METR count_odds reviewer packet
A public-safe packet for one Flywheel task-image interoperability check, not a product release.
Download
This archive preserves the source pins, runner receipts, and verification notes needed to inspect Flywheel PR234's count_odds bridge path. This is interop evidence, not METR endorsement, a model benchmark, a safety certification, or a Flywheel 1.0 release claim.
- metr-count-odds-reviewer-packet-20260913-public-v2.zip 178,203 bytes
- metr-count-odds-reviewer-packet-20260913-public-v2.zip.sha256
SHA-256: 1a1a9ca19728f11935d93eb8ef14770fa4d78f82484df356ac22b4da929dcac0. The hash checks archive identity; it does not by itself prove semantic correctness.
What the packet covers
Flywheel PR234 added a minimal METR Task Standard import path and was merged on September 13, 2026. The retained run built the upstream count_odds task image in a runner-local registry, exercised three controls through a pinned Inspect bridge, and imported the run exports through a pinned Flywheel revision.
- Correct submit: answer
6, score1. - Wrong submit: answer
1, score0. - No submit tool call: score
0, with the packet noting that upstream scorer fallback limits what this proves.
Pinned sources
| Item | Pin or URL |
|---|---|
| Flywheel PR | HarperZ9/flywheel#234 |
| PR merge commit | d28200a78c16cf2a970104d8a5f057a526b8cfa0 |
| GitHub Actions run | 34750411689, conclusion success |
| Inspect bridge source | METR/inspect-metr-task-bridge@20a430cc59283c8b463e526e77f8d040509457ec |
| Task Standard source | METR/task-standard@03236e9a1a0d3c9f9d63f6c9e60a9278a59d22ff |
| Flywheel importer source used by the run | HarperZ9/flywheel@b3ec822b4c1a942697f1370383c83beac339155e |
The public packet omits copied bridge source and built task images because the pinned bridge snapshot did not include a detected license file.
Verify locally
Extract the archive and use its bundled verifier against the exact Flywheel source checkout named in the packet README:
python .\tools\verify_packet.py --flywheel-source <flywheel-checkout> --out .\_verify\packet-verificationThe verifier checks the retained manifests, hashes, and referenced local packet structure. It does not rerun METR infrastructure or establish general conformance.
To recheck the same claim from a clean Flywheel clone instead, run python -m harness.recheck run recheck/metr-count-odds-packet.json. The runner fetches the packet, refuses it unless its sha256 matches, reruns the check and then a control that must fail. RECHECK.md describes the runner.
Scope boundary
The packet carries reported scorer outputs, source pins, and byte-level receipts for one public bridge exercise. It does not certify METR conformance, scorer semantic correctness, Flywheel installed release status, independent adoption, model safety, or Flywheel 1.0 readiness.