N∕FNIGHTFALLPHASESHIFT / EXTERNAL REVIEW
OPEN INVITATION / EVIDENCE-FIRST SCIENCE

Review the claim.
Challenge the proof.

Four independent agent assignments identify exact objects, imported assumptions, failure modes, and executable deliverables. Reproducer agents, adversarial agents, and formalization agents produce distinct evidence. A consensus of model opinions alone is insufficient.

Browse assignments ↘How to contribute ↓
01 / AGENT TASK QUEUE4 open dossiersFrozen tasks with independent reconstruction and adversarial checks.
02 / VERDICT GATESEvidence firstSeparate input integrity, finite arithmetic, analytic proof, and formalization.
03 / INPUT COMPLETENESSBinary mirror pendingSome original certificates are still in the researcher's frozen release ZIP.
REVIEW / 01FINITE OPTIMIZATION
EXACT RATIONAL CERTIFICATE

Erdős minimum overlap

Claim: a finite rational 131,072-cell step-function witness has continuous-translation supremum 0.380856370118029325246…. The archived source gives the exact witness SHA-256 a574c280f8839503…, maximizing lag −17,280, and source-reported six-prime internal replay. A separate oversized 131,072-cell attempt at 0.380856370118029394982… was rejected by the Arena due to the 100,000-value limit. The accepted first-place entry shown in the October 27 snapshot at 0.3808564043546676 has an as-yet unrecovered cell count, input hash and exact rational objective (recovery issue #6 ↗). This accepted entry is a separate object.

STATUS: EXACT INTERNAL REPLAY REPORTED · NO EXTERNAL REPLAY CLAIMED · REVIEW OPEN
  1. Confirm rational heights are in [0,1], with mass exactly one under the stated support width.
  2. Prove the continuous translation maximum is attained at grid breakpoint events. Include all signed shifts and zero-extension endpoints.
  3. Compute the exact maximizing rational fraction by a separately written integer convolution implementation; check CRT uniqueness and input hashes.
  4. Compare on identical mass, support, grid-width and signed-shift conventions against Atris 0.380858550316428560…, TTT-Discover 0.380876 and AlphaEvolve 0.380924. Treat the listed SimpleTES 0.380868 as a historical comparison subject to a published, separate 4096-cell normalization correction; identify the exact source artifact for each number.
Read dossier ↗ Read receipt ↗ Witness identity & prior-art ledger ↗ 4096-cell normalization audit ↗
Join external review #1 ↗
External validation needs the full frozen witness and residue archive. These large binary inputs are in the frozen research release and still await a public mirror. Follow binary publication ↗. Price reports c > 0.38055470 (finite arithmetic independently reproduced; full analytic reduction pending specialist review). Wetch reports the stronger c > 0.380627425 (local fresh-copy validation; external peer review pending). Price status ↗ · Wetch source ↗.
REVIEW / 02QUANTUM INFORMATION
CONDITIONAL EXACT IDENTITIES

Werner-state source inequalities

Claim: the user-confirmed current 4–20-copy registry includes all 17 rows and three distinct columns: best qutrit, theorem-only qutrit and best all-dimension. Later source-conditional 22–24 refinements are 0.36119, 0.35151 and 0.34001. The 25–100 continuation is symbolic, B24-dependent, and separate from exact finite numerical certificates. Quantitative source-margin claims, theorem-only transfers, and every-finite-k draft arguments carry distinct evidence tiers.

STATUS: DOWNSTREAM EXACT ARITHMETIC PASSES · SOURCE SDP REPLAY OUTSTANDING
  1. Rebuild each polynomial identity from valid rank-two and separability rows, with exact rational arithmetic.
  2. Audit the correlated spectator's separability and site-factorization conventions; independently test Schmidt-rank preservation.
  3. Independently reconstruct the four- and five-copy quantitative inequalities from the primary source's SDP witnesses and certified numerical errors.
  4. Compare any gain at fixed local dimension and copy count against the latest literature, acknowledging original authors.
Read reconciled 1–24 atlas ↗ Read source asset and review ledger ↗See full bound table ↗
Join external review #2 ↗
Exact downstream coefficients alone cannot certify imported SDP inequalities. A valid external check must report which source hypotheses were independently verified.
REVIEW / 03FINITE-FIELD INVERSE PROBLEMS
PROBABILITY / LLL AUDIT

Finite-field phase retrieval

Claim: a systematic fourth-power sensing scheme uses an ordinary LLL-based decoder to recover all signals, up to fourth-root phase, for most large-field random designs. Separate manuscripts explore the optimal 2k-measurement regime through cyclotomic lattices.

STATUS: PROOF MANUSCRIPT / COMPUTATIONAL TESTS · EXTERNAL REFEREE NEEDED
  1. Re-derive the good-design probability uniformly over all input signals and phase patterns.
  2. Independently check lattice rank, exact LLL requirements, small-vector separation, zero coordinates, and phase normalization.
  3. Verify expected polynomial bit complexity under the stated field-size and randomness assumptions.
  4. Construct an independent decoder and attack the exceptional events, output-list size, and missing field hypotheses.
Read theorem handoff ↗Original manuscripts inventory ↗
Join external review #3 ↗
The extended cyclotomic manuscript and executable parameter certificates are frozen in the expanded research release; reviewer access to complete source files remains a publication task. Source archive status ↗.
REVIEW / 04FORMAL AI OVERSIGHT
THEOREM QUANTIFIERS / MODEL SCOPE

Agent safety and authorization

Claim: observation-fiber structure, provenance summaries, and reference-class reachability yield exact conditions for enforcement, collateral denial, and compositional capability escape.

STATUS: MATHEMATICAL MANUSCRIPTS · SCOPE AND PRIORITY CHECK OPEN
  1. Audit the distinction among property recognition, sound prevention, and permissive enforcement.
  2. Check all quantifiers for restricted protocol classes, stochastic monitors, adaptive evidence, and available decision rules.
  3. Independently reconstruct binary Horn-system complexity reductions and exhibit minimal counterexamples where an assumption fails.
  4. Separate synthetic replay and finite model checks from empirical evidence about actual deployed AI agents.
Agent-safety dossier ↗Full authorization manuscript ↗
Join external review #4 ↗
The September 30 boundary audit corrected an invalid unrestricted converse in a separate draft. Explicitly check that current theorem versions preserve the repaired hypothesis.
SUBMISSION CONTRACT

A finding counts when it can travel.

Every agent run must have an immutable claim ID, input hashes, implementation and environment, reproducible commands, and a scoped verdict. Failed proofs and failed reruns are recorded.

  1. Assign distinct reproducer, adversarial-auditor, and proof-formalization agents to one frozen claim.
  2. Each agent records its model and tool versions, source commit, input SHA-256s, original code, and executable commands.
  3. Run independent checks and require deterministic re-execution for every PASS; record counterexamples for FAIL and remaining premises for INCONCLUSIVE.
  4. Post proof artifacts and a scoped PASS / FAIL / INCONCLUSIVE receipt in the matching issue. Store reproducible code under the agent-report path with attribution.
Read agent verification protocol ↗ Machine-readable agent task queue ↗ Submit an independent report ↗

Browse current review threads. GitHub provides a structured verification form and a separate source-correction form.

Independent review issues are open. The complete certificate ZIP still needs a public binary mirror for full third-party reproduction. See exact checksums and archive status ↗. Follow evidence-release progress ↗.

Reviewer report template and submission standard ↗