run; $0 NIM on the Llama-3.1 branch). Executes the pre-committed RESULTS_W105_W106_PLANNING_V1 Verdict C / sub-case C1 (retired class=Llama-3.3-70B, failed class=Llama-3.1-70B), per docs/RUNBOOK_W106.md locked BEFORE any NIM call. Lane 1 (claim): REGISTERS the W105 meta/llama-3.3-70b-instruct HumanEval+ Phase 3 retirement (+7.00pp; 6/6 bars; MLB-2=55.62%) as the SECOND confirmed multi-seed same-budget multi-agent superiority retirement after W89, on a different benchmark family; BOUNDED to ONE model class / ONE benchmark family (HumanEval+) / ONE parameter scale (70B). Lane 2 (margin-cap dispatch): a pre-committed TWO-GATE rule (coordpy.margin_cap_dispatch_v1; 20 unit tests) returned NO-GO on the Llama-3.1-70B FAIL_MARGIN branch -- GATE 1 (entitlement) ENTITLED (margin +2.33 in [0,+5); MLB-2 50.54%>=33%; A1 86.33%<90%; Branch-C cheap-confirmation row ~990 NIM calls) but GATE 2 (verdict-changing power) FAIL on all three sub-conditions (2a rescue-concentrated slice is an UPPER BOUND = the W102 anti-pattern; 2b the authoritative fair broad-slice multi-seed Phase 3 verdict already ran at W105 at +2.33pp; 2c the miss is a CLEAN true magnitude miss -- executor clean 100%, byte-exact budget, per-seed majority 3/3, MLB-2 50.54% healthy -- no confound). Decision CID de3dfb02...; the Llama-3.1 branch is CLOSED, not deferred. Lane 3 (graphify): graph refreshed from HEAD at start AND close (73,026 nodes / 237,304 edges / 2,336 communities; dated backup graphify-out/2026-05-28/); query/affected/explain used to confirm the new module is a sibling of the W104/W105 evaluators/comparators. COO-9 REMAINS lead path. W107 left obvious: 405B reachability gate -> alpha cross-scale-UP cheap pilot (if reachable) / beta next-code-benchmark NIM-free preflight LiveCodeBench/APPS (if 405B HTTP 404) / gamma W89->W106 consolidation in parallel. No empirical retirement added or retired; W89 + W105 remain the two confirmed retirements (both Llama-3.3-70B). 16th consecutive preflight-discipline validation (W93-W106). Stable boundary preserved: coordpy.__version__=0.5.20; SDK_VERSION=coordpy.sdk.v3.43; no PyPI publish; coordpy/__init__.py untouched; ONE new explicit-import-only coordpy.* module.",
      "commits": [
        "9ba018e"
      ],
      "docs": [
        "docs/RUNBOOK_W106.md",
        "docs/RESULTS_W106_BOUNDED_RETIREMENT_REGISTRATION_V1.md",
        "docs/RESULTS_W106_MARGIN_CAP_DISPATCH_V1.md",
        "docs/RESULTS_W106_MILESTONE_SUMMARY_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W106_V1.md"
      ],
      "linear_issues": [
        "COO-30"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W106-T-BOUNDED-SECOND-RETIREMENT-REGISTERED",
        "W106-T-MARGIN-CAP-DISPATCH-V1-SHIPS",
        "W106-L-HUMANEVAL-PLUS-LLAMA31-70B-MARGIN-CAP-CHEAP-CONFIRMATION-NOT-EARNED-CAP"
      ]
    },
    {
      "id": "W107",
      "title": "405B reachability gate (4th HTTP 404 -> CLOSED) + next-code-battlefield β preflight (LiveCodeBench PRIMARY / APPS pivot backup) + publication-grade W89→W106 consolidation",
      "outcome": "Gated branch milestone with THREE lanes, NOT a new broad benchmark tournament; /bin/zsh expensive NIM (only the sub-second free 405B side-probe). docs/RUNBOOK_W107.md locked BEFORE any NIM call. Lane α (405B gate): re-probed meta/llama-3.1-405b-instruct on NIM -> HTTP 404 (183ms), the FOURTH consecutive 404 (W104/W105/W106/W107); GATE=CLOSED (decision CID 332d4ef983313f7faf724c7d8b2ad96e8e5964a125dfd30d177bc49ba6b2111e); no 405B cheap pilot earned or launched; § 3 cheap-pilot rule NOT exercised; W104-L-...-405B-UNREACHABLE-ON-NIM-CAP refreshed. Lane β (main empirical lane): applied W101 C1-C8 rubric + W107 S1∧S2∧S3 structural-soundness pivot test -> LiveCodeBench is the structurally-sound PRIMARY (time-anchored contamination resistance is the decisive publication-grade property; functional starter_code subset has a clean deterministic subprocess executor proven offline: gold PASS / wrong FAIL / infinite-loop TIMEOUT; W89 decomposition fits), APPS the structural-pivot backup (no pivot triggered); shipped coordpy.livecodebench_loader_v1 (SHA-pinnable functional-subset loader; refuses unpinned/mismatched corpus = W102 silent-degeneration guard) + coordpy.livecodebench_executor_v1 (clean functional-form subprocess executor) + scripts/run_w107_livecodebench_preflight.py (verdict CID 55910d11e210c323fb1a393bbf8be1c3ffa2d19dd22f9e9f40e52ffc9746c6b6; 16 PASSing tests); A1 residual honestly recorded as published-baseline-grade pending operator corpus-fetch. Lane γ (consolidation): docs/CONSOLIDATED_CODE_RETIREMENT_NARRATIVE_V1.md (publication-grade W89→W106 arc: TWO confirmed retirements, both Llama-3.3-70B @ 70B, bounded on three axes) + harmonized claim surface across THEOREM_REGISTRY / RESEARCH_STATUS / HOW_NOT_TO_OVERSTATE. W107 adds NO empirical retirement and retires NO cap; W89 + W105 remain the two confirmed retirements. COO-9 stays lead path; W108 = LiveCodeBench functional-subset Phase 2 cheap pilot after operator corpus-fetch (or APPS pivot, or honest no-go). 17th consecutive preflight-discipline validation (W93-W107). Stable boundary preserved (no version bump; no PyPI; coordpy/__init__.py untouched; 2 new explicit-import-only modules).",
      "commits": [
        "9caa0d2"
      ],
      "docs": [
        "docs/RUNBOOK_W107.md",
        "docs/RESULTS_W107_405B_GATE_V1.md",
        "docs/RESULTS_W107_NEXT_BATTLEFIELD_PREFLIGHT_V1.md",
        "docs/RESULTS_W107_MILESTONE_SUMMARY_V1.md",
        "docs/CONSOLIDATED_CODE_RETIREMENT_NARRATIVE_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W107_V1.md"
      ],
      "linear_issues": [
        "COO-31"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W107-T-405B-GATE-FOURTH-404-CLOSED",
        "W107-T-LIVECODEBENCH-PREFLIGHT-V1-SHIPS",
        "W107-T-CODE-RETIREMENT-NARRATIVE-CONSOLIDATED",
        "W107-L-LIVECODEBENCH-LOADER-V1-SCHEMA-CONFIRM-AT-FETCH-CAP",
        "W107-L-LIVECODEBENCH-RESIDUAL-PUBLISHED-BASELINE-GRADE-CAP",
        "W107-L-LIVECODEBENCH-FUNCTIONAL-SUBSET-ONLY-CAP"
      ]
    },
    {
      "id": "W108",
      "title": "LiveCodeBench real-data bug-fix + cheap-pilot FAIL (B-A1=-3.33pp; MLB-2=25%) + APPS backup readiness + 405B 5th-404",
      "outcome": "Gated branch milestone with THREE lanes, NOT a new broad benchmark tournament. docs/RUNBOOK_W108.md locked BEFORE any expensive NIM call; the one expensive run was the EARNED 330-call cheap pilot ($0 expensive 405B). Lane alpha (405B gate): re-probed meta/llama-3.1-405b-instruct -> HTTP 404 (163ms), the FIFTH consecutive 404 (W104-W108); GATE=CLOSED (decision CID e1af4451e4d15e7038016f7acd50d72560a77fc874470157c88d1426ee370e8d); W104-L-...-405B-UNREACHABLE-ON-NIM-CAP refreshed. Lane beta (LiveCodeBench, main): diagnosed + fixed the partial-scaffold gold-path smoke (A0=A1=B=0.0) -> root cause was the real release_v6 corpus storing metadata as a JSON STRING, so the W107 loader left func_name='' -> executor ENTRY_NOT_FOUND on every arm; livecodebench_loader_v1._resolve_func_name now handles both encodings (+starter_code fallback); 19 regression tests lock it. Real-data preflight PASS (SHA pin bb4c364f...; 63 functional, all func_name resolved, all plain-arg, all dated 2025 post-cutoff; slice CID 2afc318c...; verdict CID 61b9961c...) -> pilot EARNED. Cheap pilot (meta/llama-3.3-70b-instruct; 1 seed x 30 x K=5 = 330 NIM calls; ~77min; 444 HTTP-429 retries survived): A0=43.33% / A1@K5=63.33% / B=60.00% / B-A1=-3.33pp; 7/9 gates (G3+G4 fail); MLB-1=53.33% PASS, MLB-2=25% FAIL -> FAIL (NON-mechanism-driven). FIRST contamination-resistant test of the W89 mechanism, and it FAILed. Lane gamma (APPS backup): real coordpy.apps_loader_v1 + apps_executor_v1 + preflight + 7 tests + codified in-milestone pivot conditions; pivot NOT triggered (LiveCodeBench passed real-data soundness); APPS is 2021 contamination-exposed (C7=C), backup/control only. Truth surface: the TWO confirmed retirements (W89 +5.56pp; W105 +7.00pp, both Llama-3.3-70B @ 70B) STAND unchanged; W108 adds NO retirement and retires NO research carry-forward; both retirements are on contamination-EXPOSED HumanEval-family (2021) problems and the contamination-confound is an OPEN hypothesis (not a finding). The two carry_forwards_retired below are W107 INFRASTRUCTURE caps discharged by the real-data fetch (schema confirmed; live A1 measured), NOT research retirements. graphify refreshed from HEAD at start + close. 18th consecutive preflight-discipline validation (W93-W108). COO-9 stays lead; W109 = APPS contaminated-control contrast or multi-seed LCB de-noise. Stable boundary preserved (no version bump; no PyPI; coordpy/__init__.py untouched; 2 new explicit-import-only APPS modules + slice selector; loader fixed in place).",
      "commits": [
        "9b1b29a",
        "e1c0ea7"
      ],
      "docs": [
        "docs/RUNBOOK_W108.md",
        "docs/RESULTS_W108_LIVECODEBENCH_PHASE2_70B_V1.md",
        "docs/RESULTS_W108_MILESTONE_SUMMARY_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W108_V1.md"
      ],
      "linear_issues": [
        "COO-32"
      ],
      "carry_forwards_retired": [
        "W107-L-LIVECODEBENCH-LOADER-V1-SCHEMA-CONFIRM-AT-FETCH-CAP",
        "W107-L-LIVECODEBENCH-RESIDUAL-PUBLISHED-BASELINE-GRADE-CAP"
      ],
      "carry_forwards_added": [
        "W108-T-LIVECODEBENCH-REAL-DATA-BUGFIX-METADATA-JSON-STRING",
        "W108-T-LIVECODEBENCH-REAL-DATA-PREFLIGHT-EARNED",
        "W108-T-405B-GATE-FIFTH-404-CLOSED",
        "W108-T-APPS-BACKUP-SCAFFOLDING-REAL-PIVOT-READY",
        "W108-L-LIVECODEBENCH-REFLEXION-PHASE2-70B-CAP",
        "W108-L-REFLEXION-NOT-DEMONSTRATED-ON-CONTAMINATION-RESISTANT-BENCH-CAP"
      ]
    },
    {
      "id": "W109",
      "title": "APPS contaminated-control contrast (B-A1=+16.67pp; PASS_NON_MECHANISM_DRIVEN) + LiveCodeBench de-noise NOT-WARRANTED + claim tightening (contamination-confound now SUPPORTED, not proven)",
      "outcome": "Gated branch milestone with THREE lanes, NOT a new broad benchmark tournament. docs/RUNBOOK_W109.md locked BEFORE any expensive NIM call; the one expensive run was the EARNED 330-call APPS cheap pilot ($0 on LCB de-noise; $0 on 405B, not re-probed). Lane alpha (APPS contaminated-control, MAIN): fetched the REAL codeparrot/apps corpus via refs/convert/parquet @ 0f10e424 (config all/test, 743MB SHA-verified shards, 5000 problems), materialized the SHA-pinned call-based subset (38 problems: 28 interview/10 introductory; apps-test.jsonl SHA f6c44d76); confirmed real schema (input_output JSON string -> {fn_name,inputs,outputs}; heterogeneous output-wrapper faithfully matched by output==expected OR output==expected[0]) -- DISCHARGES the two W108 APPS confirm-at-fetch caps. Built coordpy.apps_reflexion_bench_v1 (A0/A1/B byte-identical in shape to W89/W105/W108). Real-data preflight P1^P2^P3^P4 PASS (slice CID 783687d6; verdict CID 0cf1a8e2) -> pilot EARNED. Cheap pilot (meta/llama-3.3-70b-instruct, 1 seed x 30 x K=5 = 330 calls, ~75min, bench Merkle a571c08b): A0=73.33% / A1@K5=73.33% / B=90.00% / B-A1=+16.67pp; 9/9 gates; MLB-2=57.14% PASS, MLB-1=23.33% FAIL => PASS_NON_MECHANISM_DRIVEN (0 regressions; 4 of 5 B-wins are reflexion rescues; MLB-1 fails only because A0 is high). The W89 mechanism RECOVERED a large same-budget win on contamination-EXPOSED APPS (+16.67pp) where it FAILed on contamination-RESISTANT LiveCodeBench (-3.33pp; W108) -- a double dissociation by vintage reinforced by an A0 single-shot gap (73.33% exposed vs 43.33% resistant). Moves the contamination-confound from OPEN to SUPPORTED but NOT established (one single-seed control pair; APPS PASS non-mechanism-driven; APPS contamination-EXPOSED => CONTROL evidence only, NOT a third retirement, NOT publication-grade). Lane beta (LiveCodeBench de-noise DECISION): coordpy.livecodebench_denoise_decision_v1 two-gate rule returns NOT WARRANTED on the W108 result (negative margin + weak MLB-2; +8.33pp mean shift multi-seed cannot supply; decision CID 290afa46); $0 further LCB NIM; does NOT re-open the closed Llama-3.1 branch. Lane gamma (claim/graphify/truth): graphify refreshed from HEAD at start+close; query/explain/path/affected used; added docs/CONTAMINATION_CONTROL_FRAMING_W109_V1.md (exposed-vs-resistant 2x2 + honesty rules) and tightened RESEARCH_STATUS/THEOREM_REGISTRY/HOW_NOT_TO_OVERSTATE/CONSOLIDATED narrative/CHANGELOG. Adds NO retirement and retires NO research carry-forward (W89+W105 STAND; boundary now SHARPER). COO-9 stays lead; W110 = a SECOND contamination-RESISTANT benchmark. 19th consecutive preflight-discipline validation (W93-W109). Stable boundary preserved (no version bump; no PyPI; coordpy/__init__.py untouched; 2 new explicit-import-only modules + 3 scripts; APPS loader/executor reused unchanged).",
      "commits": [
        "2ebb15ce785009c1b83afeccbf23463c643e8c1e",
        "7954129baf7bee8130480a8bfcc85adfe1a67006"
      ],
      "docs": [
        "docs/RUNBOOK_W109.md",
        "docs/RESULTS_W109_APPS_CONTROL_PHASE2_70B_V1.md",
        "docs/RESULTS_W109_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W109_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W109_V1.md"
      ],
      "linear_issues": [
        "COO-33",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W109-T-APPS-REAL-DATA-FETCH-PINNED",
        "W109-T-APPS-CONTROL-PREFLIGHT-EARNED",
        "W109-L-APPS-CONTROL-PHASE2-70B-PASS-NON-MECHANISM-DRIVEN-CAP",
        "W109-L-APPS-CONTROL-MLB1-INVOCATION-CAP",
        "W109-T-CONTAMINATION-CONFOUND-SUPPORTED-NOT-PROVEN",
        "W109-T-LIVECODEBENCH-DENOISE-NOT-WARRANTED"
      ]
    },
    {
      "id": "W110",
      "title": "Second contamination-RESISTANT benchmark = BigCodeBench; Phase-2 FAIL (B-A1=+0.00pp; MLB-2=25%) ⇒ W108 LCB FAIL is NOT LCB-specific; resistant superiority 0/2; contamination-confound STRENGTHENED-not-proven",
      "outcome": "Gated branch milestone with THREE lanes, NOT a new broad benchmark tournament. docs/RUNBOOK_W110.md locked BEFORE any expensive NIM call; the one expensive run was the EARNED 330-call BigCodeBench cheap pilot ($0 on SWE-bench-lite, LiveBench-coding, APPS, LCB de-noise, 405B). Lane alpha (MAIN): locked the S1^S2^S3^S4 + feasibility selection rule, then on real-data HF probes REJECTED SWE-bench-lite (synthetic in-repo swe_* MiniSWEBank; real instances need Docker/per-repo-env; multi-file patches break the K=5 single-artifact byte-exact budget) and LiveBench-coding (it IS LiveCodeBench repackaged: task=LCB_generation, citation via livecodebench), and SELECTED BigCodeBench v0.1.4 (2024-06 post-cutoff release-date resistance; clean unittest oracle; single task_func K=5; novel library-composition). Fetched + SHA-pinned the corpus (refs/convert/parquet shard SHA d9a49658; materialized JSONL SHA ca4f352e; 1140 problems). Built 4 explicit-import-only modules (bigcodebench_loader_v1 + bigcodebench_executor_v1 + bigcodebench_reflexion_bench_v1 + contamination_resistant_interpretation_v1). Caught + fixed an in-milestone executor bug: the macOS interactive matplotlib backend (360/1140 BigCodeBench tasks plot) popped GUI windows AND risked a blocking plt.show() falsely TIMING-OUT correct chart solutions; forced headless Agg -> recovered +32 gold-green (936->968); added a wall-stability guard (drop golds >=20s) so the slice CID is reproducible (b69bf3a0, reproduced across runs). Real-data preflight P1^P2^P3^P4 PASS (gold-green 968/1140; verdict CID 6be9fc8e) -> pilot EARNED; canary validated the live path. Cheap pilot (meta/llama-3.3-70b-instruct, 1 seed x 30 x K=5 = 330 calls, ~106min, bench Merkle 128dfb19): A0=63.33 / A1@K5=70.00 / B=70.00 %; B-A1=+0.00pp; 7/9 gates; MLB-1=40% PASS, MLB-2=25% FAIL -> FAIL (NON-mechanism-driven); B rescued 1 (/51), regressed 1 (/26), net 0. This is the SECOND contamination-RESISTANT FAIL -> the W108 LiveCodeBench FAIL is NOT LCB-specific; the W89 mechanism fails on contamination-resistant code GENERALLY at 70B (resistant superiority 0/2 vs exposed 3/3). Lane beta (pre-committed interpretation rule, locked before the verdict): confound_direction=STRENGTHENS, earns_phase3=False; contamination-confound SUPPORTED->STRENGTHENED toward a finding, NOT proven (two single-seed resistant points; orthogonal difficulty not excluded). Lane gamma: graphify refreshed from HEAD at start (1e8f131) + mid (73a0212) + end; explain/path/affected/query run (BigCodeBench bench = 4-hop sibling of the APPS + LiveCodeBench benches). Two confirmed retirements (W89, W105) STAND, boundary now contamination-EXPOSED-specific at 70B. W110 adds NO retirement and retires NO research carry-forward. 20th consecutive preflight-discipline validation (W93-W110). Stable boundary preserved (no version bump; no PyPI; coordpy/__init__.py untouched; 4 new modules + 3 scripts + 3 test files / 17 tests). COO-9 REMAINS lead; W111 = register the tightened boundary + decide a DIFFERENT mechanism vs accept the bounded two-retirement contamination-EXPOSED-HumanEval-family claim.",
      "commits": [
        "73a0212",
        "ed3ba63",
        "88ed03b"
      ],
      "docs": [
        "docs/RUNBOOK_W110.md",
        "docs/RESULTS_W110_BIGCODEBENCH_PHASE2_70B_V1.md",
        "docs/RESULTS_W110_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W110_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W110_V1.md",
        "results/w110/bigcodebench_preflight/preflight_verdict.json"
      ],
      "linear_issues": [
        "COO-34",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_added": [
        "W110-T-BIGCODEBENCH-REAL-DATA-FETCH-PINNED",
        "W110-T-BIGCODEBENCH-SECOND-RESISTANT-PREFLIGHT-EARNED",
        "W110-T-CONTAMINATION-CONFOUND-STRENGTHENED-NOT-PROVEN",
        "W110-T-BIGCODEBENCH-EXECUTOR-V1-HEADLESS-AGG-FIX",
        "W110-L-BIGCODEBENCH-REFLEXION-PHASE2-70B-CAP",
        "W110-L-REFLEXION-FAILS-ON-CONTAMINATION-RESISTANT-CODE-GENERALLY-CAP",
        "W110-L-BIGCODEBENCH-GOLD-GREEN-WALL-STABILITY-GUARD-CAP",
        "W110-L-BIGCODEBENCH-RELEASE-DATE-RESISTANCE-NOT-CONTEST-DATE-CAP",
        "W110-L-BIGCODEBENCH-EXECUTOR-V1-EXEC-NAMESPACE-NOT-FILE-MODULE-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W111",
      "title": "Different-MECHANISM tournament on contamination-resistant code + bounded-claim fallback EARNED (M3 executor-grounded patcher sub-reflexion; resistant ceiling NOT reflexion-specific)",
      "outcome": "Gated branch milestone with THREE lanes, NOT a new benchmark tournament and NOT a reflexion rerun. docs/RUNBOOK_W111.md locked + committed (65274c4) BEFORE the only NIM call. After W110's resistant FAIL (reflexion 0/2), W111 asked the sharper question: is the resistant ceiling reflexion-SPECIFIC, or does it hold for ANY same-budget mechanism at 70B? Lane alpha (MAIN): a $0-NIM mechanism-mining census re-executed all 300 W110 BigCodeBench A1+B candidates (scripts/mine_w111_resistant_failure_modes_v1.py) -> resistant failure distribution = 81.6% SEMANTIC hidden-test-coupling / 1.8% API-grounding (both API failures on /51, already rescued); hard-core (8 both-A1+B-fail) = 6/8 mock-coupling (fix needs hidden test source = unreachable in a fair regime) + 2/8 output-value. This KILLED M2 (tool-augmented local symbol/doc introspection; attacks 1.8%) and M1 (library/spec planner; attacks comprehension, sacrifices a self-consistency sample, no executor grounding) AT $0 NIM, and admitted M3 (executor-grounded structured-failure patcher: typed expected/actual contract + minimal-patch on the latest candidate, materially different from prose reflexion, NEVER the hidden test source; coordpy.executor_grounded_patcher_v1; 9 tests incl. a test-source-non-leak fairness guard). The one earned NIM run was a smallest-decisive 143-call M3 probe on the pinned rescue-CONCENTRATED 13-problem hard-core slice (CID b611fae0; seed 111001; bench Merkle 70353e77; ~69min, heavy 429 throttling survived): A0=30.77 / A1=30.77 / M3=46.15 %; M3-A1=+15.38pp (rescue-concentrated UPPER BOUND); MLB-1=61.5%, MLB-2=12.5%. M3's patch loop rescued ONE hard-core problem reflexion B FAILED (/13) -- a genuine mechanism win -- but at 12.5% (1/8), BELOW reflexion's 25% and the 33% floor; the other M3 win (/20) is an attempt-0 SAMPLING win, not the mechanism; M3 did not hold reflexion's /51 rescue; 0 regressions vs A1. Literal pre-committed verdict = AMBIGUOUS (rescued 1 OUTPUT_VALUE but didn't hold /51); resolved: M3 did NOT earn the fair 30-slice pilot (EARN bar unmet + margin NON-mechanism-driven [MLB-2 12.5% < floor] + W104->W105 rescue-concentrated erosion + W106 margin-cap discipline => fair-pilot verdict-changing power LOW => NOT WARRANTED). Lane beta: the bounded-claim fallback rule was pre-committed (RUNBOOK section 6, LAST resort) and is now EARNED; no NIM on M1/M2/fair-pilot/APPS/reflexion/405B/Llama-3.1; BigCodeBench-primary / LiveCodeBench-secondary cross-check rule held (secondary not triggered -- gated on an M3 fair-slice PASS that did not happen). Lane gamma: graphify refreshed from HEAD at start (d41265d5) + re-ingested the M3 module mid-milestone (bench node degree 23) + refreshed at end; explain/path/affected/query run (M3 = 3-hop structural sibling of the W110 reflexion bench). BOUNDED-CLAIM FALLBACK EARNED: the resistant ceiling is NOT reflexion-specific (two mechanisms now fail to beat A1 on resistant code at 70B -- reflexion 0/2 + M3 sub-reflexion); the bounded two-retirement contamination-EXPOSED-HumanEval-family-at-70B claim is the honest code ceiling. W89 + W105 STAND. W111 adds NO retirement, retires NO carry-forward, and proves NOTHING about the confound (it tests a mechanism, not the confound); honest positive registered: M3 is not vacuous (it rescued /13), just sub-reflexion. 21st consecutive preflight-discipline validation (W93-W111). Stable boundary preserved (no version bump; no PyPI; coordpy/__init__.py untouched; 1 new explicit-import-only module + 2 scripts + 9 tests). COO-9 REMAINS lead; W112 = cross-scale-UP probe on a stronger code model if reachable / a NIM-free-earned M3 strengthening that clears the 33% floor / acceptance of the bounded claim.",
      "commits": [
        "65274c4",
        "866528b"
      ],
      "docs": [
        "docs/RUNBOOK_W111.md",
        "docs/RESULTS_W111_M3_PATCHER_PROBE_70B_V1.md",
        "docs/RESULTS_W111_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W111_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W111_V1.md",
        "results/w111/mechanism_mining/w110_bcb_failure_census.json"
      ],
      "linear_issues": [
        "COO-35",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_added": [
        "W111-T-RESISTANT-FAILURE-IS-SEMANTIC-HIDDEN-TEST-COUPLING",
        "W111-T-EXECUTOR-GROUNDED-PATCHER-V1-SHIPS",
        "W111-L-NO-DIFFERENT-MECHANISM-BEATS-A1-ON-RESISTANT-CODE-AT-70B-CHEAP-CAP",
        "W111-L-M3-PATCHER-SUB-REFLEXION-ON-RESISTANT-HARD-CORE-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W112",
      "title": "Stronger-model resistant-code gate + NIM-free M3-strengthening structural kill (Llama-4-Maverick BigCodeBench +10pp but model-EXPOSED, NOT a clean resistant reopening; contamination-resistance is model-cutoff-relative)",
      "outcome": "Gated branch milestone with THREE lanes, NOT a new benchmark tournament and NOT a 70B rerun. docs/RUNBOOK_W112.md locked BEFORE any NIM call (incl. the sub-second reachability sweep); the section 1-alpha target-selection rule was locked BEFORE probing so the catalogue could not be target-shopped. COO-9 stays lead; ultracode OFF. Lane alpha (stronger-model gate, LIVE): an honest NIM reachability sweep over the live 118-model catalogue (scripts/run_w112_stronger_model_reachability_sweep_v1.py; decision CID a654956b) found meta/llama-3.1-405b-instruct HTTP 404 for the 6th consecutive time (W104-W108, W112) but meta/llama-4-maverick-17b-128e-instruct HTTP 200 -- the first strictly-stronger, same-budget-comparable, NON-reasoning code model reachable since the cross-scale-UP axis opened (3 more eligible tier-2: Qwen3-Coder-480B, DeepSeek-V4-pro, Mistral-Small-4-119B). Selected Maverick tier-1 per the locked Llama-family rule. The §1alpha-earn canary (2 problems, 22 calls) confirmed the plain code path => the smallest honest BigCodeBench pilot was EARNED + RUN on the EXACT W110 fair 30-slice (CID b69bf3a0 re-derived + matched; only the model changed; 1 seed x 30 x K=5 = 330 calls; reflexion B): A0 73.33 / A1 73.33 / B 83.33 %; B-A1 = +10.00 pp; 9/9 core Phase-2 gates; MLB-2 = 37.5% PASS; MLB-1 = 26.67% FAIL => PASS_NON_MECHANISM_DRIVEN (3 clean rescues /15,/26,/51; 0 regressions); the reflexion margin REOPENED vs 70B's +0.00 pp (W110). DECISIVE CAVEAT (grounded): contamination-resistance is MODEL-CUTOFF-RELATIVE -- BigCodeBench 2024-06 is resistant for Llama-3.3-70B (~2024-01 cutoff) but EXPOSED for Llama-4-Maverick (Aug-2024 pretraining cutoff > release); the result is a structural twin of the W109 APPS contamination-EXPOSED control (A0 73.33% IDENTICAL; same MLB-1-fail/MLB-2-pass/PASS_NON_MECHANISM_DRIVEN shape), and the IDENTICAL slice flips +0.00pp -> +10.00pp as the model cutoff crosses the release date (first WITHIN-benchmark resistant->exposed flip) => EXPOSED-column result, NOT a clean reopening of contamination-RESISTANT superiority. Lane beta (NIM-free M3 strengthening, killed at $0): a harder fair-reachability re-mining (scripts/mine_w112_fair_reachability_v1.py re-executed the W110 transcripts) showed that on the MLB-2 denominator (12 invoked) the reliably fair-reachable ceiling is 8.3% (1/12) and the best-conceivable bound 33.3% (4/12; 58% mock/fixture-coupled) merely TOUCHES the floor; all four fair strengthenings (richer typed digest / multi-candidate aggregation / patch-rejection / doctest invariants) killed at $0 -- none expands the reliably-reachable set => NO_FAIR_STRENGTHENING_CAN_CLEAR_FLOOR (structurally strengthens the W111 empirical sub-floor finding). Lane gamma: graphify refreshed start (HEAD 2985b55) -> after adding the 2 W112 scripts (75,728 nodes / 241,040 edges) -> at end; explain/path/affected on run_executor_grounded_patcher_bench_v1; explain on the new mining script confirmed fair-regime reuse. Claim surfaces tightened (registry / status / honesty / consolidated narrative / CHANGELOG / contamination framing). NET: TWO confirmed retirements (W89 +5.56pp, W105 +7.00pp) STAND, contamination-EXPOSED-HumanEval-family-at-70B; W112 adds NONE and retires NONE; resistant superiority still unproven; the contamination-confound is STRENGTHENED a 3rd time (3 exposed margins incl. a within-benchmark flip vs 0/2 resistant) but NOT proven. 22nd consecutive preflight/earn-discipline validation (W93-W112). Stable boundary preserved (no version bump; no PyPI; coordpy/__init__.py untouched; ZERO new coordpy.* modules -- 2 new scripts only, the pilot reused the W110 driver verbatim; 26 reused-module tests pass). COO-9 REMAINS lead; W113 = a benchmark verifiably contamination-resistant FOR Llama-4 (problem dates > Aug-2024; date-filtered LiveCodeBench) to separate capability from exposure.",
      "commits": [
        "a1b7d78"
      ],
      "docs": [
        "docs/RUNBOOK_W112.md",
        "docs/RESULTS_W112_STRONGER_MODEL_BIGCODEBENCH_PILOT_V1.md",
        "docs/RESULTS_W112_STRONGER_MODEL_GATE_AND_SELECTION_V1.md",
        "docs/RESULTS_W112_FAIR_REACHABILITY_M3_STRENGTHENING_V1.md",
        "docs/RESULTS_W112_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W112_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W112_V1.md",
        "results/w112/fair_reachability/w110_bcb_fair_reachability.json"
      ],
      "linear_issues": [
        "COO-36",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_added": [
        "W112-T-405B-GATE-SIXTH-404-CLOSED",
        "W112-T-STRONGER-CODE-MODEL-REACHABLE-LLAMA4-MAVERICK",
        "W112-T-STRONGER-MODEL-BIGCODEBENCH-MARGIN-REOPENS-BUT-MODEL-EXPOSED",
        "W112-T-CONTAMINATION-RESISTANCE-IS-MODEL-CUTOFF-RELATIVE",
        "W112-T-FAIR-M3-STRENGTHENING-CEILING-SUB-FLOOR",
        "W112-L-STRONGER-MODEL-RESISTANT-SUPERIORITY-NOT-CLEANLY-DEMONSTRATED-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W113",
      "title": "Clean resistant-FOR-Llama-4 benchmark construction + earned Maverick pilot => exposure confirmed (the W112 +10pp was contamination exposure; B-A1 collapses +10.00 exposed -> +0.00 resistant at the same Maverick scale; resistant superiority 0 clean across BOTH scales; W89+W105 stand; no retirement added)",
      "outcome": "Gated branch milestone with THREE lanes, NOT a new tournament and NOT a confounded exposed rerun. docs/RUNBOOK_W113.md locked BEFORE any NIM call (incl. the canary). ultracode OFF; COO-9 stays lead; $0 expensive run except the one earned pilot. Lane alpha (clean resistant main lane, LIVE): built the machine-checkable resistance rule FIRST (coordpy.livecodebench_resistant_slice_v1: RESISTANT-for-Maverick iff contest_date > 2024-08-31, the entire ambiguous August-2024 window EXCLUDED; a KNOWN/ESTIMATED/UNKNOWN cutoff registry certifies resistance only against a KNOWN cutoff); proved date integrity NIM-free (scripts/run_w113_resistant_slice_preflight.py; verdict CID 6f30990c) -- the SHA-pinned release_v6 functional subset is 63/63 resistant (dates 2025-01-11..2025-04-05; 0 missing / 0 unparseable / 0 in-August) and the deterministic resistant 30-slice CID 2afc318c == the EXACT W108 slice (the date filter did not perturb the set => clean cross-scale; only model scale varies vs W108's 70B run); EARNED + RAN the cheapest honest Maverick pilot (scripts/run_w113_resistant_pilot.py; meta/llama-4-maverick-17b-128e-instruct; 1 seed x 30 x K=5 = 330 calls; ~55 min) => A0 30.00 / A1 50.00 / B 50.00 %; B-A1 = +0.00 pp; 7/9 gates; MLB-1 = 63.33 % PASS; MLB-2 = 21.05 % FAIL => FAIL => EXPOSURE_CONFIRMED (reflexion genuinely invoked, MORE than 70B; rescued 4 / regressed 2 / net 0; interp CID aa324208). The W112 +10.00pp on EXPOSED BigCodeBench COLLAPSED to +0.00pp on resistant LiveCodeBench at the SAME scale => the +10pp was contamination EXPOSURE, not a capability reopening -- a within-MODEL exposed->resistant flip (sharpest contamination dissociation yet; corroborated by Maverick resistant A1 50 % < 70B 63.33 %). The clean 2x2 is complete; the RESISTANT column is 0 clean across BOTH scales (70B -3.33 / +0.00; Maverick +0.00). Lane beta (tier-2 readiness, NIM-free, $0): coordpy.tier2_readiness_v1 + scripts/run_w113_tier2_readiness_v1.py lock the ranking + same-filtered-slice applicability + spend rule; all three reachable tier-2 stronger models (Qwen3-Coder-480B, DeepSeek-V4-pro, Mistral-Small-4-119B) have UNKNOWN 2025-2026 cutoffs => NONE certifiably resistant on the pinned 2025 slice => tier-2 spend BLOCKED on a missing instrument under EVERY outcome (next = release_v7+). Lane gamma (graphify): refreshed from HEAD at start (0 token cost) + close (76,435 nodes); explain/path/affected/query used for file selection + dependency checks + claim-surface location; claim surfaces tightened. NET: the two retirements (W89, W105) STAND, contamination-EXPOSED-HumanEval-family-at-70B; W113 adds NO retirement; resistant superiority 0 clean across BOTH scales; contamination-confound STRENGTHENED a fourth time (within-model flip) but NOT proven; W114 = accept the bounded contamination-EXPOSED claim + pursue a GENUINELY DIFFERENT axis; 3 new explicit-import-only modules + 3 scripts + 31 tests; pilot reused the canonical evaluator + W108 NIM generator (namespace import); no version bump; no PyPI; coordpy/__init__.py untouched; 23rd consecutive preflight/earn-discipline validation (W93-W113).",
      "commits": [
        "e98ffb6"
      ],
      "docs": [
        "docs/RUNBOOK_W113.md",
        "docs/RESULTS_W113_RESISTANT_PILOT_V1.md",
        "docs/RESULTS_W113_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W113_V1.md",
        "results/w113/resistant_slice_preflight/preflight_verdict.json",
        "results/w113/tier2_readiness/tier2_readiness.json"
      ],
      "linear_issues": [
        "COO-37"
      ],
      "carry_forwards_added": [
        "W113-T-RESISTANT-FOR-LLAMA4-SLICE-CONSTRUCTED",
        "W113-T-STRONGER-MODEL-RESISTANT-CODE-EXPOSURE-CONFIRMED",
        "W113-T-CONTAMINATION-CONFOUND-STRENGTHENED-WITHIN-MODEL",
        "W113-L-STRONGER-MODEL-RESISTANT-SUPERIORITY-FAIL-CAP",
        "W113-L-TIER2-NO-CERTIFIABLE-RESISTANT-SLICE-ON-PINNED-CORPUS-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W114",
      "title": "Bounded-ceiling registration + newer-certifiable-post-cutoff-instrument attempt + per-model certification layer => NO stronger-than-Maverick model certifiable on the latest real data ($0 NIM); W89+W105 stand; no retirement added",
      "outcome": "Gated branch milestone with THREE lanes, NOT another exposed rerun and NOT another same-scale resistant reflexion rerun. docs/RUNBOOK_W114.md locked BEFORE any NIM (the no-go branch is pre-committed by the rule => $0 spend is discipline, not omission). ultracode OFF; COO-9 stays lead. Lane alpha (bounded-ceiling registration, NIM-free): registered the W113 outcome as the honest code-superiority FLOOR across the canonical truth surfaces (RESEARCH_STATUS Last-touched; THEOREM_REGISTRY Last-touched/claim-row/new W114 section; CONSOLIDATED narrative; HOW_NOT_TO_OVERSTATE; new CONTAMINATION_CONTROL_FRAMING_W114; CHANGELOG) -- exactly TWO retirements (W89 +5.56pp, W105 +7.00pp), both meta/llama-3.3-70b-instruct @ 70B contamination-EXPOSED-HumanEval-family; resistant superiority 0 clean across BOTH 70B and Maverick scales; the W112 +10pp was EXPOSURE (W114-T-BOUNDED-EXPOSED-CODE-CEILING-REGISTERED); the truth floor a new axis must beat, NOT surrender (README untouched -- no per-milestone research claims). Lane beta (genuinely different axis = certification-supply analysis): verified from PRIMARY sources (2026-05-29) the latest LiveCodeBench release (HF file tree: test6.jsonl is the highest-numbered => release_v6 latest, no test7+; FUNCTIONAL subset 63 problems, all 2025-01-11..2025-04-05, frontier 2025-04-05) AND official model cutoffs (Llama-4-Maverick = Aug-2024 KNOWN per the official Llama-4 model card; Qwen3-Coder-480B = NO CUTOFF STATED on the official HF card + Qwen blog; DeepSeek-V4-pro = none disclosed; Mistral-Small-4-2603 = none in official Mistral docs, 2026-03 release post-dates the whole window) => verdict NO_CERTIFIABLE_STRONGER_MODEL, $0 NIM: a >=30 functional resistant slice requires a KNOWN cutoff <= ~Jan-2025; Maverick (the only reachable KNOWN cutoff) is already SETTLED (W113 resistant FAIL => redundant); every stronger-than-Maverick frontier model has an OFFICIALLY UNDISCLOSED cutoff (C1) and, where estimable, a cutoff at/after the Apr-2025 frontier (C2) => the resistant-instrument frontier has AGED OUT relative to the reachable model frontier (W114-L-RESISTANT-INSTRUMENT-FRONTIER-LAGS-MODEL-FRONTIER-CAP + W114-T-STRONGER-MODEL-CUTOFFS-OFFICIALLY-UNDISCLOSED). Lane gamma (per-model certification layer + graphify): built coordpy.stronger_model_cutoff_certification_v1 (explicit-import-only; imports the W113 registry + partition_resistant_v1 + ranking, no duplication; C1^C2^C3^C4 gate; verified instrument-frontier record + per-model official-source provenance + W113<->W114 confidence-consistency guard) + scripts/run_w114_stronger_model_certification_v1.py (re-verifies the month histogram against the SHA-pinned bb4c364f corpus: sha_ok + histogram_match; decision CID 258b6ed7); graphify refreshed from HEAD at start + close (76,573 nodes); explain/path/affected/query used for file selection + dependency checks. NET: W89 + W105 STAND, W114 adds NO retirement + retires NO carry-forward; contamination-confound UNCHANGED (W114 tests certification supply, not the confound); reachability NOT re-probed (not the binding gate; W112 facts carried); COO-9 stays lead; W115 = fires only when a resistant FUNCTIONAL instrument with >=30 problems dated strictly after a reachable frontier model's KNOWN cutoff exists. 1 new explicit-import-only module + 1 script + 12 new tests (+ 30 W113 regression tests pass); ZERO additions to coordpy/__init__.py; no version bump; no PyPI; 24th consecutive preflight/earn-discipline validation (W93-W114).",
      "commits": [
        "76f8b0a"
      ],
      "docs": [
        "docs/RUNBOOK_W114.md",
        "docs/RESULTS_W114_STRONGER_MODEL_CERTIFICATION_V1.md",
        "docs/RESULTS_W114_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W114_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W114_V1.md",
        "results/w114/certification/certification_verdict.json"
      ],
      "linear_issues": [
        "COO-38"
      ],
      "carry_forwards_added": [
        "W114-T-BOUNDED-EXPOSED-CODE-CEILING-REGISTERED",
        "W114-T-STRONGER-MODEL-CUTOFFS-OFFICIALLY-UNDISCLOSED",
        "W114-T-STRONGER-MODEL-CERTIFICATION-V1-SHIPS",
        "W114-L-RESISTANT-INSTRUMENT-FRONTIER-LAGS-MODEL-FRONTIER-CAP",
        "W114-L-NO-NEW-STRONGER-MODEL-RESISTANT-PILOT-EARNED-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W115",
      "title": "External-frontier refresh (LIVE) + durable future-fire certification/instrument-supply pipeline => external frontier UNCHANGED, NO certifiable stronger model ($0 NIM); DeepSeek V4 card now exists (2026-04-27) but discloses no cutoff; W89+W105 stand",
      "outcome": "Gated branch milestone with THREE lanes, NOT another exposed rerun and NOT another same-scale resistant reflexion rerun on the same release_v6 slice. Executes the W114 RUNBOOK_W114.md s9 NO_CERTIFIABLE_STRONGER_MODEL branch (COO-38 close). docs/RUNBOOK_W115.md locked BEFORE any NIM (the no-go branch is pre-committed by the rule => $0 spend is discipline, not omission). ultracode OFF; COO-9 stays lead. Lane alpha (LIVE external-frontier refresh): re-verified from PRIMARY sources (2026-05-29; WebSearch/WebFetch against official HF dataset/model cards + vendor PDFs) that the latest LiveCodeBench release is STILL release_v6 (HF file tree highest = test6.jsonl, 'add v6' ~1yr ago; NO test7+; functional frontier 2025-04-05 UNCHANGED) and that NO reachable model stronger than Maverick has a KNOWN cutoff <= ~Jan-2025 (Qwen3-Coder-480B official HF card states NO cutoff; the DeepSeek V4 official model card NOW EXISTS [published 2026-04-27, Pro=1.6T/49B] but contains NO 'cutoff' string and no training-data date => still UNKNOWN; Mistral-Small-4 UNKNOWN; Maverick Aug-2024 KNOWN but already SETTLED at W113). Both binding conditions still fail => verdict re-derives NO_CERTIFIABLE_STRONGER_MODEL (decision CID 258b6ed7, byte-identical to W114), $0 NIM; the one external change since W114 (DeepSeek V4 card publication) does NOT move the verdict (W115-L-EXTERNAL-FRONTIER-UNCHANGED-NO-CERTIFIABLE-SLICE-REVERIFIED-CAP). Reachability NOT re-probed (not the binding gate; W112 facts carried). Lane beta (future-fire certification/instrument-supply pipeline): shipped coordpy.frontier_certification_pipeline_v1 (explicit-import-only; reuses the W113 registry + the W114 certify_model_v1/decide_certification_v1 gate + the loader's LIVECODEBENCH_KNOWN_RELEASES, no duplication -- graphify-confirmed 2-hop reuse chain run_frontier_certification_v1 -> decide_certification_v1 -> certify_model_v1) + scripts/run_w115_frontier_certification_v1.py: a latest-official-release detector (newer_release_available=False), a generalised frontier-date summary + threshold table (max KNOWN cutoff month admitting a >=30 slice = 2025-01), a per-model go/no-go matrix, a disclosure-consistency guard (live disclosures vs the encoded registry -- a divergence is the W116 update signal), and a structured W116FireConditionV1 -- all driven by a FrontierSnapshotV1 (external state as DATA) so W116 is push-button (W115-T-FUTURE-FIRE-CERTIFICATION-PIPELINE-SHIPS); the script re-verifies the histogram against the SHA-pinned bb4c364f corpus (sha_ok + histogram_match); result CID 6890419c; artifact results/w115/frontier_certification/frontier_certification_verdict.json. Lane gamma (claim/graphify/readiness): graphify refreshed from HEAD at start (f8b085d) + re-ingested the new module/script (76,727 nodes) + close; explain/path/affected/query used for file selection + dependency checks; truth surfaces tightened (RESEARCH_STATUS Last-touched; THEOREM_REGISTRY Last-touched/claim-row/new W115 section; CONSOLIDATED narrative; HOW_NOT_TO_OVERSTATE; new CONTAMINATION_CONTROL_FRAMING_W115; new FRONTIER_RELEVANCE_AUDIT_W115; CHANGELOG). NET: W89 + W105 STAND, W115 adds NO retirement + retires NO carry-forward; contamination-confound UNCHANGED (W115 tests certification supply, not the confound); COO-9 stays lead; COO-39; W116 fires when the pipeline trigger flips (a newer admitted release_v7+ with >=30 post-frontier functional problems for a KNOWN-cutoff stronger model, OR a reachable stronger-than-Maverick model disclosing a KNOWN cutoff month <= 2025-01). 1 new explicit-import-only module + 1 script + 10 new tests (+ 42 W113/W114 regression tests pass); ZERO additions to coordpy/__init__.py; no version bump; no PyPI; 25th consecutive preflight/earn-discipline validation (W93-W115).",
      "commits": [
        "84a9009"
      ],
      "docs": [
        "docs/RUNBOOK_W115.md",
        "docs/RESULTS_W115_FRONTIER_CERTIFICATION_V1.md",
        "docs/RESULTS_W115_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W115_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W115_V1.md",
        "results/w115/frontier_certification/frontier_certification_verdict.json"
      ],
      "linear_issues": [
        "COO-39"
      ],
      "carry_forwards_added": [
        "W115-T-FUTURE-FIRE-CERTIFICATION-PIPELINE-SHIPS",
        "W115-L-EXTERNAL-FRONTIER-UNCHANGED-NO-CERTIFIABLE-SLICE-REVERIFIED-CAP",
        "W115-L-NO-NEW-PILOT-EARNED-FRONTIER-UNCHANGED-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W116",
      "title": "Upstream instrument-supply ATTACK (LIVE, four upstream surfaces) + primary-source model-cutoff ATTACK + durable upstream-ADMISSION pipeline => NO admissible new instrument beyond release_v6, NO certifiable stronger model ($0 NIM); Mistral-Small-4-119B-2603 CONFIRMED REAL with a primary card disclosing no cutoff; W89+W105 stand; no retirement added",
      "outcome": "Gated branch milestone with THREE lanes, NOT another exposed rerun and NOT another same-scale resistant reflexion rerun on the same release_v6 slice. Executes the W115 RUNBOOK_W115.md s9 NO_CERTIFIABLE_STRONGER_MODEL branch (COO-39 close): instead of passively waiting for release_v7, ATTACK the upstream instrument supply + the model-cutoff disclosure side, operationalise the next clean shot, and run a pilot only if certification is genuinely clean. docs/RUNBOOK_W116.md locked BEFORE any NIM (the no-go branch is pre-committed by the rule => $0 spend is discipline, not omission). ultracode OFF; COO-9 stays lead. Lane alpha (LIVE upstream instrument-supply attack): went one level UPSTREAM of release_v6 and re-verified the LiveCodeBench frontier at FOUR authoritative surfaces -- (1) HF code_generation_lite file tree (highest test6.jsonl, no test7+, lastModified 2025-06-05); (2) loader code_generation_lite.py ALLOWED_FILES (v_list=[v1..v6], release_latest -> release_v6 files, DEFAULT_CONFIG_NAME=release_latest); (3) full code_generation dataset README (release_v6 = May 2023..Apr 2025, 1055 problems, frontier 2025-04-05); (4) GitHub repo README (tops out at release_v6, no v7 tag) => NO admissible new instrument beyond release_v6; the functional frontier is conclusively 2025-04-05. The 'planned v7' mentioned only by a non-primary search summary is INADMISSIBLE under the pre-committed A1..A5 rule (no artifact, no SHA) and is recorded only as the W117 watch signal (W116-L-UPSTREAM-SUPPLY-NO-ADMISSIBLE-NEW-INSTRUMENT-FOUR-SURFACE-CAP). Lane beta (primary-source model-cutoff attack): re-checked official cutoffs from PRIMARY sources => no reachable stronger-than-Maverick model has a primary-KNOWN cutoff <= 2025-01: Qwen3-Coder-480B + DeepSeek-V4-pro UNKNOWN (official cards: NO CUTOFF STATED); mistralai/mistral-small-4-119b-2603 now CONFIRMED REAL (119B MoE, 128 experts/4 active, released 2026-03-16) with its official Mistral docs model card + official announcement (mistral.ai/news/mistral-small-4) disclosing NO cutoff -- the only figure is a non-primary aggregator (OpenRouter '2025-06') that is itself C2-exposed (post-dates the Apr-2025 frontier); Mistral-Small-3.2-24B KNOWN ~Oct-2023 but sub-70B (C3) (W116-T-MISTRAL-SMALL-4-CONFIRMED-REAL-PRIMARY-NO-CUTOFF). Lane gamma (upstream-admission pipeline + graphify): shipped coordpy.upstream_instrument_admission_v1 (explicit-import-only; reuses the W113 registry + the W114 certify_model_v1 gate + the W115 run_frontier_certification_v1 pipeline + the loader's LIVECODEBENCH_KNOWN_RELEASES, no duplication -- graphify-confirmed reuse chain run_upstream_admission_v1 -> run_frontier_certification_v1) + scripts/run_w116_upstream_admission_v1.py: a pre-committed A1..A5 admissibility rule (REFUSES aggregator/mirror/website-intro/rumor instruments), a multi-surface upstream-change detector (detect_upstream_change_v1; richer than W115's single boolean), a certifiable-slice builder, a four-way disclosure-status matrix (KNOWN/ESTIMATED-but-unusable/UNKNOWN/contradictory-or-stale), and a structured W117FireConditionV1 -- all driven by a UpstreamSupplySnapshotV1 so W117 is push-button (W116-T-UPSTREAM-ADMISSION-PIPELINE-SHIPS); the script re-verifies the histogram against the SHA-pinned bb4c364f corpus (sha_ok + histogram_match), asserts the decision-CID-drift guard, and emits results/w116/upstream_admission/upstream_admission_verdict.json (result CID 193164c4; decision CID 258b6ed7 byte-identical to W114/W115). graphify refreshed from HEAD at start (5b3f75d) + re-ingested the new module/script (77,010 nodes) + close; explain/path/affected/query used for file selection + dependency checks. NET: verdict re-derives NO_CERTIFIABLE_STRONGER_MODEL, $0 NIM, no pilot earned (a pilot needs BOTH an admissible new instrument AND a non-redundant primary-KNOWN-cutoff stronger model -- neither exists; n_admissible_new_instruments=0; W117 fires_now=False); W89 + W105 STAND, W116 adds NO retirement + retires NO carry-forward; contamination-confound UNCHANGED (W116 tests certification supply, not the confound); reachability NOT re-probed (not the binding gate; W112 facts carried); COO-9 stays lead; COO-40; W117 fires the moment detect_upstream_change_v1 flags an admissible change (a newer admitted release_v7+ / release_latest re-point / new upstream functional dataset with >=30 post-frontier problems for a KNOWN-cutoff stronger model, OR a reachable stronger-than-Maverick model disclosing a primary-KNOWN cutoff month <= 2025-01). 1 new explicit-import-only module + 1 script + 16 new tests (+ 52 W113/W114/W115 regression tests pass = 68 total); ZERO additions to coordpy/__init__.py; no version bump; no PyPI; 26th consecutive preflight/earn-discipline validation (W93-W116).",
      "commits": [
        "14ad122"
      ],
      "docs": [
        "docs/RUNBOOK_W116.md",
        "docs/RESULTS_W116_UPSTREAM_ADMISSION_V1.md",
        "docs/RESULTS_W116_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W116_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W116_V1.md",
        "results/w116/upstream_admission/upstream_admission_verdict.json"
      ],
      "linear_issues": [
        "COO-40"
      ],
      "carry_forwards_added": [
        "W116-T-UPSTREAM-ADMISSION-PIPELINE-SHIPS",
        "W116-T-MISTRAL-SMALL-4-CONFIRMED-REAL-PRIMARY-NO-CUTOFF",
        "W116-L-UPSTREAM-SUPPLY-NO-ADMISSIBLE-NEW-INSTRUMENT-FOUR-SURFACE-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W117",
      "title": "Upstream-DERIVED instrument CONSTRUCTION attack (LIVE, eight provenance surfaces) + deeper primary-source model-cutoff attack + durable upstream-DERIVED construction/admission pipeline => NO construction-admissible instrument can be built from authoritative provenance beyond release_v6, NO certifiable stronger model ($0 NIM); W89+W105 stand; no retirement added",
      "outcome": "Gated branch milestone with THREE lanes, NOT another exposed rerun, NOT a packaged-release watch-and-wait check, NOT another same-scale resistant reflexion rerun. Executes the W116 RUNBOOK_W116.md s9 branch (COO-40 close). W116 answered 'is there an admissible PACKAGED release_v7?' (No); W117 escalates to the CONSTRUCTION question: even before release_v7 is packaged, can a post-cutoff functional instrument be CONSTRUCTED from official upstream PROVENANCE? docs/RUNBOOK_W117.md locked BEFORE any NIM (the no-go branch is pre-committed by the rule => $0 spend is discipline, not omission). ultracode OFF; COO-9 stays lead. Lane alpha (LIVE upstream-derived construction attack): attacked the upstream CONSTRUCTION provenance at EIGHT authoritative surfaces (the revision history + collection mechanism, not just the release label) -- (1) HF dataset commit/revision log (latest data commit 'add v6' 2025-04-21, HEAD 'fix typos' 2025-06-05); (2) HF refs (1 branch/0 tags); (3) HF discussions/PRs (newest = 'LCB pull request' #14 + a v6-size clarification #13, no v7 thread); (4) GitHub commits (newest 2025-07-16 runner-maintenance, no data commit); (5) GitHub tags (0); (6) GitHub repo pipeline structure (lcb_runner/+assets/ only, NO scraper/collection dir); (7) dataset README provenance (LeetCode/AtCoder/Codeforces, documents only loading published releases, NO generation tool/manifest); (8) runner loader code_generation.py (loads HF-only via load_dataset(version_tag=...), no local scraping) => the authoritative construction provenance IS the packaged HF release; LCB publishes no collection pipeline or forward problem-id manifest; no post-v6 LCB-published artifact exists at any surface; the only post-v6 path (raw-contest hand-assembly) is CONSTRUCTION-INADMISSIBLE (refused by the pre-committed B1 authoritative-LCB-provenance + B2 no-operator-curation criteria) => 0 construction-admissible NEW instruments -- sharper than 'no packaged v7': a post-v6 instrument cannot be CONSTRUCTED from authoritative provenance, only hand-curated, which is refused (W117-L-NO-CONSTRUCTION-ADMISSIBLE-POST-V6-INSTRUMENT-EIGHT-SURFACE-CAP + W117-T-LCB-CONSTRUCTION-PROVENANCE-IS-PACKAGED-RELEASE). Lane beta (deeper primary-source model-cutoff attack): no reachable stronger-than-Maverick model has a primary-KNOWN cutoff <= 2025-01 and nothing newly-disclosed since W116 -- Maverick 'August 2024' re-confirmed VERBATIM (Meta MODEL_CARD.md) but C4-settled; Qwen3-Coder-480B UNKNOWN (official HF card raw README: no cutoff); DeepSeek-V4-pro UNKNOWN from primary -- official V4 model-card PDF re-checked still states NO cutoff, the only figure is a non-primary aggregator 'Apr 2026' that is itself C2-exposed, now matching Mistral-Small-4-119B-2603's UNKNOWN-primary + C2-exposed-aggregator pattern (W117-T-DEEPSEEK-V4-PRIMARY-PDF-RECONFIRMED-NO-CUTOFF). Lane gamma (construction/admission pipeline + graphify): shipped coordpy.upstream_derived_instrument_construction_v1 (explicit-import-only; reuses + WRAPS the W116 run_upstream_admission_v1 -> W115 run_frontier_certification_v1 -> W114 certify_model_v1 chain, no duplication; graphify-confirmed run_upstream_construction_v1 --calls--> run_upstream_admission_v1) + scripts/run_w117_upstream_construction_v1.py: the construction rule (A1..A5 reused + B1 + B2), an eight-surface provenance snapshot, a candidate-instrument constructor (raw-contest assembly triply refused; LCB-pipeline template construction-admissible-in-principle but artifact-absent => constructed=False + the EXACT missing artifact named), a provenance validator, a sharpened disclosure matrix, and a structured W118FireConditionV1 (packaged / construction-provenance / cutoff); the script re-verifies the histogram against the SHA-pinned bb4c364f corpus (sha_ok + histogram_match), asserts the decision-CID-drift guard, and emits results/w117/upstream_construction/upstream_construction_verdict.json (result CID c3c60483; decision CID 258b6ed7 byte-identical to W114/W115/W116). graphify refreshed from HEAD at start (dcec243) + re-ingested the new module/script (77,252 nodes) + close; explain/path/affected/query used for file selection + dependency checks. NET: verdict re-derives NO_CERTIFIABLE_STRONGER_MODEL, $0 NIM, no pilot earned (needs BOTH a construction-admissible new instrument AND a non-redundant primary-KNOWN-cutoff stronger model -- neither exists; constructed=False; W118 fires_now=False); W89 + W105 STAND, W117 adds NO retirement + retires NO carry-forward; contamination-confound UNCHANGED (W117 tests construction supply, not the confound); reachability NOT re-probed (not the binding gate; W112 facts carried); COO-9 stays lead; COO-41; W118 fires the moment a packaged release_v7+ is admitted, OR an LCB-PUBLISHED post-v6 construction provenance (a dataset revision/commit/PR with post-2025-04 functional problems, OR a published collection pipeline + problem-id manifest) enables a B1+B2 reproducible >=30-problem post-cutoff slice, OR a reachable stronger-than-Maverick model discloses a primary-KNOWN cutoff month <= 2025-01. 1 new explicit-import-only module + 1 script + 13 new tests (+ 68 W113/W114/W115/W116 regression tests pass = 81 total); ZERO additions to coordpy/__init__.py; no version bump; no PyPI; 27th consecutive preflight/earn-discipline validation (W93-W117).",
      "commits": [
        "113c456"
      ],
      "docs": [
        "docs/RUNBOOK_W117.md",
        "docs/RESULTS_W117_UPSTREAM_CONSTRUCTION_V1.md",
        "docs/RESULTS_W117_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W117_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W117_V1.md",
        "results/w117/upstream_construction/upstream_construction_verdict.json"
      ],
      "linear_issues": [
        "COO-41"
      ],
      "carry_forwards_added": [
        "W117-T-UPSTREAM-DERIVED-CONSTRUCTION-PIPELINE-SHIPS",
        "W117-L-NO-CONSTRUCTION-ADMISSIBLE-POST-V6-INSTRUMENT-EIGHT-SURFACE-CAP",
        "W117-T-LCB-CONSTRUCTION-PROVENANCE-IS-PACKAGED-RELEASE",
        "W117-T-DEEPSEEK-V4-PRIMARY-PDF-RECONFIRMED-NO-CUTOFF"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W118",
      "title": "CoordPy-OWNED post-v6 functional-instrument CONSTRUCTION (LIVE, official Codeforces API) + deeper primary-source model-cutoff attack + durable construction/admission/pilot-readiness pipeline => 894 official post-v6 functional problem IDENTITIES constructed (identity axis SOLVED at scale), but the executable GRADER is ABSENT family-wide (Codeforces/AtCoder/LeetCode), NO certifiable stronger model ($0 NIM); W89+W105 stand; no retirement added",
      "outcome": "Gated branch milestone with THREE lanes, NOT another exposed rerun, NOT a 'check if v7 exists yet' watch-and-wait, NOT another same-scale resistant reflexion rerun, NOT a bounded-context/compaction job. Executes docs/RUNBOOK_W118.md (locked BEFORE any NIM; the no-go branch is pre-committed by the rule => $0 spend is discipline). ultracode OFF; COO-9 stays lead; COO-42 is the W118 issue. W118 escalated W117 from proving no post-v6 instrument can be inherited to BUILDING one from official sources. Lane alpha (LIVE CoordPy-owned construction): a real constructor (coordpy.coordpy_frontier_functional_v1) ran LIVE against the official Codeforces API (contest.list + problemset.problems; raw-fetch SHA b6342fd1...), applying a total deterministic inclusion rule (PROGRAMMING ^ FINISHED ^ contest_date strictly after the release_v6 frontier 2025-04-05) => 894 admitted post-v6 functional problems from 11223 candidates, 2025-04-07..2026-05-30, 130 contests; manifest CID fb4185a6...; the IDENTITY axis (O1..O6) is SOLVED at scale (894 >> MIN_SLICE 30) (W118-T-COORDPY-OWNED-POST-V6-FUNCTIONAL-IDENTITY-CONSTRUCTIBLE). BUT the executable functional GRADER (O7) is ABSENT family-wide: Codeforces API exposes problem metadata only (no test field/endpoint); LeetCode hidden tests deliberately private; AtCoder system tests Dropbox-only (no official API); sample-only grading is non-credible and operator-synthesised tests are operator curation (refused by O5) => identity-admissible but NOT pilot-admissible (W118-L-OFFICIAL-SOURCE-FAMILY-NO-EXECUTABLE-GRADER-CAP). Maverick (KNOWN Aug-2024) has all 894 problems resistant on a genuinely-new instrument it never ran => C1^C2^C3^C4 identity-CERTIFIABLE, blocked ONLY by the missing grader (W118-T-MAVERICK-IDENTITY-CERTIFIABLE-GRADER-BLOCKED). Lane beta (deeper primary-source cutoff attack): DeepSeek V4 PDF re-fetched directly => NO CUTOFF STATED; Maverick 'August 2024' verbatim; Qwen3-Coder-480B + Mistral-Small-4 v26.03 NO CUTOFF STATED; GLM-5 NEWLY NOTED but UNKNOWN-from-primary + C2-exposed + reachability-unverified => nothing newly primary-disclosed since W117 (W118-T-GLM5-NEWLY-NOTED-UNCERTIFIABLE). Lane gamma (construction/admission/pilot-readiness pipeline): shipped coordpy.coordpy_frontier_functional_v1 (explicit-import-only; reuses the W113 registry + W114 certify_model_v1/LatestResistantInstrumentV1/STRONGER_MODEL_CANDIDATES + the W117 run_upstream_construction_v1 + the W116 disclosure types, no duplication; graphify-confirmed reuse edge run_frontier_functional_construction_v1 --calls--> run_upstream_construction_v1) + scripts/run_w118_frontier_functional_construction_v1.py: the O1..O7 instrument rule, the official-source-family grader registry, the deterministic manifest constructor + thin live fetch, the reused C1..C4 + O7 certification, the W118 disclosure matrix, and a structured W119FireConditionV1 with a falsifiability test (W118-T-FRONTIER-FUNCTIONAL-CONSTRUCTION-PIPELINE-SHIPS); the LCB-inherited verdict re-derives NO_CERTIFIABLE_STRONGER_MODEL (decision CID 258b6ed7..., byte-identical to W114/W115/W116/W117; result CID 3ab0d186...; artifacts results/w118/frontier_functional/frontier_functional_verdict.json + coordpy_frontier_functional_v1_manifest.json). graphify refreshed from HEAD at start (32d3498) + re-ingested the new module/script (77572 nodes) + close. NET: the blocker MOVED from W117's 'no post-v6 identities can be constructed' to W118's 'abundant official post-v6 identities (894), no official executable grader'; verdict NO_CERTIFIABLE_STRONGER_MODEL, $0 NIM, no pilot; W89 + W105 STAND, W118 adds NO retirement, retires NO carry-forward; contamination-confound UNCHANGED (W118 tests construction + grader supply, not the confound); reachability NOT re-probed (not the binding gate; W112 facts carried); COO-9 stays lead; W119 fires the moment an OFFICIAL executable per-problem test suite for >=30 post-v6 functional problems appears on a clean official surface (Maverick is already identity-certifiable => a grader alone unlocks the cheapest honest verdict-changing pilot), OR a packaged release_v7+ / LCB-published post-v6 construction provenance appears, OR a reachable stronger-than-Maverick model discloses a primary-KNOWN cutoff <= the manifest frontier. 1 new explicit-import-only module + 1 script + 15 new tests (+ 81 W113/W114/W115/W116/W117 regression tests pass = 96 total); ZERO additions to coordpy/__init__.py; no version bump; no PyPI; 28th consecutive preflight/earn-discipline validation (W93-W118).",
      "commits": [
        "1dbce7f"
      ],
      "docs": [
        "docs/RUNBOOK_W118.md",
        "docs/RESULTS_W118_FRONTIER_FUNCTIONAL_CONSTRUCTION_V1.md",
        "docs/RESULTS_W118_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W118_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W118_V1.md",
        "results/w118/frontier_functional/frontier_functional_verdict.json",
        "results/w118/frontier_functional/coordpy_frontier_functional_v1_manifest.json"
      ],
      "linear_issues": [
        "COO-42"
      ],
      "carry_forwards_added": [
        "W118-T-COORDPY-OWNED-POST-V6-FUNCTIONAL-IDENTITY-CONSTRUCTIBLE",
        "W118-L-OFFICIAL-SOURCE-FAMILY-NO-EXECUTABLE-GRADER-CAP",
        "W118-T-MAVERICK-IDENTITY-CERTIFIABLE-GRADER-BLOCKED",
        "W118-T-FRONTIER-FUNCTIONAL-CONSTRUCTION-PIPELINE-SHIPS",
        "W118-T-GLM5-NEWLY-NOTED-UNCERTIFIABLE"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W119",
      "title": "official ICPC public-package pivot DISSOLVES the W118 grader blocker (real non-LFS data/secret + output validators + 119 accepted refs; NIM-free grader self-test 16/16 PASS) => the executable functional grader W118 proved ABSENT family-wide is now PRESENT + verified on a genuinely-new post-Maverick-cutoff resistant ICPC battlefield; but the post-cutoff resistant pass-fail slice = 24 < MIN_RESISTANT_SLICE 30 blocks both slice-admissibility AND the reused C2 cert gate => 0 certifiable models, NO pilot ($0 NIM); W89+W105 stand; no retirement added",
      "outcome": "Gated branch milestone with THREE lanes (alpha official ICPC public-package construction, beta primary-source stronger-model cutoff attack, gamma package-to-pilot pipeline/graphify/truth), NOT another LCB-packaging/provenance blocker memo, NOT another exposed rerun, NOT a bounded-context/compaction job. Executes docs/RUNBOOK_W119.md; no NIM spent (pilot not earned: 24<30), so the no-go is discipline by the rule. ultracode OFF; COO-9 stays lead; COO-43 is the W119 issue. W119 made the aggressive pivot W118 set up: off the grader-LESS LiveCodeBench source family (Codeforces/AtCoder/LeetCode) onto the official ICPC problem-package family (github.com/icpc) that already ships the executable grader. LANE alpha (LIVE official ICPC public-package construction): a real constructor (coordpy.coordpy_icpc_public_functional_v1; fetch_icpc_package_listing_v1 -> build_icpc_manifest_v1) enumerated the official ICPC GitHub org (33 repos) and applied a total deterministic inclusion rule (official ICPC package ^ dated ^ contest date strictly after Maverick's Aug-2024 cutoff ^ NOT interactive ^ ships a usable grader) => 24 admitted post-cutoff resistant pass-fail problems from 26 candidates across 2 official repos (Rocky Mountain Regional 2025-2026 created 2025-11-13 + 2024-2025 created 2024-12-03; excluded interactive 1 + custom-without-validator 1 + pre-cutoff 0); months 2024-12 (12) + 2025-11 (12); manifest CID 2b337377... The executable GRADER is PRESENT + EXECUTABLE (the W118 family-wide blocker DISSOLVED): the official ICPC packages ship real non-Git-LFS data/secret/*.in+*.ans (verified by direct content fetch) + 8 output validators + 119 accepted reference solutions; a NIM-free grader self-test ran official accepted Python solutions against the official secret cases = 16/16 cases PASS (videogames 8/8 + whattimeisitmrfox 8/8) => a real executable oracle (W119-T-OFFICIAL-ICPC-PACKAGE-FAMILY-SHIPS-EXECUTABLE-GRADER + W119-T-ICPC-GRADER-SELF-TEST-PASSES). The SLICE COUNT is the binding blocker, load-bearing at BOTH levels: 24 < 30 blocks slice-admissibility AND the reused W114 C2 cert gate (needs >=30 resistant after cutoff), so even KNOWN-cutoff Maverick is NOT identity-certifiable on the 24-slice, tier-2 C1-blocked (UNKNOWN cutoffs) => 0 certifiable models; the NWERC-2024 official static-package second surface 404s (W119-L-OFFICIAL-ICPC-RESISTANT-PASSFAIL-SLICE-COUNT-CAP). LANE beta (primary-source cutoff attack, reused W118 matrix): Maverick KNOWN Aug-2024; Qwen3-Coder-480B / DeepSeek-V4-pro / Mistral-Small-4 v26.03 / GLM-5 all NO CUTOFF STATED (primary); {KNOWN: 1, UNKNOWN: 4}; nothing newly primary-disclosed since W118; NVIDIA_API_KEY present => the no-pilot is a clean count-gate no-go, not reachability. LANE gamma (package-to-pilot pipeline): shipped coordpy.coordpy_icpc_public_functional_v1 (explicit-import-only; reuses the W113 registry + W114 certify_model_v1/LatestResistantInstrumentV1/STRONGER_MODEL_CANDIDATES + the W117 run_upstream_construction_v1 + the W116/W118 disclosure types, no duplication; graphify-confirmed reuse edges to run_upstream_construction_v1 + certify_model_v1) + scripts/run_w119_icpc_public_construction_v1.py: the P1..P8 rule, the official ICPC source-family grader registry, the deterministic manifest constructor + thin live GitHub-API fetch, a real fresh-subprocess stdin/stdout executor (run_icpc_stdin_executor_v1), the grader self-test, the reused C1..C4 + grader + slice certification, and a structured W120FireConditionV1 with 2 falsifiability tests; the LCB-inherited verdict re-derives NO_CERTIFIABLE_STRONGER_MODEL (decision CID 258b6ed7..., byte-identical to W114/W115/W116/W117/W118; result CID 577f7633...; artifacts results/w119/icpc_public/icpc_public_verdict.json + coordpy_icpc_public_functional_v1_manifest.json). graphify refreshed start + close (graph content includes the W119 module: coordpy_icpc_public_functional_v1 120 nodes). NET: the blocker MOVED from W118's 'abundant official identities, no official grader' to W119's 'official grader present + self-test-passing, +6 post-cutoff resistant pass-fail tasks short of a clean pilot'; verdict NO_CERTIFIABLE_STRONGER_MODEL, $0 NIM, no pilot; W89 + W105 STAND, W119 adds NO retirement + retires NO carry-forward; contamination-confound UNCHANGED; COO-9 stays lead; W120 fires the moment the post-cutoff resistant pass-fail count reaches 30 on a clean official surface (next official ICPC regional drop / clean official second-surface aggregation - grader already present, so a slice alone unlocks the cheapest honest verdict-changing Maverick pilot), OR a reachable stronger-than-Maverick model discloses a primary-KNOWN cutoff. 1 new explicit-import-only module + 1 script + 18 new tests (110 across W113-W119); ZERO additions to coordpy/__init__.py; no version bump; no PyPI; 29th consecutive preflight/earn-discipline validation (W93-W119).",
      "commits": [
        "73bf3e1"
      ],
      "docs": [
        "docs/RUNBOOK_W119.md",
        "docs/RESULTS_W119_ICPC_PUBLIC_CONSTRUCTION_V1.md",
        "docs/RESULTS_W119_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W119_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W119_V1.md",
        "results/w119/icpc_public/icpc_public_verdict.json",
        "results/w119/icpc_public/coordpy_icpc_public_functional_v1_manifest.json"
      ],
      "linear_issues": [
        "COO-43"
      ],
      "carry_forwards_added": [
        "W119-T-OFFICIAL-ICPC-PACKAGE-FAMILY-SHIPS-EXECUTABLE-GRADER",
        "W119-T-ICPC-GRADER-SELF-TEST-PASSES",
        "W119-L-OFFICIAL-ICPC-RESISTANT-PASSFAIL-SLICE-COUNT-CAP",
        "W119-T-ICPC-PUBLIC-FUNCTIONAL-CONSTRUCTION-PIPELINE-SHIPS"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W120",
      "title": "closed the official-ICPC count gap (RMRC exclusion audit [draftlottery->float] + NEW official surface icpc/na-ecna-archive) => coordpy_icpc_battlefield_v1 = 45 tier-1 pure pass-fail >= MIN_RESISTANT_SLICE 30; grader self-test 165/165 each surface; certified Maverick (KNOWN Aug-2024, reachable; C2 flips 24->45); LOCKED the runbook + ran a clean canary, then RAN the EARNED pilot (Maverick x 30-slice, 330 NIM calls): A0 20.00 / A1 23.33 / B 23.33 %; B-A1 = +0.00 pp; MLB-1 83.33% PASS / MLB-2 8.00% FAIL; 6/9; verdict FAIL => bounded resistant ceiling HOLDS; resistant superiority 0 clean across FOUR settings; W114-W119 'no certifiable resistant instrument' escape CLOSED; NO third retirement (W89+W105 stand); contamination-confound STRENGTHENED not proven",
      "outcome": "Gated branch milestone with THREE lanes (alpha official-ICPC multi-surface count-gap closure, beta primary-source stronger-model certification, gamma executable battlefield-to-pilot infra), executing docs/RUNBOOK_W120.md (LOCKED before any NIM). ultracode OFF; COO-9 stays lead; COO-44 is the W120 issue. NOT another '24<30' memo and NOT an exposed rerun. LANE alpha: Route alpha1 re-derived all 26 RMRC problems from each official problem.yaml+tree (clean correction: W119's draftlottery is float_relative_tolerance 1e-6 = tier-2 float, not pure pass-fail; W119 headline-24 holds, composition sharpened to 22 pure + 1 float + 1 custom-with-validator; exclusions confirmed load-bearing: poetictournament interactive, alwaysknowwhereyourtowelis custom-without-validator). Route alpha2 admitted a NEW official surface icpc/na-ecna-archive (NA East Division 2024-11-11 + 2025-11-10, 25 post-cutoff Kattis packages; grader self-test 6/6 Python-self-testable problems, 149/149 cases PASS incl. valleygulls 40/40 under a deterministic float oracle). Combined coordpy_icpc_battlefield_v1: 51 seen -> 49 admitted -> 45 tier-1 PURE pass-fail >=30 with margin, NO loosening (+3 float +1 custom = 49 gradeable; 2 typed exclusions; snapshot SHA b212866f; manifest CID bf55bb6c; 30-slice CID 01bf9ef8; combined grader self-test 165/165 each surface; dates 2024-11-11..2025-11-13). LANE beta: Maverick KNOWN Aug-2024 re-verified verbatim (Meta llama4 MODEL_CARD.md), reachable in the 118-model NIM catalogue => C2 flips 24->45 => first certifiable on this family; Qwen3-Coder-480B/DeepSeek-V4-pro/Mistral-Small-4/GLM-5 all UNKNOWN-from-primary; {KNOWN:1, UNKNOWN:4}; nothing newly disclosed since W119. LANE gamma: shipped coordpy.coordpy_icpc_battlefield_v1 (multi-surface aggregator + R1..R8 tiered admission + exclusion audit + combined manifest + deterministic float oracle + per-model certification + core-slice selector + W121FireConditionV1) + coordpy.icpc_reflexion_bench_v1 (stdin/stdout A0/A1/B; report-shape compatible with the VERBATIM W108 gate evaluator) + 2 scripts + 13 tests (2 falsifiability); reuses W119 executor + W114 certify_model_v1 + W117 run_upstream_construction_v1 => decision CID 258b6ed7 re-derives byte-identically. PILOT (earned + ran; runbook locked first; canary clean): Maverick x the deterministic tier-1 core 30-slice, 1x30xK5 = 330 NIM calls, ~54 min; grader = official secret cases (token-diff, NO LLM judge); reflexion feedback = public samples + judge verdict + stderr only. A0 20.00 / A1 23.33 / B 23.33 %; B-A1 = +0.00 pp; MLB-1 83.33% (25/30) PASS / MLB-2 8.00% (2/25) FAIL; Phase-2 6/9; verdict FAIL -> BOUNDED_CEILING_HOLDS_ON_RESISTANT_ICPC (reflexion invoked but net-zero: 2 rescues - 1 regression). NET: NO third retirement (W89 +5.56pp + W105 +7.00pp STAND, the only two, contamination-EXPOSED-HumanEval-family @ 70B); resistant superiority now 0 CLEAN across FOUR settings (W108 -3.33 / W110 +0.00 / W113 +0.00 / W120 +0.00); the W114-W119 'no certifiable resistant instrument' escape is CLOSED (built it, certified Maverick, ran it, mechanism did not transfer); contamination-confound STRENGTHENED (4th + cleanest resistant null on a >=30 grader-clean OFFICIAL battlefield) but NOT proven (single seed 120001; ICPC-vs-HumanEval difficulty + Python-TLE floor unexcluded). W121 = accept the bounded resistant ceiling / genuinely different axis (optional cheap multi-seed) OR a reachable stronger-than-Maverick model with a primary-KNOWN cutoff (prefer strongest honest target on this same >=30 battlefield). 30th consecutive preflight/earn-discipline validation (W93-W120). No version bump; no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "0e146c4"
      ],
      "docs": [
        "docs/RUNBOOK_W120.md",
        "docs/RESULTS_W120_ICPC_BATTLEFIELD_V1.md",
        "docs/RESULTS_W120_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W120_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W120_V1.md",
        "results/w120/icpc_battlefield/battlefield_verdict.json",
        "results/w120/icpc_battlefield/battlefield_snapshot.json",
        "coordpy/coordpy_icpc_battlefield_v1.py",
        "coordpy/icpc_reflexion_bench_v1.py",
        "scripts/run_w120_icpc_battlefield_v1.py",
        "scripts/run_w120_icpc_pilot.py",
        "tests/test_w120_icpc_battlefield_v1.py"
      ],
      "linear_issues": [
        "COO-44"
      ],
      "carry_forwards_added": [
        "W120-T-OFFICIAL-ICPC-MULTI-SURFACE->=30-BATTLEFIELD-CONSTRUCTIBLE",
        "W120-T-MAVERICK-CERTIFIABLE-ON->=30-OFFICIAL-ICPC",
        "W120-L-RESISTANT-SUPERIORITY-0-CLEAN-ON-OFFICIAL-ICPC->=30-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W121",
      "title": "matched EXPOSED official-ICPC control on the SAME package family as W120 (RMRC 2021 + ECNA 2022-2023 + RMRC 2022-2023 + ECNA 2023-2024, pre-Aug-2024) + Maverick same-model same-mechanism contrast RAN => coordpy_icpc_exposed_control_v1 = 42 tier-1 pure pass-fail >= 30; grader self-test 30 all-pass / 637 official secret cases each surface; Maverick EXPOSED-certifiable (W114 gate C2->C2e); LOCKED runbook + clean canary, then RAN the EARNED pilot (Maverick x exposed 30-slice, 330 NIM calls): A0 6.67 / A1 26.67 / B 30.00 %; B-A1 = +3.33 pp; MLB-1 93.33% PASS / MLB-2 25.00% FAIL; 8/9; verdict FAIL => the matched-family exposure flip did NOT reproduce the retirement-grade HumanEval-family margins (+5.56/+7.00) => contamination-confound WEAKENED, difficulty/family-ease implicated, bounded ceiling HARDENS to HumanEval-family-(ease/structure)-specific @ 70B; NOT refuted (faint sub-floor exposure-consistent gradient; single seed); NO third retirement (W89+W105 stand); paired seed NOT earned",
      "outcome": "Gated branch milestone with THREE lanes executing docs/RUNBOOK_W121.md (LOCKED before any NIM; the W120 section-10 FAIL-branch 'genuinely different axis'). Removed the last loophole on the W120 resistant null by holding family+difficulty FIXED and flipping ONLY exposure. LANE alpha: SAME two official github.com/icpc families W120 used (na-ecna-archive + na-rocky-mountain-*-public), pre-cutoff editions of the same regionals; byte-identical R1/R2/R4..R8 with the date rule FLIPPED (E3: date <= 2024-08-31 = EXPOSED) => 50 seen, 48 admitted, 42 tier-1 pure pass-fail >= 30 (+5 float +1 custom; 2 custom-no-validator excluded); SHA 653e3682; manifest CID 8acbc7cc; 30-slice CID 32d15db5; grader self-test 30 all-pass / 637 cases each surface; RMRC 2023-2024 excluded (typed: minimal package, no shipped problem_statement). LANE beta: W114 gate with C2->C2e (>=30 problems at/before cutoff) certifies Maverick on the EXPOSED side; tier-2 all UNKNOWN; EARNED pilot RAN (Maverick x exposed 30-slice, 1 seed x 30 x K=5 = 330 NIM calls, ~70 min): A0 6.67 / A1 26.67 / B 30.00 %; B-A1 = +3.33 pp; MLB-1 93.33% PASS / MLB-2 25.00% FAIL; 8/9; FAIL (net +1 problem: rescues rsamistake+isbnconversion, regression icouldhavewon; bench merkle 618e270a). LANE gamma: shipped coordpy.coordpy_icpc_exposed_control_v1 (reuses W120 classifier/oracle/slice-selector + W114 cutoff registry + W117 run_upstream_construction_v1 => decision CID 258b6ed7 invariant) + MatchedFamilyComparisonV1 (differs_only_in_cutoff_side=True; same org/format/grader/tiers/difficulty/model) + the pre-committed three-branch interpret_exposed_vs_resistant_v1 + W122FireConditionV1 + 3 scripts + 17 tests. THE ANSWER: EXPOSED +3.33pp vs RESISTANT +0.00pp (W120) both within the +-3.34pp null band => the contamination hypothesis predicted a clean exposed margin (absent), difficulty/family-ease predicted a null even when exposed (matched) => contamination-confound WEAKENED (first within-family within-model exposure control), difficulty/family-ease implicated, bounded ceiling HARDENS; NOT refuted (faint sub-floor gradient: exposed +3.33 > resistant +0.00, exposed rescue 25% > resistant 8%; single seed); difficulty comparability supported (exposed A0 6.67% <= resistant A0 20.00%). NO third retirement (W89+W105 STAND); paired seed NOT earned (null-side of band; W106 discipline); COO-9 lead; COO-45; W122 = accept hardened bounded ceiling / different axis OR primary-KNOWN stronger-than-Maverick on BOTH battlefields OR (optional) one paired seed on BOTH. 31st preflight/earn-discipline validation (W93-W121). No version bump; no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "d49237a"
      ],
      "docs": [
        "docs/RUNBOOK_W121.md",
        "docs/RESULTS_W121_EXPOSED_CONTROL_V1.md",
        "docs/RESULTS_W121_MILESTONE_SUMMARY_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W121_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W121_V1.md",
        "results/w121/exposed_control/exposed_control_verdict.json",
        "results/w121/exposed_control/exposed_listing_live.json",
        "coordpy/coordpy_icpc_exposed_control_v1.py",
        "scripts/build_w121_exposed_listing_v1.py",
        "scripts/run_w121_exposed_control_v1.py",
        "scripts/run_w121_exposed_pilot.py",
        "tests/test_w121_exposed_control_v1.py"
      ],
      "linear_issues": [
        "COO-45"
      ],
      "carry_forwards_added": [
        "W121-T-MATCHED-EXPOSED-ICPC-CONTROL-CONSTRUCTIBLE",
        "W121-T-MAVERICK-EXPOSED-CERTIFIABLE",
        "W121-L-MATCHED-FAMILY-EXPOSURE-DOES-NOT-REPRODUCE-MARGIN-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W122",
      "title": "matched-family 3-seed paired-seed CLOSURE on official ICPC (seeds 120002 + earned tiebreaker 120003 on BOTH fields; SAME W120 resistant 30-slice CID 01bf9ef8 + W121 exposed 30-slice CID 32d15db5; Maverick; only the seed changed; 1320 NIM calls) => FINAL 3-seed aggregate AMBIGUOUS_THIRD_PAIRED_SEED_EARNED (B4): seed 120003 spiked +10.00pp on BOTH fields (PASS_NON_MECHANISM_DRIVEN) => 3-seed means RESISTANT +4.44pp (OUT of +-3.34 band, in the 3.34-5.00 gap) / EXPOSED +8.89pp (OUT of band); both all_seeds_clean_pass=false, neither shows_margin => B1/B2/B3 all off; closure NOT achieved (the 2-seed 'resistant null vs exposed popped' asymmetry DISSOLVED -- resistant ALSO spiked -- so both fields show rescue-concentrated non-mechanism variance and the matched contrast is unresolvable at n=30); the single-seed caveat is retired (3 seeds/side) but REPLACED by a small-n-variance limitation; M3 KILLED NIM-free (m3_exclusive_signal_fraction=0.000; ICPC secret token-diff => ceiling mechanism-robust); stronger-model gate STRUCTURALLY CLOSED; NO third retirement (W89+W105 STAND); no 4th seed",
      "outcome": "Gated branch milestone with THREE lanes executing docs/RUNBOOK_W122.md (LOCKED before any NIM; the W121 section-9 CONFOUND_WEAKENS branch's optional paired-seed tightening, PROMOTED to the main empirical lane). ultracode OFF; COO-9 stays lead; COO-46 is the W122 issue. LANE alpha (matched-family multi-seed closure): reused the EXACT W120 resistant 30-slice + W121 exposed 30-slice (re-derived NIM-free == provenance == pilot guard); Maverick x BOTH x seeds 120002+120003 (1320 calls). Per-seed B-A1: RESISTANT 120001 +0.00 FAIL / 120002 +3.33 FAIL / 120003 +10.00 PASS_NON_MECHANISM_DRIVEN (9/9, MLB-2 16.7%, 3 rescues/0 regr, merkle adf55ff9); EXPOSED 120001 +3.33 FAIL / 120002 +13.33 PASS_NON_MECHANISM_DRIVEN / 120003 +10.00 PASS_NON_MECHANISM_DRIVEN (9/9, MLB-2 18.5%, 3 rescues/0 regr, merkle d88d025f). FINAL 3-seed means RESISTANT +4.44 / EXPOSED +8.89, both out of band, neither all-clean-per-seed => interpret_paired_closure_v1 emits AMBIGUOUS_THIRD_PAIRED_SEED_EARNED (B4), caveat_closed=false; B4-after-the-3rd-seed is terminal (no 4th). The driver was fixed first so the 3-seed aggregate gathers ALL prior seeds keyed by seed (NIM-free verified). CLOSURE NOT achieved: the 3rd seed dissolved the 2-seed asymmetry (resistant also spiked +10.00); at n=30/K=5 the per-field B-A1 swings +-10pp on ~3 rescues => the matched contrast is unresolvable at n=30; the W121 'weakened' read is NOT multi-seed-confirmed and contamination is NOT established (the 3-seed data UNDERCUTS a clean contamination read: contamination-RESISTANT code spikes just like exposed => sampling variance, not exposure; faint exposed>resistant +8.89>+4.44 is exposure-consistent but non-mechanism-driven/noisy). LANE beta (same-family different-mechanism M3): NIM-free audit (audit_icpc_mechanism_signal_v1) over the real W120+W121 sidecars (240 reflexion turns) => m3_exclusive_signal_fraction=0.000 < 0.33 floor => KILL_M3_LANE_NIM_FREE (ICPC hidden oracle is SECRET token-diff, denying M3 its expected/actual differentiator => same-family ceiling mechanism-robust, not merely reflexion-specific); $0. LANE gamma (stronger-model gate): resistant field anchored to Maverick Aug-2024 => only a primary-KNOWN cutoff <= ~Aug-2024 could be resistant-certified, Maverick the unique reachable such model => gate STRUCTURALLY CLOSED; {KNOWN:1, UNKNOWN:4}; decision CID 258b6ed7 invariant; $0. NO third retirement (W89+W105 STAND, the only two); W122 adds none. 1 new explicit-import-only module + 2 scripts + 17 tests (falsifiability-first); fixed a W121 doc CID typo (f7cdc917->01bf9ef8); graphify refreshed START (6b05ccd2) + END; no version bump; no PyPI; coordpy/__init__.py untouched; 32nd consecutive preflight/earn-discipline validation (W93-W122). W123 = accept the bounded HumanEval-family ceiling / escalate to larger n PER FIELD (>=100/field, not more n=30 seeds, no 4th) OR a primary-KNOWN stronger-than-Maverick model on BOTH matched ICPC battlefields. (2026-05-31)",
      "commits": [
        "45e08ea"
      ],
      "docs": [
        "docs/RUNBOOK_W122.md",
        "docs/RESULTS_W122_PAIRED_SEED_CLOSURE_V1.md",
        "docs/RESULTS_W122_MILESTONE_SUMMARY_V1.md",
        "docs/RESULTS_W122_MECHANISM_AUDIT_V1.md",
        "docs/CONTAMINATION_CONTROL_FRAMING_W122_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W122_V1.md",
        "results/w122/paired_seed/w122_paired_seed_120003_20260531T184401Z/paired_seed_closure_verdict.json",
        "results/w122/mechanism_audit/mechanism_audit_verdict.json",
        "coordpy/coordpy_icpc_paired_seed_closure_v1.py",
        "scripts/run_w122_paired_seed_pilot.py",
        "scripts/run_w122_mechanism_signal_audit_v1.py",
        "tests/test_w122_paired_seed_closure_v1.py"
      ],
      "linear_issues": [
        "COO-46"
      ],
      "carry_forwards_added": [
        "W122-T-PAIRED-SEED-CLOSURE-RULE-PRECOMMITTED",
        "W122-T-PAIRED-SEED-3SEED-AMBIGUOUS-B4",
        "W122-T-BOTH-FIELDS-NONMECHANISM-SPIKE",
        "W122-T-THIRD-PAIRED-SEED-RAN-RESOLVED-B4",
        "W122-T-ICPC-CEILING-MECHANISM-ROBUST",
        "W122-T-STRONGER-MODEL-GATE-STRUCTURALLY-CLOSED",
        "W122-L-NO-THIRD-RETIREMENT-CAP",
        "W122-L-MATCHED-CONTRAST-UNRESOLVABLE-AT-N30-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W123",
      "title": "large-n matched ICPC closure ATTEMPT (>=100/field): official-family supply census",
      "outcome": "alpha-CAP — the >=100/field matched battlefield is UNREACHABLE from the official github.com/icpc family, blocked SOLELY on the post-cutoff axis. Lane-alpha (NIM-free, live gh-API-verified all_match=true, census_cid bf7b2efbc724): RESISTANT 4 post-cutoff surfaces 51 raw/~45 tier-1 (51==W120 n_seen; all mined by W120; no 5th; <100 even at 100% yield = HARD CAP) vs EXPOSED 11 pre-cutoff surfaces 135 raw/~113 tier-1 (REACHES >=100). Lane-beta NOT earned (needs BOTH>=100) => $0 NIM, no pilot, no n=30 seeds. Lane-gamma stronger-model gate CLOSED+MOOT {KNOWN:1,UNKNOWN:4}; decision CID 258b6ed7 invariant. n=30 caveat now POST-CUTOFF-SUPPLY-BOUND (W123-L). W89+W105 STAND (only two retirements). No version bump; no PyPI; coordpy/__init__.py untouched. Self-caught + corrected a phantom-output census-count error mid-milestone.",
      "commits": [
        "c023599",
        "388032b"
      ],
      "docs": [
        "docs/RUNBOOK_W123.md",
        "docs/RESULTS_W123_LARGEN_SUPPLY_CENSUS_V1.md"
      ],
      "linear_issues": [
        "COO-48"
      ],
      "carry_forwards_added": [
        "W123-L-LARGEN-MATCHED-BATTLEFIELD-SUPPLY-UNREACHABLE",
        "W123-T-STRONGER-MODEL-GATE-CLOSED-AND-MOOT"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W124",
      "title": "transformer-native code-intervention line on matched ICPC (M4 AST-boundary hidden-state probe on distilgpt2) + learned-memory/controller line + hosted gate",
      "outcome": "M4_CLOSE_BLIP_NOT_A_GAIN — distilgpt2 AST-boundary hidden state adds nothing over surface (AUC_hidden 0.6345 ≈ AUC_surface 0.6343, Δ=+0.0001 ≪ +0.05); transformer-native hidden-state line BLOCKED AT THE PRECURSOR on local code-model-encoder supply (no code-competent local model; tf 4.28.1 too old + repo transformers_runtime_v1 needs ≥~4.36); learned-memory controller at chance 0.502 (TOO_SYNTHETIC_NOT_WARRANTED); stronger-model gate CLOSED, decision CID 258b6ed7 invariant {KNOWN:1,UNKNOWN:4}; M6 tool-substrate controller contract shipped (text-translatable, contract-only); NO hosted Maverick probe earned ($0 NIM); W89+W105 STAND (no third retirement); ultracode OFF; no version bump; no PyPI; coordpy/__init__.py untouched",
      "commits": [
        "e3348fd"
      ],
      "docs": [
        "docs/RUNBOOK_W124.md",
        "docs/RESULTS_W124_TRANSFORMER_NATIVE_CODE_INTERVENTION_V1.md"
      ],
      "linear_issues": [
        "COO-49"
      ],
      "carry_forwards_added": [
        "W124-L-TRANSFORMER-NATIVE-HIDDEN-STATE-ADDS-NOTHING-OVER-SURFACE",
        "W124-L-TRANSFORMER-NATIVE-LINE-BLOCKED-AT-PRECURSOR-LOCAL-ENCODER-SUPPLY",
        "W124-L-LEARNED-MEMORY-CONTROLLER-AT-CHANCE-ON-ICPC-RESCUE",
        "W124-T-M6-TOOL-SUBSTRATE-CONTROLLER-CONTRACT-SHIPS",
        "W124-T-STRONGER-MODEL-GATE-CLOSED-AND-HOSTED-MOOT",
        "W124-L-NO-THIRD-RETIREMENT-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W125",
      "title": "hosted controller-native code mechanism on the official RESISTANT ICPC family (router/logprob-router/cache-aware-planner v12 + tool_call_substrate_v1 + executor_grounded_patcher_v1) + resistant-first $0 headroom replay + stronger-model gate",
      "outcome": "controller arsenal is REAL + contract-clean (mechanism question YES) but the resistant field is GENERATION-CAPPED for $0 re-routing (spend question NOT earned). Lane alpha: controller_native_code_mechanism_v1 (first bridge of the hosted-controller stack to the audited tool-call substrate; graphify-confirmed 6 hops apart, no edge); structural fake-different test BITES (reflexion B + C0 control => FAKE_DIFFERENT; C1/C2/C3 => REAL; lead C3 cid 7989655f); all 4 NIM-free contract checks PASS. Lane beta ($0 replay over 330 already-paid Maverick gens on the W120 resistant 30-slice CID 01bf9ef8): A1 7/30, pool-union 8/30, oracle_headroom +1 (in null band), C2 7/30 + C3 7/30 => blind_selection_headroom=0; reflexion_divergence 23/30 but looks_right_fails_hidden 10 (public-sample signal non-discriminating) => FRESH_RESISTANT_PILOT_NOT_EARNED_HEADROOM_CAP; $0 NIM; exposed control NOT bought (resistant-first). Lane gamma: NO_CERTIFIABLE_STRONGER_MODEL, decision CID 258b6ed7 invariant, {KNOWN:1,UNKNOWN:4}; local transformer-native line CLOSED. W89+W105 STAND (no third retirement). ultracode OFF; no version bump; no PyPI; coordpy/__init__.py untouched",
      "commits": [
        "97232eb"
      ],
      "docs": [
        "docs/RUNBOOK_W125.md",
        "docs/RESULTS_W125_CONTROLLER_NATIVE_CODE_MECHANISM_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W125_V1.md"
      ],
      "linear_issues": [
        "COO-50"
      ],
      "carry_forwards_added": [
        "W125-T-CONTROLLER-NATIVE-MECHANISM-REAL-NOT-FAKE-DIFFERENT",
        "W125-L-RESISTANT-GENERATION-CAP",
        "W125-T-RESISTANT-PUBLIC-SAMPLE-SIGNAL-NON-DISCRIMINATING",
        "W125-L-NO-THIRD-RETIREMENT-CAP"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W126",
      "title": "family-adapted repair SYNTHESIS on the official RESISTANT ICPC family (adversarial_consensus_repair_v1.trust_weighted_consensus_v1 + executor_grounded_patcher_v1 digest bridged onto the official-ICPC code path) + resistant-first $0 synthesis precursor + stronger-model gate",
      "outcome": "synthesis slate REAL + arsenal-native + contract/leakage-clean (mechanism question YES) but the resistant field is SYNTHESIS-CAPPED for $0 NEW-trajectory creation (oracle ceiling 0/22; spend question NOT earned); FRESH_RESISTANT_PILOT_NOT_EARNED_SYNTHESIS_DEAD; $0 NIM; W89+W105 STAND (no third retirement)",
      "commits": [
        "01b456a"
      ],
      "docs": [
        "docs/RUNBOOK_W126.md",
        "docs/RESULTS_W126_FAMILY_ADAPTED_REPAIR_SYNTHESIS_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W126_V1.md"
      ],
      "linear_issues": [
        "COO-51"
      ],
      "carry_forwards_added": [
        "W126-L-RESISTANT-SYNTHESIS-CAP",
        "W126-T-SYNTHESIS-OF-CAPABILITY-FAILURES-IS-DEAD-NOT-BLIND-CAPPED"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W127",
      "title": "resistant capability atlas + family-specific algorithm-scaffold GENERATION + targeted resistant fresh-generation probe (operator-greenlit dev spend; coordpy.resistant_capability_atlas_v1 + coordpy.family_scaffold_generation_v1 wired onto the official-ICPC path under a no-leakage rule)",
      "outcome": "atlas (22 resistant failures = 95% wrong-algorithm + algorithmically DIVERSE, no dominant cluster, surface labels 47% concordant) + scaffold line REAL; EXPOSED dev bench EARNED (scaffold 3/8 vs baseline 1/8, net +2 / 2 families / 0 regressions / leakage-clean after a documented boilerplate-FP recalibration to a contiguous-block tripwire, positive control preserved) but the earned R1 and R2 targeted resistant probe = 0/6 new solves => RESISTANT_SCAFFOLD_FRESH_GEN_CAP (the exposed +2 did NOT transfer to the non-memorizable resistant field; most parsimoniously memorization/variance/framing, not scaffold-taught algorithm); stronger-model gate CLOSED 258b6ed7 {KNOWN:1,UNKNOWN:4}; exposed control NOT bought; $0 beyond authorized dev(80)+probe(30)=110 NIM; W89+W105 STAND (no third retirement); COO-9 lead; no version bump; no PyPI; coordpy/__init__.py untouched",
      "commits": [
        "95f55a8"
      ],
      "docs": [
        "docs/RUNBOOK_W127.md",
        "docs/RESULTS_W127_CAPABILITY_ATLAS_AND_SCAFFOLD_GENERATION_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W127_V1.md"
      ],
      "linear_issues": [
        "COO-52"
      ],
      "carry_forwards_added": [
        "W127-L-RESISTANT-SCAFFOLD-FRESH-GEN-CAP",
        "W127-T-RESISTANT-FAILURES-ARE-DIVERSE-WRONG-ALGORITHM-NOT-SURFACE",
        "W127-T-EXPOSED-SCAFFOLD-DEV-BENCH-EARNS-WEAKLY-CONFOUNDED"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W128",
      "title": "role-diverse algorithm SEARCH on the non-scaffoldable resistant ICPC clusters + same-family hard-cluster dev bench + targeted resistant probe (operator-greenlit EXPOSED dev spend; coordpy.role_diverse_algorithm_search_v1 bridging the W41/W42 synthesis decisions + the executor digest onto the official-ICPC path under a no-leakage rule)",
      "outcome": "mechanism REAL (all 11 EXPOSED hard-cluster dev runs classify genuine REAL-diversity after a documented sketch-parser-bug recalibration + clean re-run) and the diverse SEARCH LIFTS the generation ceiling (pool 3/11 > plain baseline 2/11, reaching a simulation_grid program i.i.d. sampling missed) BUT NOT EARNED: RDA4 committed 2/11, net +0 (=+1 unique blueberrywaffle - 1 regression pawnshop), ties baseline + the W127 scaffold line => ROLE_DIVERSE_HARD_CLUSTER_DEV_BENCH_NOT_EARNED; the cap is localized to the verification-based SELECTION layer (abstains 7/11, mis-commits the one pool-only win), NOT generation (W128-T-ROLE-DIVERSE-SEARCH-LIFTS-GENERATION-CEILING-BUT-SELECTION-CAPPED). T1 FALSE => targeted resistant probe NOT launched (T2 technically matches via the simulation_grid unique solve but T1 fails the net bar) => $0 resistant NIM; exposed control NOT bought; stronger-model gate CLOSED 258b6ed7 {KNOWN:1,UNKNOWN:4}; W79 substrate-controller literal bridge examined + KILLED as fake (W128-T-SUBSTRATE-CONTROLLERS-NOT-CODE-CONSENSUS-APPLICABLE); graph_flow EXPOSED supply 0 (W128-L-GRAPH-FLOW-EXPOSED-SUPPLY-CAP); 360 EXPOSED dev calls (30 validation + 165 parser-bug-invalidated + 165 valid), $0 resistant; W89+W105 STAND (no third retirement); COO-9 lead; no version bump; no PyPI; coordpy/__init__.py untouched",
      "commits": [
        "03501ef"
      ],
      "docs": [
        "docs/RUNBOOK_W128.md",
        "docs/RESULTS_W128_ROLE_DIVERSE_ALGORITHM_SEARCH_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W128_V1.md"
      ],
      "linear_issues": [
        "COO-53"
      ],
      "carry_forwards_added": [
        "W128-L-ROLE-DIVERSE-HARD-CLUSTER-DEV-BENCH-CAP",
        "W128-T-ROLE-DIVERSE-SEARCH-LIFTS-GENERATION-CEILING-BUT-SELECTION-CAPPED",
        "W128-L-GRAPH-FLOW-EXPOSED-SUPPLY-CAP",
        "W128-T-SUBSTRATE-CONTROLLERS-NOT-CODE-CONSENSUS-APPLICABLE"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W129",
      "title": "public-signal SELECTION ORACLE on the W128 hard-cluster miss pattern + same-family EXPOSED dev-bench validation + targeted resistant probe only if earned (operator-greenlit EXPOSED dev spend; coordpy.public_signal_selection_oracle_v1 = SO1 falsifier stack / SO2 differential-disagreement via integrated_synthesis / SO3 verifier-final mined from mathvista_bench_v2 / SO4 trust-weighted-abstain realizing the integrity_trust_coupled_consensus_v1 ABSTAIN concept natively, substrate TrustWeightedConsensusController literal bridge KILLED)",
      "outcome": "the W128 pawnshop 'selection miss' is a COMPLEXITY bug a code-reading verifier BREAKS (a $0 recon shows hidden-correct B1 and hidden-wrong A0 are byte-identical on every public+derived case, but A0 is O(N^2) and B1 O(N) under N<=300000 so A0 fails secret by TLE; the SO3 verifier-final chose B1 citing complexity, refuting 'non-discriminating even for a model judge'), NOT a hard public-signal information limit. The NIM-free SOLEAD is abstain-disciplined (keeps the separable blueberrywaffle win, abstains the under-determined pawnshop+sunandmoon ties => 0 SELECTION mis-commits vs W128-RDA4's 1; only the generation-level single-survivor doubleup remains); a generic stress scale-up does NOT crack pawnshop (adversarial worst-case ~ solving the problem; not hand-tuned). BINDING: committed <= the GENERATION pool ceiling (3/11 = baseline+1) < the +2 earn bar regardless of selector quality => R2' structurally unreachable by selection => beta1 fresh dev bench NOT fired (hope-funded) => the binding cap MOVED from W128's SELECTION to W129's GENERATION ceiling; T1 FALSE => $0 resistant NIM. Stronger-model gate NO_CERTIFIABLE_STRONGER_MODEL, decision CID 258b6ed7 invariant, {KNOWN:1,UNKNOWN:4}; exposed control NOT bought; W123-W128 caps stay closed. W89(+5.56)+W105(+7.00) STAND as the only two retirements (none added, none retired). 3 NIM total (beta0 SO3 verifier probe); 1 module + 5 scripts + 29 tests; graphify refreshed START 0f47803 + END; COO-9 lead; no version bump; no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "f621120"
      ],
      "docs": [
        "docs/RUNBOOK_W129.md",
        "docs/RESULTS_W129_PUBLIC_SIGNAL_SELECTION_ORACLE_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W129_V1.md"
      ],
      "linear_issues": [
        "COO-54"
      ],
      "carry_forwards_added": [
        "W129-T-PAWNSHOP-SELECTION-MISS-IS-COMPLEXITY-VERIFIER-BREAKABLE",
        "W129-T-ABSTAIN-DISCIPLINE-ELIMINATES-SELECTION-MISCOMMIT",
        "W129-L-NIMFREE-GENERIC-STRESS-CANNOT-CONSTRUCT-ADVERSARIAL-WORSTCASE",
        "W129-L-HARD-CLUSTER-GENERATION-CEILING-CAPS-SELECTION-EARN"
      ],
      "carry_forwards_retired": []
    },
    {
      "id": "W130",
      "title": "generation-ceiling attack on the hard ICPC clusters + same-family EXPOSED dev bench + targeted resistant probe only if earned (operator-greenlit EXPOSED dev spend; 275 NIM dev; coordpy.generator_failure_atlas_v1 generator-failure atlas + coordpy.stronger_generator_slate_v1 GG1 complexity-gated handoff / GG2 counterexample-to-rewrite / GG3 family anti-pattern coach / GG4 budget router / GGLEAD, W129 selector held FIXED downstream)",
      "outcome": "the generator-failure atlas decomposed the 8/11 pool-DEAD set (3 WRONG_ALGORITHM_ADMISSIBLE + 2 HIDDEN_EDGE_STATE_MISS + 2 WRONG_ALGORITHM_NO_SKETCH capability + 1 PARSE_IO_FAILURE; 2 pool-bearing misses SELECTION_FIXABLE); the stronger same-budget generator slate created exactly 1 NEW pool solve (GG2 cracked doubleup, a HIDDEN_EDGE problem the entire old W128/W129 pool missed = FIRST W120-W130 generation crack of a pool-DEAD problem) but NOT EARNED (1 < the +2 earn bar, does not span >=2 families/modes); the dominant WRONG_ALGORITHM_ADMISSIBLE mode cracked by no arm (admissible idiom-overlap = UPPER BOUND, capability-bound); stored regression trio PRESERVED (0 new mis-commits); GG3 killed (boilerplate leakage); T1 FALSE => $0 resistant NIM; stronger-model gate CLOSED 258b6ed7 {KNOWN:1,UNKNOWN:4}; W89+W105 STAND (no third retirement)",
      "commits": [
        "cccc56a"
      ],
      "docs": [
        "docs/RUNBOOK_W130.md",
        "docs/RESULTS_W130_GENERATION_CEILING_ATTACK_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W130_V1.md"
      ],
      "linear_issues": [
        "COO-55"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W130-L-GENERATION-CEILING-DEV-BENCH-CAP",
        "W130-T-COUNTEREXAMPLE-REWRITE-LIFTS-ONE-HIDDEN-EDGE",
        "W130-T-ADMISSIBLE-SKETCH-IS-CAPABILITY-NOT-GENERATION-FIXABLE",
        "W130-T-ANTI-PATTERN-COACH-RISKS-BOILERPLATE-LEAKAGE"
      ]
    },
    {
      "id": "W131",
      "title": "code-competent MODEL supply census + stronger-generator hard-cluster dev bench (W129 selector FIXED) + targeted resistant probe only if earned (operator-greenlit EXPOSED dev spend; ~250 NIM dev; coordpy.code_model_supply_census_v1 three-surface supply census + coordpy.generator_model_bench_v1 B0 Maverick-reuse / B1_PLAIN / B2_RDIV(GG1) / B3_GG2 / B4_GGLEAD, W129 NIM-free SOLEAD selector held FIXED downstream)",
      "outcome": "Lane alpha census: local-HF DEAD (no torch/transformers under Python 3.14.5) but 13 reachable STRONGER-than-Maverick code models on the NIM catalogue (Qwen3-Coder-480B / DeepSeek-V4-pro / Qwen3.5-397B / Mistral-Large-3-675B / GLM-5.1), ALL UNKNOWN-from-primary on cutoff => 32 DEV_ONLY, FRONTIER_ELIGIBLE=NONE => the model-axis supply gap MOVED from W124's 'no strong code model exists/loads' to 'no PRIMARY-KNOWN-cutoff stronger model on the ICPC family' (disclosure, not existence; census_cid d360c117); a load-bearing parsing-fairness fix normalize_fence_v1 (the Qwen-Coder lone-language-tag fence quirk crashes the audited extractor; NOT a capability lever). Lane beta (220 NIM valid): the strongest reachable code model Qwen3-Coder-480B-a35b (~28x Maverick's active params), full B1-B4 slate at MATCHED K=5 with the W129 selector held FIXED so GENERATION is the only variable, created 0 genuinely-new EXPOSED hard-cluster solves (B1_PLAIN reproduced exactly the W130 doubleup; B2/B3/B4 added 0) => NOT EARNED (1 < the +2-spanning bar); the dominant WRONG_ALGORITHM_* failures unmoved by a frontier code model = capability-bound; stored regression trio PRESERVED (0 new mis-commits; the 480B even committed a CORRECT pawnshop via SINGLE_SURVIVOR; role-diverse reached correct blueberrywaffle/sunandmoon but the fixed selector correctly ABSTAINED); local 32B reachable+code-competent (smoke PASS) but CPU-bound on this host (~1 tok/s) = throughput-impractical. Lane gamma: T1 FALSE => NO targeted resistant probe => $0 resistant NIM; T2 moot (FRONTIER_ELIGIBLE=NONE); stronger-model gate NO_CERTIFIABLE_STRONGER_MODEL, decision CID 258b6ed7 invariant, {KNOWN:1, UNKNOWN:4}; W123-W130 caps stay closed. W89 (+5.56) + W105 (+7.00) STAND as the only two retirements (none added, none retired). MODEL-axis sibling of the W123->W130 cap taxonomy. No version bump; no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "3708ea5"
      ],
      "docs": [
        "docs/RUNBOOK_W131.md",
        "docs/RESULTS_W131_MODEL_SUPPLY_AND_GENERATOR_BENCH_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W131_V1.md"
      ],
      "linear_issues": [
        "COO-56"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W131-L-MODEL-AXIS-GENERATION-CEILING-DEV-BENCH-CAP",
        "W131-T-STRONGER-MODEL-DOES-NOT-CRACK-NEW-HARD-CLUSTER-PROBLEMS",
        "W131-T-FRONTIER-ELIGIBLE-SUPPLY-IS-CUTOFF-DISCLOSURE-BOUND"
      ]
    },
    {
      "id": "W132",
      "title": "resistant-by-construction hard-family battlefield + exact-oracle quality gates + EXECUTED frontier pilot on Llama-3.3-70B (Maverick infra-down) = clean FAIL B-A1=+3.33pp (coordpy.resistant_by_construction_battlefield_v1 framework + _slate_v1 33-template slate)",
      "outcome": "W123 bounded official resistant-package SUPPLY and W131 found the model axis blocked on cutoff DISCLOSURE; both share one upstream dependency (a resistant battlefield we inherit or certify against a disclosed cutoff). W132 REMOVED it by MINTING the battlefield AND RAN the frontier pilot. Lane alpha ($0): 33 admitted (>=30) resistant-by-construction problems (COMPLEXITY_BLIND 9 / HIDDEN_EDGE_STATE_MISS 8 / WRONG_ALGORITHM_ADMISSIBLE 8 / SEARCH_ENUM 8, 33 distinct families) each with an exact ref_source answer-key oracle cross-checked by an INDEPENDENT exhaustive brute_source, an admissible-wrong naive_source trap (looks-right-on-public / fails-hidden discriminating gate), a near-dup+official-identity novelty guard, deterministic regeneration, manifest_cid 562aafbd / core 30-slice f6a2ebed; emitted as IcpcPilotProblemV1 for the already-validated W120 bench. Resistance-certified by date AND construction => W132-T-RESISTANT-BY-CONSTRUCTION-BATTLEFIELD-MINTABLE. Lane beta: battlefield EARNED the pilot; the locked target Maverick was INFRA-DOWN (machine-checked: same key, llama-3.1-8b 0.42s / llama-3.3-70b 7s respond, maverick chat hangs 0-bytes 45-75s) => frontier target substituted PRE-SPEND (RUNBOOK 8d) to meta/llama-3.3-70b-instruct, the EXACT W105 retirement model (KNOWN cutoff ~Dec-2023, resistant by date AND construction). Executed pilot (30-slice, 1 seed, K=5, 330 calls, 80min): A0=A1=76.67% B=80.00% B-A1=+3.33pp (< +5pp; 7/9 gates; MLB-1 26.7% FAIL, MLB-2 25% FAIL) => VERDICT FAIL => W132-L-RESISTANT-BY-CONSTRUCTION-PILOT-CAP; +3.33pp = 1 B-unique COMPLEXITY rescue / 0 regressions, 6 capability-bound traps resist all arms; CLOSES the 'wrong-test' escape (+3.33pp = SAME as the W121 matched-exposed-ICPC control). Stronger-model gate CLOSED (258b6ed7 invariant); Maverick = push-button CROSS-SCALE check on the same slice. No third retirement — W89 (+5.56) + W105 (+7.00) STILL the only two.",
      "commits": [
        "1dd9582",
        "517b015",
        "0dfa7f6"
      ],
      "docs": [
        "docs/RUNBOOK_W132.md",
        "docs/RESULTS_W132_RESISTANT_BY_CONSTRUCTION_BATTLEFIELD_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W132_V1.md"
      ],
      "linear_issues": [
        "COO-57",
        "COO-9"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W132-T-RESISTANT-BY-CONSTRUCTION-BATTLEFIELD-MINTABLE",
        "W132-L-RESISTANT-BY-CONSTRUCTION-PILOT-CAP"
      ]
    },
    {
      "id": "W133",
      "title": "exact-oracle WITNESS curriculum + held-out witness-guided mechanism bench on Llama-3.3-70B (Maverick infra-down) — a REAL but SINGLE-MODE complexity-witness gain; frontier rerun NOT earned (coordpy.exact_oracle_witness_v1 + witness_curriculum_corpus_v1)",
      "outcome": "W132 closed the 'wrong-test' escape but left the BLIND reject bit un-pulled; W133 turned the owned battlefield into a TEACHER. Lane alpha ($0): exact_oracle_witness_v1 (EW1 minimal-counterexample / EW2 complexity-timing / EW3 edge / EW4 search-enum slate + same-budget witness arm + fake-different test) + witness_curriculum_corpus_v1 (seed-disjoint train/dev/eval over RBC_SLATE_V1) => 99 admitted = 33/33/33 (>=96; curriculum_cid e5e44a43; eval split CID 88b9b79a LOCKED before any eval spend); witnesses fire leakage-clean + genuinely-new on every admitted problem AND all six W132 traps; no-leakage enforced BY WITNESS TYPE (EW1 disjoint-from-graded, EW2 size/timing-only), graded on a DISJOINT hidden bank (generalisation not memorisation); 9 tests incl. a positive control => W133-T-EXACT-ORACLE-WITNESS-INSTRUMENT-MINTABLE. Lane beta (held-out dev spend; meta/llama-3.3-70b-instruct, Maverick infra-down): same-budget A0/A1/B0 + C1/C2/C3, 858 NIM, A0=72.73 A1=75.76 B0=81.82 (B0-A1=+6.06); C1=81.82 (arm-B0=+0.00, 0 rescues vs B0); C2=C3=87.88 (arm-A1=+12.12 arm-B0=+6.06, MLB-2 60%, 4 ALGORITHMIC complexity rescues). The EW2 complexity witness is REAL + load-bearing (cracks O(N^2) COMPLEXITY traps B0 + counterexample-only C1 cannot) => W133-T-COMPLEXITY-WITNESS-IS-REAL-AND-LOAD-BEARING-ON-HELD-OUT-DEV; but the EW1 counterexample witness adds +0.00 over blind reflexion (wrong-algorithm capability ceiling). SINGLE-MODE gain => meets -A1>=+5 + -B0>+3.33 but FAILS the >=2-mode condition => §7a dev gate FAILS => $0 eval, $0 frontier (NOT earned) => W133-L-WITNESS-FEEDBACK-SINGLE-MODE-CAP; LOCALISES the W132 cap (part feedback-fixable complexity + part capability-bound value/algorithm). Lane gamma ($0): primary-source research confirms the mechanism EXECUTABLE-HERE + PREDICTED the two limits; stronger-model gate NO_CERTIFIABLE_STRONGER_MODEL, decision CID 258b6ed7 invariant {KNOWN:1,UNKNOWN:4}; graphify START 66c0b38 + END (exact_oracle_witness_v1 = first bridge to battlefield + bench + grader); no 405B. W89 (+5.56) + W105 (+7.00) STILL the only two retirements; W133 retires none. COO-9 stays lead. No version bump (0.5.20 / coordpy.sdk.v3.43); no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "39ab1a0"
      ],
      "docs": [
        "docs/RUNBOOK_W133.md",
        "docs/RESULTS_W133_WITNESS_FEEDBACK_BENCH_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W133_V1.md"
      ],
      "linear_issues": [
        "COO-58",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W133-T-EXACT-ORACLE-WITNESS-INSTRUMENT-MINTABLE",
        "W133-T-COMPLEXITY-WITNESS-IS-REAL-AND-LOAD-BEARING-ON-HELD-OUT-DEV",
        "W133-L-WITNESS-FEEDBACK-SINGLE-MODE-CAP"
      ]
    },
    {
      "id": "W134",
      "title": "deployable complexity witness (constraint-derived, oracle-free) + dedicated complexity-only held-out corpus + held-out dev bench — REAL but SUB-ORACLE + SINGLE-FAMILY; frontier rerun NOT earned (coordpy.deployable_complexity_witness_v1 + complexity_only_corpus_v1)",
      "outcome": "Distilled the W133 exact-oracle EW2 complexity witness into a DEPLOYABLE oracle-free public-signal witness (DW1 constraint-derived budget via parse_max_constraint_v1 + DW2 public-format stress-growth ladder/log-log exponent/extrapolation + DW3 rewrite + DW4 KEEP/REWRITE/ABSTAIN controller; same-budget arm consumes ONLY code+statement+public-samples) and validated on a dedicated COMPLEXITY-ONLY held-out corpus (9 cb_ families x 5 seed-disjoint seeds/split; admitted 45/45/45/45 >= 36/36/36/30; corpus_cid 191d9954 timeout-invariant; eval_split 748dd6fa + frontier_slice 31a81304 LOCKED; naive/ref separation 45/45 == exact-oracle EW2). Held-out DEV bench (18 problems, Llama-3.3-70B, 522 NIM): A0 44.44 / A1 55.56 / B0 88.89 / C0(exact-oracle EW2) 100.00 / D1 94.44 / D2 94.44 / D3(lead) 88.89. Blind reflexion B0 is already strong on a complexity-only field (B0-A1=+33.33pp); even C0's full +11.11pp gain over B0 is SINGLE-FAMILY; deployable D1/D2 +5.56pp over B0 (half of C0, 1 family, 1 witness-induced regression); controller D3 ties B0. §7a dev gate FAILS for every arm => $0 eval, $0 frontier (frontier rerun NOT earned). Stronger-model gate CLOSED (258b6ed7 invariant). W89 (+5.56) + W105 (+7.00) STILL the only two retirements; W134 retires none. COO-9 stays lead. No version bump (0.5.20 / coordpy.sdk.v3.43); no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "cd3cd8e"
      ],
      "docs": [
        "docs/RUNBOOK_W134.md",
        "docs/RESULTS_W134_DEPLOYABLE_COMPLEXITY_WITNESS_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W134_V1.md"
      ],
      "linear_issues": [
        "COO-59",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W134-T-DEPLOYABLE-COMPLEXITY-WITNESS-INSTRUMENT-MINTABLE",
        "W134-T-DEPLOYABLE-WITNESS-FAITHFUL-TO-EXACT-ORACLE-ON-NAIVE-REF-SEPARATION",
        "W134-T-DEPLOYABLE-COMPLEXITY-CORPUS-CID-TIMEOUT-INVARIANT",
        "W134-T-DEDICATED-COMPLEXITY-FIELD-B0-CEILING-HIGH-AND-SINGLE-FAMILY-HEADROOM",
        "W134-T-DEPLOYABLE-COMPLEXITY-WITNESS-IS-REAL-BUT-SUB-ORACLE-ON-HELD-OUT-DEV",
        "W134-T-CONTROLLER-ABSTAIN-DISCIPLINE-NEUTRALISES-WITNESS-ON-COMPLEXITY-FIELD",
        "W134-L-DEPLOYABLE-COMPLEXITY-WITNESS-DEV-CAP"
      ]
    },
    {
      "id": "W135",
      "title": "oracle-derived solution-STRUCTURE witness + dedicated NON-COMPLEXITY (WRONG_ALGORITHM + SEARCH_ENUM) held-out corpus + held-out dev bench — instrument REAL but the structure witness TIES the counterexample; frontier rerun NOT earned (coordpy.solution_structure_witness_v1 + noncomplexity_structure_corpus_v1)",
      "outcome": "W133 showed the EW1 COUNTEREXAMPLE witness adds +0.00 over blind reflexion on WRONG_ALGORITHM/SEARCH_ENUM (the wrong-algorithm capability ceiling); W135 asked whether oracle-derived SOLUTION STRUCTURE breaks it. Lane alpha (SUCCESS, $0): solution_structure_witness_v1 (SW1 greedy-failure certificate / SW2 optimal-substructure ladder [exploratory+leak-constrained per Pu OOPSLA'11 + KNARsack] / SW3 search-frontier exact-vs-naive count / SW4 controller [LEAD] = union + SolidCoder shift-left), anchored on a token-minimal fresh disjoint counterexample (reuses W133 EW1); research CHANGED the FORM (PGS arXiv:2506.18315 minimal+attribution+property; LDB arXiv:2402.16906 => genuinely new in FORM). No-leakage ENFORCED+TESTED (executes ref/naive/brute but never renders source; counterexample + every sub-instance byte-disjoint from secret_cases; never the recurrence/state; graded on the DISJOINT hidden bank). noncomplexity_structure_corpus_v1: 16 families x 5 seed-disjoint seeds/split => admitted 80/79/80/78 (>= 36/36/36/30); corpus_cid 306610ae / eval_split 3f6e3e59 / frontier_slice 8aa53564 LOCKED; held-out integrity TRUE after cross-split dedup of seed-independent fib_no_adjacent collisions; naive/ref separation 80/80, genuinely-new-vs-EW1 54/80, all leakage-clean. Held-out DEV (16 problems = 8 SE + 8 WA, Llama-3.3-70B, 336 NIM, decisive LEAD subset A0/A1/B0/C1/S4 per a §9 latency amendment): A0=A1=B0=C1=S4=81.25% (13/16); S4-C1=S4-B0=+0.00; 0 rescues over C1 (0 modes/0 families). A0 single-shot already 13/16 => near-zero feedback headroom at 70B; the SAME 3 traps (se_lattice_paths_blocked/wa_knapsack_01/wa_weighted_interval_scheduling) are CAPABILITY-bound for EVERY arm; the structure witness fired genuinely-new+leakage-clean on the 4/16 failed problems but the model converted NONE. §7a dev gate FAILS => $0 eval, $0 frontier; the wrong-algorithm ceiling is CAPABILITY-bound NOT feedback-form-bound at 70B (sharpens W133). Stronger-model gate CLOSED (258b6ed7 invariant {KNOWN:1,UNKNOWN:4}; new MiniMax-M3 no published card). W89 (+5.56) + W105 (+7.00) STILL the only two retirements; W135 retires none. No version bump (0.5.20 / coordpy.sdk.v3.43); no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "075bbd1"
      ],
      "docs": [
        "docs/RUNBOOK_W135.md",
        "docs/RESULTS_W135_STRUCTURE_WITNESS_BENCH_V1.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W135_V1.md"
      ],
      "linear_issues": [
        "COO-60",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W135-T-SOLUTION-STRUCTURE-WITNESS-INSTRUMENT-MINTABLE",
        "W135-T-NONCOMPLEXITY-STRUCTURE-CORPUS-MINTABLE",
        "W135-T-NONCOMPLEXITY-FIELD-LOW-HEADROOM-AT-70B",
        "W135-T-STRUCTURE-WITNESS-TIES-COUNTEREXAMPLE-ON-HELD-OUT-DEV",
        "W135-L-SOLUTION-STRUCTURE-WITNESS-DEV-CAP"
      ]
    },
    {
      "id": "W136",
      "title": "machine-structured algorithm-state trace instrument + ROOT-CAUSE of the W132-W135 'wrong-algorithm ceiling' = an I/O-FORMAT CONFOUND (coordpy.algorithm_state_trace_v1 + algorithm_state_trace_corpus_v1)",
      "outcome": "Lane alpha SUCCESS ($0): the machine-structured algorithm-state trace instrument (AT1 decision-path / AT2 subproblem-state incl. the full 2-D StateGridV1 DP table from black-box oracle calls / AT3 search-frontier / AT5 typed capsule; AT4 EXCLUDED for leakage) + family-aware _typed_subinstances_v1 + execution-grounded I/O-repair arm T_IO + fresh-seed corpus (80/79/79/78 >= floors; corpus_cid ce1a6bc6 / eval 13519353 / frontier 3f75b302 LOCKED; 18 tests). Lane beta: T1 (incl. the full 2-D 0/1-knapsack DP table + recurrence scaffold) cracks 0/3 traps like prose S4 and counterexample C1; root-cause diagnostic found the W132-W135 'wrong-algorithm capability ceiling' is an I/O-FORMAT CONFOUND -- the model writes CORRECT algorithms but misparses the battlefield's whitespace-flattened input. Quadruple-confirmed: (1) corpus easy/hard split = standard/non-standard I/O; (2) same 0/1-knapsack DP passes 7/7 robust vs fails 7/7 per-line; (3) T_IO I/O-repair cracks 3/3 traps (+18.75pp, both modes); (4) standard-I/O A0 one-shots 3/3. Parsing-only gains => section 7b condition 4 EXCLUDES them => NOT a frontier earn ($0 frontier). REVISES W133-L + W135-L; the generated WA/SE battlefield is I/O-confounded + (I/O-fixed) low-algorithm-headroom. Lane gamma: stronger-model gate CLOSED (258b6ed7 invariant); graphify START+END (83922 nodes; new module = REAL 1-hop bridge to 6 modules). W89 (+5.56) + W105 (+7.00) STAND as the only two retirements (REAL HumanEval, outside the confound); W136 retires none, earns no frontier, but corrects a four-milestone misinterpretation + lands the I/O-repair mechanism. No version bump (0.5.20 / coordpy.sdk.v3.43); no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "7bbcb2b"
      ],
      "docs": [
        "docs/RUNBOOK_W136.md",
        "docs/RESULTS_W136_ALGORITHM_STATE_TRACE_BENCH_V1.md"
      ],
      "linear_issues": [
        "COO-61",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W136-T-ALGORITHM-STATE-TRACE-INSTRUMENT-MINTABLE",
        "W136-T-MACHINE-STRUCTURED-STATE-FEEDBACK-DOES-NOT-LIFT-APPARENT-CEILING",
        "W136-T-WRONG-ALGORITHM-CEILING-ON-GENERATED-WA-SE-IS-IO-FORMAT-CONFOUND",
        "W136-T-EXECUTION-GROUNDED-IO-REPAIR-CRACKS-ALL-TRAPS",
        "W136-T-STANDARD-IO-A0-ONESHOTS-ALL-TRAPS",
        "W136-L-GENERATED-WA-SE-BATTLEFIELD-IS-IO-CONFOUNDED-AND-ALGORITHM-LOW-HEADROOM"
      ]
    },
    {
      "id": "W137",
      "title": "parser-neutral HARD battlefield v2 + model-ladder hardness calibration + bounded complexity-mechanism diagnostic (coordpy.parser_neutral_io_v1 + hard_battlefield_slate_v2 + model_ladder_calibration_v1 + hard_battlefield_corpus_v2 + repaired_field_mechanism_bench_v1)",
      "outcome": "Lane alpha ($0 build SUCCESS): the parser-neutral I/O kernel (canonical one-item-per-line normal form + HC1 dual-parser-agreement gate => the W136 confound is machine-impossible to recur; confound-regression fixture BITES) + 17 parser-neutral templates / 4 modes (slate_fingerprint_cid 2ce207c5; all 17 pass HC1+HC2+HC5; deterministic + timeout-invariant) + the model-ladder hardness-calibration harness (HC3/HC4) + corpus assembler. Lane alpha-CALIBRATION (NIM; ladder llama-3.1-8b + llama-3.3-70b; calibration_cid 8b0f6231; wall 2735s) -- DECISIVE: at 70B the field's difficulty is BIMODAL, every template's strong pass@5 (A1) is EXACTLY 1.00 or 0.00: 12/17 HC3-SATURATED (one-shot; ALL wrong-algorithm+hidden-edge+search-enum modes), 4/17 HC4-DEAD (A0=A1=0), 1/17 band => 1/17 surviving << 40/40/40/30 floors = THRESHOLD MISS; no template has 0<A1<1. GENERALIZES W136. Lane beta (bounded complexity-mechanism diagnostic, n=3; A0=0.00 A1=B0=33.33% M1[complexity witness]=66.67%, M1-B0=+33.33pp): M1 RESCUED a dead complexity template reconfirming the W133 complexity witness on a PARSER-NEUTRAL field, but SINGLE-MODE => §7a FAILS by SPAN => NOT earned, $0 frontier (small-N). Lane gamma: stronger-model gate re-derived IN CODE => NO_CERTIFIABLE_STRONGER_MODEL, decision CID 258b6ed7 byte-identical. NET: at 70B clean algorithmic difficulty is BIMODAL (one-shot or impossible), leaving NO mechanism-sensitive multi-mode headroom -- the structural reason resistant code superiority is 0-clean; W89 (+5.56) + W105 (+7.00) STAND BECAUSE HumanEval(+) HAS a middle band. W89 + W105 remain the only two retirements; W137 retires none. No version bump; no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "7bed040"
      ],
      "docs": [
        "docs/RUNBOOK_W137.md",
        "docs/RESULTS_W137_PARSER_NEUTRAL_BATTLEFIELD_V2.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W137_V1.md"
      ],
      "linear_issues": [
        "COO-62",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W137-T-PARSER-NEUTRAL-IO-KERNEL-MINTABLE",
        "W137-T-HARD-BATTLEFIELD-SLATE-V2-MINTABLE",
        "W137-T-MODEL-LADDER-CALIBRATION-INSTRUMENT-MINTABLE",
        "W137-T-REPAIRED-FIELD-DIFFICULTY-IS-BIMODAL-AT-70B",
        "W137-T-COMPLEXITY-WITNESS-IS-REAL-BUT-SINGLE-MODE-ON-PARSER-NEUTRAL-FIELD",
        "W137-L-REPAIRED-FIELD-NO-MECHANISM-HEADROOM-BAND-AT-70B"
      ]
    },
    {
      "id": "W138",
      "title": "headroom-band battlefield search + cross-scale mechanism validation + architecture-requirements extraction (coordpy.headroom_band_slate_v3 + headroom_band_calibration_v2 + headroom_band_corpus_v3 + band_mechanism_bench_v1)",
      "outcome": "REVISES W137. The W137 'BIMODAL field / no mechanism-headroom band at 70B' reading was an ARTIFACT: A1 was measured with n_a1=1 (a single any-of-K draw => a1_rate in {0,1} by construction) and the hidden-N knob was pinned past the complexity cliff. Lane alpha (4 new leaf modules reusing the W137/W120 machinery verbatim): measuring A1 as a POPULATION RATE over n_a1=8 + Wilson + HB3 intermediate-band + HB4 cross-scale-discrimination, over a CONTINUOUS hidden-N grid (metabench arXiv:2407.12844 / CLRS-Text arXiv:2406.04229). Calibration (8b+70b; 180 NIM; calibration_cid ff87cb60) lands a REAL 0<A1<1 band: 5/18 cells => 3 families / 2 modes (count_pairs_sum_le_t@20000 a1=0.60 -- it read A1=1 SATURATED in W137) => REVISES W137-T-REPAIRED-FIELD-DIFFICULTY-IS-BIMODAL-AT-70B + W137-L (the field is NOT inherently no-headroom). Lane beta: a $0 de-risk found the counterexample witness (EW1) cannot fire on functional families (small-probe-only search) => the 2nd mode is firing-bound (mechanistically explains W133's EW1 +0; R3/R7/R9); operative lead = complexity witness C0 (the X1 controller is DOMINATED by C0). On the FRONTIER ANCHOR (llama-3.3-70b): §7a dev (5 complexity families, n=3) C0-A1=+53.33, C0-B0=+46.67, rescues span ALL 5 families, PASS; §7b HELD-OUT eval (n=4) C0-A1=+40.00, C0-B0=+25.00, span >=3 families = the FIRST time the W133/W137 single-family complexity witness clears the §7a/§7b SPAN gate (multi-FAMILY). BUT the §7b two-tier robustness check FAILS: the witness HURTS the small tier (llama-3.1-8b C0-A1=-25.00, capability-gated; arXiv:2404.17140/arXiv:2310.01798) AND the witness-over-REFLEXION edge does NOT reproduce on a capable proxy (llama-3.1-70b near-saturated A1=70 => C0-B0=+0.00) => CAPABILITY-GATED + ANCHOR-SPECIFIC => per the discipline NO frontier rerun (§7c not run), NO third retirement. Lane gamma: docs/ARCHITECTURE_REQUIREMENTS_W138_V1.md (R1-R12); stronger-model gate re-derived CLOSED (258b6ed7 invariant). W89 (+5.56) + W105 (+7.00) STAND as the only two retirements; W138 retires none but REVISES W137-L. W139 = per-model band recalibration + capability-matched abstain/KEEP controller. No version bump (0.5.20 / coordpy.sdk.v3.43); no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "ed52f58"
      ],
      "docs": [
        "docs/RUNBOOK_W138.md",
        "docs/RESULTS_W138_HEADROOM_BAND_V1.md",
        "docs/ARCHITECTURE_REQUIREMENTS_W138_V1.md"
      ],
      "linear_issues": [
        "COO-63",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W138-T-RESISTANT-HEADROOM-BAND-IS-CONSTRUCTIBLE-AT-70B",
        "W138-T-COMPLEXITY-WITNESS-EARNS-MULTI-FAMILY-AT-THE-ANCHOR",
        "W138-T-WITNESS-GAIN-IS-CAPABILITY-GATED-AND-ANCHOR-SPECIFIC",
        "W138-T-COUNTEREXAMPLE-2ND-MODE-IS-FIRING-BOUND-AT-70B",
        "W138-L-RESISTANT-BAND-WITNESS-NOT-CROSS-TIER-ROBUST"
      ]
    },
    {
      "id": "W139",
      "title": "per-tier band recalibration + capability-matched witness compiler + cross-tier mechanism validation (coordpy.per_tier_band_calibration_v1 + capability_matched_witness_compiler_v1)",
      "outcome": "W138 found the complexity witness earns multi-family AT THE ANCHOR but is capability-gated (8B -25) + anchor-specific (mid +0). W139 attacked both. Lane alpha (per_tier_band_calibration_v1 measures A1-as-a-RATE at EVERY tier; ladder V2 8b/3.1-70b/3.3-70b; 221 NIM; per_tier_calibration_cid 72c4a4f1) ROOT-CAUSES the mid +0: llama-3.1-70b is SATURATED on the complexity family (a1=1.00 at N=20000 AND 50000) => STRONGER than the anchor => no headroom (R4' per-ability; Fluid Benchmarking arXiv:2509.11106); witness-eligible: strong only. Lane beta (capability_matched_witness_compiler_v1; dev n_per_cell=2, diagnostics all tiers, 210 NIM): the capability-matched controller ACHIEVES NON-NEGATIVITY on every tier (KEEP on the ineligible 8B+mid => Cm-A1=+0 on the 8B, the W138 -25 harm ELIMINATED; also protects the 8B from blind reflexion's harm B0 25%->0%, Cm held 25%); R2'/R5'. The large-probe counterexample (build_large_probe_set_v1, cap 400->6000) REVIVES the W138-dead 2nd mode (Nb=75% at the anchor; W138 N0=0 was a probe-budget artifact). BUT the §7a dev gate FAILS by SPAN_FAIL (anchor Cm-A1=+75, Cm-B0=+25 margins pass but rescues over strong blind reflexion span 1 mode/1 family); the apparent 2nd positive tier (mid Cm-A1=+50) is KEEP-tier SAMPLING NOISE (n=2; Cm==plain self-consistency; Nb=0). Structural cause: NO second tier on this ladder has BOTH headroom AND witness-usability (mid saturated, 8B capability-bound) => $0 eval, $0 frontier, NO third retirement. Sample sizes tiny (n=2-4) => noisy DEV signals. Lane gamma: ARCHITECTURE_REQUIREMENTS_W139_V2 (R2'/R4'/R5'); stronger-model gate re-derived CLOSED (258b6ed7 invariant {KNOWN:1,UNKNOWN:4}). 431 total NIM. W89 (+5.56) + W105 (+7.00) STAND as the only two retirements; W139 retires none but eliminates the W138 harm, root-causes the mid +0, revives the 2nd mode. W140 = a denser/different model ladder (a tier between 8B and 70B with both headroom and usability) / a mechanism independent of the single-tier headroom-and-usability coincidence / a primary-KNOWN stronger model when the gate opens. No version bump (0.5.20 / coordpy.sdk.v3.43); no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "a13189d"
      ],
      "docs": [
        "docs/RUNBOOK_W139.md",
        "docs/RESULTS_W139_CROSS_TIER_V1.md",
        "docs/ARCHITECTURE_REQUIREMENTS_W139_V2.md"
      ],
      "linear_issues": [
        "COO-64",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W139-T-PER-TIER-BAND-CALIBRATION-MINTABLE",
        "W139-T-MID-TIER-SATURATED-ON-COMPLEXITY-ROOT-CAUSES-W138-MID-PLUS-ZERO",
        "W139-T-CAPABILITY-MATCHED-CONTROLLER-ACHIEVES-NON-NEGATIVITY",
        "W139-T-LARGE-PROBE-REVIVES-COUNTEREXAMPLE-2ND-MODE",
        "W139-L-CAPABILITY-MATCHED-NON-NEGATIVITY-ACHIEVED-BUT-WITNESS-GAIN-STAYS-ANCHOR-CONCENTRATED"
      ]
    },
    {
      "id": "W140",
      "title": "teacher-compiled family-level witness tutors + weak-model lift + cross-tier usefulness (coordpy.family_tutor_compiler_v1 + cross_tier_tutor_bench_v1)",
      "outcome": "W139 left the blocker as weak-model witness USABILITY (8B raw-witness usability 0.00). W140 compiled the anchor's witness into a family-LEVEL holed-skeleton TUTOR and ACHIEVED the milestone's stated goal: a same-budget WEAK-MODEL lift of +75-100pp on 3 needed-and-unknown-technique families (subarrays_sum_and_range@30000 two-deque 25->100, sum_nearest_smaller_left@50000 monotonic-stack 0->100, kth_smallest_pair_distance@20000 binary-search-on-answer 0->75; max_j@50000 A1=50% substitute DROPPED). The tutor REVERSES the W139 usability floor (all tiers eligible vs 0.00; tutor_usability_cid 849f1ebf). The dev TIE was a SUBSTITUTE regime (a $0 failure analysis proved the tutor genuinely shifts the 8B's distribution to the technique, two-deque 0/60->24/24 public; the lift hides where A1~=0). LEAD re-identified as T2 always-apply-scaffold (the W139-style controller T4 is dominated - its public-revert can't see a hidden-only bug, W125/W129). The mechanism is ADDITIVE iff A1~=0, SUBSTITUTIVE otherwise. Anchor (frontier target) ALSO qualifies on cell-1 (A1=25%) so a strict anchor^weak earn is PLAUSIBLE but the NIM endpoint degraded mid-run => PARTIAL/endpoint-blocked. Lane gamma: ARCHITECTURE_REQUIREMENTS_W140_V3 adds R6' (cross-capability teaching-compilation); stronger-model gate 258b6ed7 invariant {KNOWN:1,UNKNOWN:4}; graphify START+END. W89(+5.56)+W105(+7.00) STAND as the only two retirements (a retirement is multi-agent frontier superiority; W140's win is a teacher-compiled weak-model lift, distinct+complementary). 2 modules + 5 scripts + 27 tests; no version bump; no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "225784b",
        "2a28d61"
      ],
      "docs": [
        "docs/RUNBOOK_W140.md",
        "docs/RESULTS_W140_CROSS_TIER_V1.md",
        "docs/RESULTS_W140_RESEARCH_V1.md",
        "docs/RESULTS_W140_TUTOR_DESIGN_V1.md",
        "docs/RESULTS_W140_MILESTONE_SUMMARY_V1.md",
        "docs/ARCHITECTURE_REQUIREMENTS_W140_V3.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W140_V1.md"
      ],
      "linear_issues": [
        "COO-65",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W140-T-FAMILY-TUTOR-COMPILER-MINTABLE",
        "W140-T-TUTOR-LEAK-GATE-FALSIFIABLE",
        "W140-T-COMPILED-TUTOR-REVERSES-WEAK-TIER-USABILITY-FLOOR",
        "W140-T-COMPILED-TUTOR-ACHIEVES-CROSS-TIER-SAFETY-BY-CONTENT",
        "W140-T-COMPILED-TUTOR-LIFTS-WEAK-TIER-+75-100PP-IN-NEEDED-UNKNOWN-REGIME",
        "W140-T-TUTOR-GAIN-ADDITIVE-IN-NEEDED-UNKNOWN-SUBSTITUTIVE-OTHERWISE",
        "W140-T-ALWAYS-APPLY-SCAFFOLD-BEATS-PUBLIC-REVERT-CONTROLLER-ON-HIDDEN-FAILURE",
        "W140-T-CROSS-TIER-USEFULNESS-IS-FAMILY-DEPENDENT-CONFIRMED-ON-THE-SHARED-DIFFICULTY-TECHNIQUE",
        "W140-L-STRICT-ANCHOR-WEAK-≥2-SHARED-FAMILY-EARN-1-OF-2-CONFIRMED-CELL4-ENDPOINT-BLOCKED"
      ]
    },
    {
      "id": "W141",
      "title": "emergent no-oracle self-tutoring (self-derived scaffold from the team's own verified-correct sample) (coordpy.self_tutoring_technique_extractor_v1 + no_oracle_verifier_v1 + self_tutoring_controller_v1)",
      "outcome": "W140 lifted the frontier with an EXACT-ORACLE-compiled scaffold (teaching). W141 removes the oracle: the team generates K diverse -> a NO-ORACLE verifier (S1 self-written-brute agreement on a constraint-adversarial bank AND S2 worst-case dual-large-input timeout, oracle-free) picks the correct+efficient winner -> the holed-skeleton scaffold is SELF-DERIVED from that winner (AST blanks the discriminating decision; correct_fill = the winner's own expressions => completable by construction; needs INTER-PROCEDURAL return-tracing for real models' helper-fn shape; 8/9 real 70B programs self-derive a clean leak-passing scaffold; buggy DISCARDED -> KEEP==A1) -> amortized. THREE established: (1) DISCOVER runs end-to-end on the live frontier 70B with ZERO oracle (first oracle-free technique mint on a frontier model); (2) the verifier is BYTE-IDENTICAL to the hidden grader (0 FP/0 FN on 115 REAL 70B samples; S1 catches 8/8 real fast-but-wrong the efficiency signal alone accepts = W125 escape; adversarial bank 28/30 vs public 0/30); (3) the scaffold lifts generation to q=100%. WIN: on count_pairs_sum_le_t (p=0.08, a family the 70B LACKS) the scaffold lifts 8%->100% = +70.6pp per-problem same-budget superiority over no-oracle verified-selection (ST 3/3 vs B 1/3), the FIRST clean resistant same-budget win (W125-W130 got zero). WALL: NOT a §7b retirement -- the amortization win needs a MODERATE-p band but the frontier's generation is BIMODAL (W137 wall at generation level): 70B KNOWS the technique (high-p, verified-selection saturates) or is CLUELESS (0-supply); exactly 1 of the tested slate is moderate-p => >=3-family span UNMET; advantage concentrates at low per-problem budget (ST-B=(1-p)^K). W89+W105 STAND as the only two retirements. 3 modules + 2 scripts + offline analyzer + 10 tests; no version bump; no PyPI; coordpy/__init__.py untouched (one additive spec_override).",
      "commits": [
        "21f61f8"
      ],
      "docs": [
        "docs/RUNBOOK_W141.md",
        "docs/RESULTS_W141_SELF_TUTORING_V1.md"
      ],
      "linear_issues": [
        "COO-66",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W141-T-NO-ORACLE-SELF-TUTORING-DISCOVER-WORKS-ON-THE-FRONTIER",
        "W141-T-NO-ORACLE-VERIFIER-IS-BYTE-IDENTICAL-TO-THE-HIDDEN-GRADER",
        "W141-T-SELF-DERIVED-SCAFFOLD-LIFTS-GENERATION-TO-Q-100PCT",
        "W141-T-EXTRACTOR-REQUIRES-INTERPROCEDURAL-RETURN-TRACING",
        "W141-T-FIRST-CLEAN-NO-ORACLE-RESISTANT-SAME-BUDGET-SUPERIORITY",
        "W141-T-SELF-TUTORING-IS-SUPPLY-LIMITED-NOT-VERIFIER-LIMITED",
        "W141-T-NEUTRAL-BASELINE-REQUIRED",
        "W141-L-RESISTANT-RETIREMENT-BLOCKED-BY-BIMODAL-FRONTIER-GENERATION"
      ]
    },
    {
      "id": "W142",
      "title": "moderate-p family screen + high-K discover reliability + ≥3-family retirement attempt (coordpy.moderate_p_family_slate_v1 + moderate_p_family_screen_v1 + discover_amortize_accounting_v1)",
      "outcome": "Lane α SUCCESS (the moderate-p band EXISTS: 3 families / 3 veins / 2 modes) but Lane β CAP (RETIREMENT NOT EARNED, ST-B0=+0.0pp). W141 diagnosed the binding constraint as moderate-p family SUPPLY; W142 attacked it. Lane α DECISIVE $0 result (lean validation FAILS:NONE): the G3 gated-accumulator gate (compiled AND n_pred_holes>=1) ADMITS counting-pair (sort+two-pointer) + two-deque veins and REJECTS prefix-hash (technique in dict-maintenance) + binary-search-on-answer (printed answer is a reassignment) => the teaching-compiler representation R6' is a binding constraint on R7'. NIM screen (7 parallel, K=12, fair neutral-prompt bank, band [0.10,0.50]): admits exactly 3 families — count_pairs_sum (p=0.333), subarrays_sum_and_range (two-deque HIDDEN_EDGE, p=0.167), sum_nearest (monotonic, p=0.500) — while ALL 4 new/sibling surfaces = p=0.000 (0/12) => a THIN FIXED band (advances W141's 1 family, confirms the bimodal structure). Lane β (parallel discover->amortize, K_a=4, M=4, raised discover-K, alien-vein NEG): earns ZERO families, each a DISTINCT wall — sum_nearest B0-SATURATES at p=0.5; count_pairs_sum low-p S1 FALSE-NEGATIVE (K_d=48 => 0 winners despite screen 4/12 secret-pass; self-brute-convention-sensitive); subarrays 2nd-mode VERIFIER FALSE-POSITIVE (W125 looks-right-fails-hidden, non-extractable, discarded). NEG no-lift holds. => the conversion is VERIFIER-RELIABILITY + BASELINE-SATURATION bound: win-size and discovery-difficulty are COUPLED via p. Lane γ: ARCHITECTURE_REQUIREMENTS_W142_V4 adds R7' (discover-then-amortize head + per-family supply estimator); stronger-model gate CLOSED (258b6ed7 invariant {KNOWN:1,UNKNOWN:4}). W89 (+5.56) + W105 (+7.00) STAND as the only two retirements; W142 retires none but converts W141's single-family observation into a systematic machine-diagnosed understanding. no version bump (0.5.20); no PyPI; coordpy/__init__.py untouched.",
      "commits": [
        "e4f1067"
      ],
      "docs": [
        "docs/RUNBOOK_W142.md",
        "docs/RESULTS_W142_MODERATE_P_SCREEN_V1.md",
        "docs/ARCHITECTURE_REQUIREMENTS_W142_V4.md",
        "docs/FRONTIER_RELEVANCE_AUDIT_W142_V1.md"
      ],
      "linear_issues": [
        "COO-67",
        "COO-9",
        "COO-6"
      ],
      "carry_forwards_retired": [],
      "carry_forwards_added": [
        "W142-T-MODERATE-P-FAMILY-SCREEN-MINTABLE",
        "W142-T-SELF-TUTORING-EXTRACTABILITY-NARROWS-TO-GATED-ACCUMULATOR-VEINS",
        "W142-T-MODERATE-P-BAND-IS-A-THIN-FIXED-SET-3-VEINS-2-MODES",
        "W142-T-DISCOVER-THEN-AMORTIZE-FAILS-FOR-THREE-DISTINCT-REASONS-AT-THE-FRONTIER",
        "W142-L-DISCOVER-THEN-AMORTIZE-CONVERSION-IS-VERIFIER-RELIABILITY-AND-BASELINE-SATURATION-BOUND"
      ]
    }
  ]
}