cross-model-review-gate
Two-model review before an irreversible step. A reviewer from a second lab checks the work, and every review is recorded against the exact commit and artifact hash. Learned from gnuchev's Open Research Club (a Claude author, an OpenAI reviewer, about three rounds per artifact). Use before submitting a structural benchmark row, publishing a tool or skill, declaring a route dead, or deploying.
- License 许可
- MIT
- 最近更新
- 2026-10-30
- Stars
- ★ 4
- Forks
- 3
01 前言 Front Matter
- name
- cross-model-review-gate
- description
- Two-model review before an irreversible step. A reviewer from a second lab checks the work, and every review is recorded against the exact commit and artifact hash. Learned from gnuchev's Open Research Club (a Claude author, an OpenAI reviewer, about three rounds per artifact). Use before submitting a structural benchmark row, publishing a tool or skill, declaring a route dead, or deploying.
02 资源文件 1
- 参考文档agents/openai.yaml176 B
03 SKILL.md 正文
Cross-model review gate
A receipt is one reviewer's record of exactly what it checked about one exact revision. It reports checks; it does
not certify the claim. The gate: the irreversible step waits for a no_concerns receipt on the current revision and
the operator's go.
Loop
- Author commits the artifact. Record the commit id and every artifact hash the claim rests on (e.g. ops.bin sha256, note file hash).
- Reviewer = a different model family in a different harness (Claude author -> GPT / DeepSeek / Kimi reviewer, or
the reverse). Brief it adversarially ("refute this or name what was missed") with paths, not conclusions. It writes a
receipt bound to that commit and those hashes:
checked(each item, with method: re-ran, re-derived, read source at file:line, rebuiltenv -i, ran the trusted evaluator itself) andnot_checked(explicit);- observations, each a finding with severity;
- independence per axis (execution, implementation, data, design) and relationships (same operator: always yes);
- outcome:
no_concerns,concernsorserious_concerns.
- Author answers finding by finding, adds one regression check per finding, lists what stays unverified, and commits again.
- Reviewer rechecks at the new commit (only the delta plus the regression checks). Loop until
no_concerns; about three rounds is normal. - Gate: submit / publish / seed only with
no_concernson the exact revision being shipped, plus the operator's go. A changed byte means a new revision and a new receipt.
Benchmark variant (ecdsa.fail)
The strongest receipt is an independent rebuild: the reviewer rebuilds env -i from the committed defaults, checks the
ops sha256, runs the evaluator itself, and runs a fixed-input A/B against the parent. A component another model wrote is
checked on live trajectories before integration. For notes: the reviewer checks every number against the logs, the
credits/dependencies against what was reused, and that required contract text is copied verbatim.
Wiring
Run the reviewer read-only at its highest effort: a second CLI agent from another lab, or a Claude Fable subagent
(model: "fable"; confirm the transcript reports the model you asked for, since a silent fallback voids the review).
Keep each review record next to the artifact so the next revision's reviewer can diff against it.
04 同仓库更多 Skills
verify-research-transfer
Check whether a literature finding supports a proposed performance optimization. Read the primary figures and the implementation source, then run a cheap falsifier before writing code or benchmarks.
verify-paired-native-measurements
Audit the process, sample and archive evidence behind paired native benchmarks, especially wrappers around an unchanged official client. Use before admitting benchmark results or when reviewing the controller that produced them. Ordinary unit tests do not need this.
squares-in-squares-campaign
Pack n = 1..1000 unit squares into the smallest square (Kaggle "Squares in Squares", score sum L/sqrt(n) - 1). Use when improving, certifying or submitting packings, or when choosing the next search route. Covers the official-metric gates, the solver and MIS tools, the structure of the frontier, and the routes measured dead. The competition closed on 25 Oct 2026.
shot-autopsy
When a candidate fails shots its parent passes, find the structural cause before any nonce search. Reconstruct the exact failing shot, localize the first divergence, repair it by paying precision, and keep the case as a permanent regression. Learned from bulengerk's ecdsa.fail rows. Use when an edited candidate fails shots its parent passed, or when counts drift between builds.
send-to-claude-fable
Send a bounded task or review packet to the local Claude Code CLI on the Fable model and bring back its verdict. Use when you want an independent second opinion from a different model.
rust-archive-build-binding
Prevent stale Cargo artifacts when a source archive is restored over a reused Rust target directory. Use for reproducible benchmark comparisons and for moving source archives between machines.