Skip to content

GCL OAI-26 WP04 — adversarial source and mathematics review of high-risk families #333

Open

Description

@fyremael

WP04: executable zero-context assignment

Programme: grandchallenge/MATH-PROGRAMME#2850
Exact upstream: openai/math@1581d4cbdd69a0faf34b2089614fe2a814316a6a; GCL catalog sources/OPENAI-MATH-2026/, provider catalog/providers/OAI-MATH-1581D4CB.json. The provider is metadata-only SOURCE_LOCKED. GCL's original 372 families / 719 PDF links, crosswalk of 200 comparator/declaration pairs, publisher's 300 formalized top-line claims must not be conflated. Start by reading AGENTS.md and relevant issue evidence, reuse independent logical roles within existing authenticated identity, protected PRs only; no MATHCERT or research promotion. Return RESULT/1 with exact SHA, evidence, scoped disposition, reproducible commands, PR and protected readback.

Task: Independently challenge bounded WP02/WP03 semantic verdicts using non-authoring read-only adversarial and mathematical specialist role records. Select at least 12 families or all high-risk ones if fewer are currently documented; select at least 3 from revision/withdrawal dependencies, 3 from claimed formal matches, 3 from unfamiliar mathematical domains, and 3 random baseline (overlap permitted only with disclosed deduplication and reason). Source-lock primary literature used to assess known results and novelty (novelty is NOT certified here).

For every reviewed case, ask if Lean statement actually implies the manuscript's conclusion under those assumptions; attempt to find vacuous hypotheses, sign/index/convention errors, reversed implication, reliance on imported strong axioms, unsound literature attribution, excluded edge cases, or withdrawn historical results. Provide exact countermodel/counterexample when available, otherwise a sharply typed unresolved semantic or mathematical proof obligation. Do not attempt wholesale original proof repair unless a specific outcome can change a verdict and Programme authorizes a bounded new target.

Acceptance: >=12 distinct case records or a documented smaller exhaustive risk set, review identity/role separated from each author's output, source-linked adversarial findings, at least one deliberate hostile semantic fixture demonstrating verdict rollback, and replay/revision demands where warranted. Distinguish mathematical refutation from evidence of mismatch, and independent role-scoped audit from independently authenticated parties. All unresolved cases remain open. MATHCERT reserves higher assurance gates.

Prerequisites: consume WP02 risk register and WP03 first wave; may start source review of already known 342 revision divergence before those reports without reopening its full conjecture. Terminal: ADVERSARIAL_WAVE_ACCEPTED or precise evidence blocker, never blanket certification.

RESULT format: RESULT/1\nASSIGNMENT_ID: GCL-OAI-26-WP04\nSOURCE_SHA: 1581d4cbdd69a0faf34b2089614fe2a814316a6a\nDISPOSITION: ...\nEXACT_HEAD: ...\nEVIDENCE: ...\nREPLAY: ...\nUNRESOLVED: ...\nCLAIM_BOUNDARY: NO_MATHCERT_CERTIFICATION. Do not mark a purported protected admission complete without independent readback.

Activity

  1. fyremael commented on Oct 25, 2026

    @fyremael
    CollaboratorAuthor

    GCL-OAI26-WP04/BOUNDED-ADVERSARIAL-REVIEW-INPUT-001

    Read-only source criticism preparation, not a completed independent mathematical certification or a review of unpublished sibling returns.

    Exact upstream openai/math@1581d4c. Inspected:

    • lean/docs/003.md blob cf794e93da2e1a494434153f5b91f2990fe9fd60;
    • lean/OAI/NumberTheory/DirichletL/Nonvanishing.lean blob 46c67f4000906b1391a471d6e6396ac0da3c16a5;
    • lean/OAI/NumberTheory/DirichletL/Detector/FinalAssemblyUnconditional.lean blob e7c20910424837fb954796e1216f5aa453bbd5da;
    • lean/ComparatorChallenges/{QuasiRiemannHypothesis,DirichletSevenEighths,SiegelZeros}.lean pinned at that source SHA.

    Initial hostile questions / exact falsification gates:

    1. Source-level Family 003 zero-free Re(s)>7/8 does not prove full Riemann Hypothesis (Re(s)=1/2): reject any automated verdict or public summary making that inference.
    2. Dirichlet theorem has ¬(χ=1 ∧ s=1), a real exceptional condition; reject any comparator/manuscript mapping deleting it. Examine whether the actual Lean definition of LFunction at a principal pole matches manuscript conventions.
    3. The doc's uniform real-zero gap is c>0 with c≤(1-β)log q for q≥3; it excludes only a logarithmic neighborhood near β=1, not every real zero β∈(0,1). Reject stronger summaries.
    4. QuasiRiemannHypothesis and SiegelZeros comparator challenge files contain intentional sorry placeholders; reject any validator that counts those as proof of solution. Inspect solution theorem and full imported axiom closure instead, including Foundation, CertifiedExistence and FinalAssemblyCertifiedBands dependencies.
    5. Family 342 September/October cone-identity change is a separate claim from the unchanged main existence theorem. A single matched main-result comparator must not bless the corrected cone theorem. Historical September references must be version-bound.
    6. Three upstream withdrawn Hodge/Weil results are not in the active 719-title index; reject any automatic inclusion or inheritance of mathematical certainty through similarly named surviving families such as 032.
    7. Multiple Lean declarations can share one Comparator config. Eight repeated config paths were observed; no universal 1:1 assumption may be used in pass-count accounting.

    Prioritized adversarial targets: Family 003 imported-axiom boundary and proposition semantics; Family 342 cone revision; withdrawn Hodge dependency network; Family 002 BSD citation/source-edition drift; Families 232–233 Lipschitz/Ashkin–Teller repairs; 034/036/056 Kähler repairs; 376 torus-projection/clock corrections.

    Not yet reviewed: twelve distinct family verdicts, primary-paper proof bodies, non-authoring hostile fixture results, or post-WP03 independent critical disposition. No claim of an accepted adversarial wave.

    Disposition: CRITIC_TEST_PLAN_BOUNDED, NOT_ADJUDICATED.

  2. fyremael commented on Nov 10, 2026

    @fyremael
    CollaboratorAuthor

    RESULT/1
    ASSIGNMENT_ID: GCL-OAI-26-WP04/FIRST-WAVE-12
    SOURCE_SHA: 1581d4cbdd69a0faf34b2089614fe2a814316a6a
    DISPOSITION: FIRST_WAVE_12_SOURCE_BOUND_LOGICAL_CRITICAL_CASES_PROTECTED; FULL_WP04_OPEN
    EXACT_HEAD: grandchallenge/MATHFORGE 5384ae9
    EVIDENCE: PR #543 merged at 5384ae9; exact merged first-wave machine ledger sources/OPENAI-MATH-2026/WP04_FIRST_WAVE_12_ADVERSARIAL_SCOPE_REVIEW.json blob 7059e1aa848794139428dc3a95557a435cbeec70; hostile checks tests/test_openai_math_wp04_first_wave.py blob 32a93bb31e6a1dffbf3fb9fa40bf1945afb30e3a; CI owner governance/ci_components/oai26_wp04.json blob 8c40176dc119e9faeb19d37cc5cc91661848635c; non-authoring logical role critical pass PR comment 6763133365.
    REPLAY: All 12 indexed case identities and their source-record blobs read back at exact protected merge. Selection strata are 3 revision/dependency, 3 formal-match, 3 unfamiliar domains and 3 disclosed random baseline; no duplicates. Hostile checks include unsupported Family 011 epsilon>=1-to-all-positive-epsilon promotion, Family 009 injectivity-to-bijectivity inference, Family 342 old unconditional versus new conditional cone equality, false source identity, false certification, false mathematical counterexample, and false externally independent role. Protected routing, reconnaissance, replay, policy, security and CodeQL checks passed before merge. New mathematical refutations 0; new certifications 0.
    UNRESOLVED: These 12 records are logically adversarial reviews of already-protected WP02/WP03 scope evidence, not fresh primary-paper complete mathematical proof and bibliography/literature reviews. No separately independently authenticated reviewer. Obtain per-case pinned primary source proof examinations and source-linked falsifiable bridges, then an actual terminal WP04 mathematical adjudication. Parent WP04 #333 and Programme #2850 remain open; WP05 final family verdict review remains separate.
    CLAIM_BOUNDARY: FIRST_WAVE_DOCUMENTARY_AND_LOGICAL_CRITIC_ONLY; NO_MATHCERT_CERTIFICATION.

  3. fyremael commented on Nov 10, 2026

    @fyremael
    CollaboratorAuthor

    GCL-OAI26-WP04/F011-PRIMARY-SOURCE-CRITICAL-READBACK/1

    Fresh primary-source inspection, distinct from the preceding WP03 ledger reuse: the upstream openai/math@1581d4cbdd69a0faf34b2089614fe2a814316a6a manuscript introduction preprints/Weighted-Dilation-Graphs-Smooth-Shifted-Primes-and-Totient-Fibers-September-24-2026/build/sections/00-introduction.tex (Git blob 3d9d4c55c1f30d7736e7c7427b97cb0461f06040) states ∀ ε>0, infinitely many positive integers n with g(n)>n^(1-ε), and separately a smooth prime predecessor count at every fixed 0<δ<1/4, # {p prime: 2x<p≤5x, P⁺(p-1)≤x^δ}≥x^(1-o(1)). Both are source assertions, NOT independently proved in this review.

    Actual Lean source inspection: lean/OAI/NumberTheory/TotientFibers/FiniteFiber.lean (Git blob 217cfe4d7f9c39981eabe9bbeeb8458febb0fc76) defines the fiber multiplicity g, proves two_le_g_prime_pred (at least 2 for p-1), and totient_large_fibers_of_one_le {ε : ℝ} (hε : 1 ≤ ε). Its proof invokes infinitely many primes, obtains n≥1, then applies Real.rpow_le_one_of_one_le_of_nonpos to 1-ε≤0 before using g(p-1)≥2. Consequently this exact proof technique depends essentially on ε≥1; it does not justify the open interval 0<ε<1, where n^(1-ε) grows with n.

    The adjacent finite_pigeonhole theorem has explicit hprime, hsize, hsmooth and hcount premises. It proves a conditional fiber lower bound given a finite set of smooth prime predecessors and a combinatorial inequality; it does not itself assert the manuscript's unconditional x^(1-o(1)) supply. This constitutes a specific statement-scope witness, not a counterexample to the manuscript theorem or a proof that no other matching Lean theorem exists.

    Critical disposition: DIRECT_PRIMARY_STATEMENT_VS_SELECTED_LEAN_SCOPE_GAP. Both primary Git blobs checked at the exact same source SHA. Next mathematically decisive work is to locate and inspect any upstream Lean result establishing the prime predecessor supply and the missing 0<ε<1 transfer (or prove a no-bridge result in a genuinely exhaustive bounded target); only then consider semantic equivalence. No new refutation, independent kernel replay, external independent reviewer, or MATHCERT certification is asserted.

  4. fyremael commented on Nov 10, 2026

    @fyremael
    CollaboratorAuthor

    Family 011 — revised mathematical criticism after Lean proof

    The prior review identified that upstream totient_large_fibers_of_one_le proves only ε≥1, not the manuscript's all-ε>0 claim. This selected-theorem scope defect has now been repaired at the conditional reduction level: protected MATHFORGE PR #546 at 6259760a3f7e4a7d3f33f7196176fc6e8da914a1 contains the complete Lean 4.34.1 proof of SmoothPrimeCount → ∀ε>0, infinitely many n with g(n)>n^(1−ε), with exact source/axiom replay evidence and 15 hostile tests (see WP03 #332).

    This does not vindicate the source manuscript's analytic proof. SmoothPrimeCount is an explicitly assumed, fixed-δ asymptotic theorem; its analytic proof and comparison with the literature remain unverified. The correct current Family 011 critical status is CONDITIONAL_SEMANTIC_BRIDGE_KERNEL_VERIFIED / UNCONDITIONAL_ANALYTIC_INPUT_PRIMARY_REVIEW_OUTSTANDING, not either NO_FORMAL_BRIDGE or FULL_SOURCE_CLAIM_PROVED. No independent external reviewer or MATHCERT certification. Retain WP04 #333 open for source-analytic, literature, and other-family criticism.

  5. fyremael commented on Nov 11, 2026

    @fyremael
    CollaboratorAuthor

    GCL-OAI26-F011/SEMANTIC-IDENTITY-BRIDGES-PROTECTED-MERGE-READBACK/1 — 2026-11-11T13:58Z

    The two remaining external-to-GCL formal statement bridge obligations have been protected-merged and read back from exact Git blobs:

    1. MATHFORGE PR #548 merged main@e1ee3957b54d4081bcbbda4bf3866056f41c935a: real-x fixed-δ smooth shifted-prime count -> GCL natural-x count -> conditional GCL all-ε>0 totient-fiber conclusion. Source WP03_FAMILY011_EXTERNAL_REAL_TO_NAT_ADAPTER.lean blob fc4e394ccb55ce08ebb8dca21b29baa103a8374f; two Lean theorem axioms exactly [propext, Classical.choice, Quot.sound]; pinned Lean exit 0; nine hostile tests pass. This code mirrors, does not import, external supply theorem.
    2. MATHFORGE PR #549 merged main@3e83fd41c47b96a31ac7327f58182d7def60114e: external g(n):=({m | Nat.totient m=n}).ncard equals GCL g(n) for every n>0; removes the exceptional singleton target {0} from infinite external conclusion; proves all-ε>0 external Erdős 821 statement implies GCL all-ε>0 positive-target statement. Source WP03_FAMILY011_EXTERNAL_FIBER_COUNT_EQUIV.lean blob f649ebb960794a48e5668531fd38449f7a046fc0 (SHA256 30886e87944de53e0f681ccfe8d6d4a71bd5fe6607abe16e44b524bf3153c31c), exact upstream openai/math@1581d4cbdd69a0faf34b2089614fe2a814316a6a. All three new theorem axiom reports exactly [propext, Classical.choice, Quot.sound], Lean exit 0 and 9/9 hostile custody tests pass. Required exact-head PR CI PASS; nonrequired Family342 jobs were pending at protected merge, not promoted to success. Verified protected readback of source, replay JSON, tests and CI routing file.

    LIVE EXTERNAL ANALYTIC PROOF REPLAY: One real Colab CPU worker gcl-f011-xu-replay4-20261111, session gcl-f011-xu-replay4-20261111-a1, at source-locked checkpoint 2/4 and actively building cadamcat/totient-fibers-lean4@f75d4f2c2d6685a0547f17fa07240377aa2c2ed4. Read-only remote log downloaded at 13:58Z shows success through at least [8974/9924], no Lean compiler errors in currently visible log. Stage 3 full build and stage 4 scripts/verify.sh plus leanchecker --fresh Erdos821.Theorem are NOT YET TERMINAL. Active hourly automation OAI26 F011 Lean Closure (job Automation_018c525c56c48191b3d7ced4e2f95225) is bound to the same owned session with no duplicate claim.

    Disposition: F011_FORMAL_EXTERNAL_TO_GCL_STATEMENT_AND_FIBER_EQUIVALENCE_COMPLETE / EXTERNAL_UNCONDITIONAL_ANALYTIC_KERNEL_REPLAY_PENDING. Both protected PRs resolve formal definition/quantifier conversion, not the external author's analytic proof review or MATHCERT certification. Retain WP03, WP04 and programme parent open for terminal independent kernel replay, source-to-formal criticism and governed proof acceptance. No external mathematical reviewer or new MATHCERT authority is claimed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    Priority

    None yet

    Projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions