AGENT SKILL

referee-report

Adversarial line-by-line referee-grade review of OUR OWN manuscript or notes, at prover level, delivered to ai/ — never edits the tex. A green build resolves nothing mathematical (mechanism 5). Use when the user says "referee my section N", "referee-report on the draft", "adversarial review of the manuscript". NOT for handling a RECEIVED referee report (triage-referee-report / manual frontier-fable response per ADR 0005) and NOT for applying fixes (revise, after human adjudication).

License 许可
GPL-3.0
最近更新
2026-11-13
Stars
★ 0
Forks
0

01 前言 Front Matter

name
referee-report
description
Adversarial line-by-line referee-grade review of OUR OWN manuscript or notes, at prover level, delivered to ai/ — never edits the tex. A green build resolves nothing mathematical (mechanism 5). Use when the user says "referee my section N", "referee-report on the draft", "adversarial review of the manuscript". NOT for handling a RECEIVED referee report (triage-referee-report / manual frontier-fable response per ADR 0005) and NOT for applying fixes (revise, after human adjudication).

03 SKILL.md 正文

referee-report

Read-only toward the tex. This leaf's procedure:

Produce the report

  1. Resolve the repo docs (subdomains/repo-docs/RESOLUTION.md); scope = the named section(s) of the canonical file (WRITERS.md's Target clause defines it).
  2. Dispatch the review at prover level per subdomains/lean/PROVERS.md, with the mandate: line-by-line; verify every proof step; attack every statement (counterexample hunting per mechanism 6); check every citation's plausibility (route suspicious ones to track-down-reference); hypothesis/notation drift; the mechanism-5 rule verbatim — compile success is not mathematical correctness.
  3. Report to ai/<date>-referee-<slug>/report.md: per-finding severity (BLOCKING / MAJOR / MINOR), location, the objection, and what evidence would resolve it. Every finding also has a stable ID, an explicit disposition, an exact source/report locator, a closure test, and a destination: RESOLVED-IN-REPORT, OPEN-TODO, IMPORTED-OBLIGATION, OUT-OF-SCOPE, or INVALIDATED. Claims in the report follow the five-way taxonomy; contradictions found route through contradiction-check (its report + refusal semantics).
  4. Run a report-completeness pass before issuing a verdict. The report must contain a coverage table listing every material objection found during the review. Every row must use exactly one destination and the corresponding evidence fields: RESOLVED-IN-REPORT gives the exact resolving passage and closure test; OPEN-TODO gives a named owner, evidence requirement, next source or computation, and closure test; IMPORTED-OBLIGATION gives the precise source, hypotheses, boundary of the import, and an explicit statement that the fixed criterion permits it; OUT-OF-SCOPE gives the criterion and the reason for exclusion; INVALIDATED gives the evidence that defeats the objection. An objection may not remain only in worker scratch, an ancillary calculation, or an unlinked note. ACCEPT is forbidden while any material finding is OPEN-TODO, or while an IMPORTED-OBLIGATION row lacks its source, hypotheses, boundary, or criterion permission. A material row may support ACCEPT only after its stated closure test passes or its permitted import/out-of-scope status is explicit.
  5. Apply the example admissibility gate. An example used as evidence for correctness, a gap, or a verdict must either be a source-native object with its exact hypotheses and source interface, or an interpretation-compatible object with an explicit typed interpretation map preserving the operations under discussion. Do not use the labels "toy model", "toy example", or "naive calculation" as mathematical evidence. If an illustrative construction has no such interpretation, omit it from the proof argument; at most record it as a rejected diagnostic in the coverage table, explicitly excluded from the verdict.
  6. Return for adjudication under the caller's authorized route.

Red flags

Thought Reality
"It compiles and reads well" thm1p10: green builds, 5 blocking findings. Attack the math.
"Fixing it while reviewing saves a round" Reviewer edits nothing. Adjudication first.

Negative results travel with the positive ones (LX11, LX12)

This skill renders previous work into a new artifact, so the negative-result floor of ../../POLICY.md applies. Before finishing, enumerate the inputs' refutations, counterexamples, ill-posed proposals, and failed expectations, and carry each into the output as a numbered statement (or a labelled question when unproved), naming the expectation it corrects: "it is natural to expect $X$; in fact $Y$". Report the count carried and, for anything deliberately left out, the scope reason. Dropping a refuted claim together with its refutation is a revise-level violation: the refutation is the surviving result.

04 同仓库更多 Skills