Skip to content

Erdős Proof Verification — end-to-end workstream

This is a separate substantive verification lane, not a rename of the compendium source-audit lane. It is controlled by ERDOS-PROOF-VERIFICATION-001 and the protected registry governance/erdos_proof_verification/registry.json. The workstream is currently pilot-registered, global triage incomplete. No proof has been verified by creating these manifests.

Task-specific executable entrypoints

Problem Source audit (completed administratively) Independent verification (open) Intended check
246 #1515 #3202 Birch source-level fidelity and theorem-to-formal bridge
247 #1516 #3203 Positive-index normalization and stronger growth-interface audit; open conjecture
248 #1517 #3204 External Lean kernel/axiom replay and Tao–Teräväinen transfer
249 #1518 #3205 Independent lacunary-subseries irrationality proof, not full open conjecture
250 #1519 #3206 Nesterenko bridge and external Lean kernel/axiom replay

All six issues are on the GCL Worker Queue, Project #2, Queue Class Mathematics. The five successors initially carry gcl-state:blocked: verify the registry and manifests on protected main before changing them to gcl-state:available. Their Project status is Todo, not Done. No protected merge, compiler output, mathematical claim, or certification is implied by source-adjudication closure.

Mandatory execution lifecycle

  1. Triage and deduplicate. Each completed catalogue source audit receives an explicit decision: SELECTED_WITH_MANIFEST, REUSED_EXISTING_VERIFICATION, DECLINED_WITH_REASON, or PENDING. Reconcile pre-existing Erdős 14 verification #2702 and Erdős 470/593 MATHSOLVE tasks before creating duplicates. The existing compendium batch snapshots are the ingestion population, not proof evidence. Record the actual distinction between proved literature, open conjecture, elementary lemma, and reported external formal proof.
  2. Lock sources and statements. Check exact primary paper theorem/page, source snapshot blob, all hypotheses/quantifiers, proof-provider commit, original Lean theorem type and any admitted axioms. Explicitly separate classic published proof from source-to-Lean interpretation. A live tag such as proved (Lean) is only a lead.
  3. Independently inspect mathematics. Reconstruct at least one bounded decisive proof step or demonstrate a precise equivalence/counterexample; inspect dimension, positivity, indexing, finiteness, quantification, and quantifier transfer. The same system may staff distinct role-scoped Adversary and Referee passes, but must not invent a new source or claim substantive independence from its own construction.
  4. Reproducible formal replay when applicable. Pin Lean, Mathlib, imports, solver source and command. Capture exact exit status, stdout/stderr hashes, full theorem type, #print axioms or equivalent, occurrence of sorry/admit and any imported axiom, and a hostile changed-goal test. CPU/Lean resources are suitable for early formal tasks; use integrated Colab when necessary. For empirical tests, prefer one GPU initially, log with W&B, and require multi-seed CLI reproduction. A compiler success is not proof of source-semantic equivalence or mathematical certification.
  5. Separate non-authoring critical review. Challenge the source, acceptance conditions and proof dependency boundary; distinguish independently reconstructed mathematics, borrowed literature claims, and formal kernel facts. A proof that an auxiliary lemma holds cannot close the original conjecture.
  6. Internal adjudication and protected documentary admission. Record reviewer and exact immutable evidence refs, proof gaps, negative tests, disposition and next action. Use an ordinary branch, protected PR, applicable checks, merge queue and protected readback. Keep evidence receipts and continuity updated. An evidence-backed exact blocker is a valid bounded result, but is not a positive verification result.
  7. Optional MATHCERT submission. Only separately authorize certification with independently sufficient evidence through MATHCERT's current policy. Protected merge, source audit, proof replay and reviewer passes confer no certification authority. Leave unsubmitted work marked NOT_SUBMITTED.

Validation and continuity

Run python3 ci/validate_erdos_proof_verification.py and python3 ci/test_erdos_proof_verification.py. The campaign policy shard includes both commands. The validator fails on duplicate issues, mutable pins, claim-scope drift, self-certifying statuses, and any green replay without a formal evidence receipt and semantic predecessor. Nine hostile fixture tests accompany the five manifests. The registry does not claim a fully reconciled catalogue population; global triage is explicitly pending.

The long-horizon multi-session current-state authority is governance/bounded_operation_checkpoints/ERDOS-PROOF-VERIFICATION-001.json. Follow its exact next_action, run its live freshness command before a material transition, and rebind changed PR/branch identities. Existing child source-audit issues remain closed as source work; proof-verification issues remain open until they reach a genuinely supported bounded proof disposition. Do not mark an open conjecture solved or close the parent on the basis of source-audit volume.

Exhaustive source-to-verification queue

The historical source registry has exactly 1,221 distinct Erdős problems, numbered 1–1,221. The first five explicit proof-verification pilots are 246–250; the remaining 1,216 now have separately identifiable work IDs in the protected-scope coverage registry at governance/erdos_proof_verification/coverage/REGISTRY.json. The original catalogue registry is a frozen source snapshot: 593 records were labeled open and 509 had formal-provider hints, neither of which establishes present theorem status or a completed Lean proof.

The execution granularity is one substantive problem per worker return, not one instant proof per 100-problem issue. Thirteen bounded cohorts each have four independently auditable phases: A source and scope reconciliation; B mathematical proof or bounded lemma verification; C formal applicability/kernel replay and nonauthoring adversarial review; D protected documentary evidence, public status and closure. Phase B/C/D for the five existing pilots refers to their existing individual issues; these are not duplicate proof requests. Every problem has its own four stage identifiers, source-audit issue, source-body SHA256, historical status, formal-provider hint, evidence fields and terminal guard. Initially all new queue issues are blocked until protected ledger admission; after that only Phase A and applicable queue-controls are globally available, while downstream stages are enabled per problem as evidence arrives.

Problem IDs A: sources B: mathematical check C: formal and adversarial D: protected documentary closure
1–100 #3212 #3225 #3238 #3251
101–200 #3213 #3226 #3239 #3252
201–300 #3214 #3227 #3240 #3253
301–400 #3215 #3228 #3241 #3254
401–500 #3216 #3229 #3242 #3255
501–600 #3217 #3230 #3243 #3256
601–700 #3218 #3231 #3244 #3257
701–800 #3219 #3232 #3245 #3258
801–900 #3220 #3233 #3246 #3259
901–1000 #3221 #3234 #3247 #3260
1001–1100 #3222 #3235 #3248 #3261
1101–1200 #3223 #3236 #3249 #3262
1201–1221 #3224 #3237 #3250 #3263

Cross-cutting acceptance work is separately tracked: closed-world coverage gate #3264, fair queue dispatch and atomic-claim continuity #3265, prior-proof deduplication #3266, claim-scoped MATHCERT handoff #3267, public evidence status #3268, and final 1,221-problem closure #3269. The existing global source-audit reconciliation #3208 stays active. Existing specialist work on 14, 470 and 593 must be critically reconciled, not blindly duplicated or treated as proof certification.

The queue is programmatically queryable from protected source using:

python3 ci/erdos_pv_queue.py summary
python3 ci/erdos_pv_queue.py next --stage A --limit 1
python3 ci/erdos_pv_queue.py next --stage A --after 100 --cohort 0101-0200
python3 ci/erdos_pv_queue.py get --problem 248 --stage B

The pickup command is intentionally read-only; a candidate is not a lease, independent review, or a claim of successful execution. Before doing work, acquire the current authorized atomic claim, check all issue comments and protected artifact state, and record the exact source and receipt. Validate the inventory and negative cases with python3 ci/validate_erdos_pv_full_coverage.py, python3 ci/test_erdos_pv_full_coverage.py, and python3 ci/test_erdos_pv_queue.py; these are added to the campaign policy shard. The complete inventory must be present for the gate to pass, while mathematical completion is separately fail-closed until individual evidence is actually admitted.

Protected inventory admission receipt

The exhaustive inventory was protected-admitted through PR #3271, merge commit 9b8067556d116ac8c09a7651e52d6f10c6952832. Readback confirmed the coverage registry and all 13 cohort packets (14 exact blobs), and the 58 Project #2 issues in the Mathematics class and Todo state. Of these, 17 are AVAILABLE: 13 source-scope cohorts (Phase A) plus coverage gate #3264, dispatcher #3265, prior-work reconciliation #3266 and public-status work #3268. The other 41 remain dependency-blocked: 39 cohort B/C/D phases, MATHCERT handoff #3267 and terminal closure #3269. The existing five explicit proof pilots are reused and remain separate.

These are issue and evidence-readback facts, not completed proofs. The exhaustive registry records 1,216 unsatisfied new per-problem triage decisions and no newly verified full mathematical claims. The next substantive proof obligation remains Erdős #248 verification #3204; source scope pickup can proceed concurrently via the governed read-only queue selector and an authorized atomic claim.