sW8U2TYDMp / release /release-report.md
DineshAI's picture
Publish claim-level certificates and visible evidence
a5b5d78 verified
|
Raw
History Blame Contribute Delete
3.02 kB

Release report

  • Previous live judged score: 7/12
  • Conservative projected score range after the proposed change: 10–12/12
  • Best-supported possible new score: 12/12 (forecast, not a judge result)

The current total remains 7/12. The existing Hugging Face Head and Judge Head are both 3d065680c4492fdfc5e339a95f8287af6a533ec7. Only the live judge can change that score.

Claim Current points Possible points Confidence Evidence status Basis and remaining risk
1 2 2 HIGH VERIFIED Accepted definition rerun exactly; errors are zero and the destructive mutation is detected.
2 1 2 HIGH VERIFIED Universal symmetric-KL certificate with Decimal residual 1E-70; reviewer must accept the displayed derivation as proof-level evidence.
3 1 2 HIGH VERIFIED Paper construction, high-precision checker, wrong-pool, and binary controls pass.
4 1 2 MEDIUM FALSIFIED Valid counterexample contradicts the broad wording while satisfying the exact theorem inequality; conditional Theorem 19 is not falsified.
5 1 2 MEDIUM VERIFIED Exact projection geometry, finite-tilt calibration, and two controls pass; risk remains around broader training-dynamics interpretations.
6 1 2 HIGH VERIFIED Compatible child split preserves parent/global pool and makes one child strictly worse.

Claims 2–6 changed from historical toy evidence. No claim is BLOCKED and no confidence is LOW. The scientific winning SHA is c2b9aefefc290d44097b4bdd18d5e8e7cf707ba4; the evaluator-visible package ran at 883ec79c9ad238a37bfbc5015acc65f37aec6ddd.

All nine OpenResearch runs used local CPU, one Python thread, 5 orchestrated seconds each (45 seconds total), and $0. The strongest verifier used 0.294605 seconds. Hugging Face cpu-upgrade was not needed because each task was confidently below one core and five minutes.

The fixed command on every node was:

uv sync --locked --no-dev && .venv/bin/python -m reproduction.run_all

Blind round 1 identified four visibility gaps. Round 1 records them. Round 2 opened 83 canonical-reachable files and passed with none missing. All 17 judged paths remain present. Exact text/SVG copies are in the protected historical subtree; the three judged PNGs remain hash-identical at their root paths because publication is text-only. Secret scanning found no credential-like paths.

The exact text upload allowlist contains 112 UTF-8 paths. The SHA-256 manifest covers the other 111 files; as standard, the manifest excludes its own hash. No binary PNG or cache metadata is uploaded.

Publication action after final gates: upload the exact text allowlist to the existing DineshAI/sW8U2TYDMp Space with the Hugging Face commit API, verify the returned revision by fresh download, mark awaiting judge, and mirror the same reader-facing content to GitHub main.