菜单
Use when a manuscript or draft needs an independent referee-style assessment before submission, revision, or handoff. Generates a clean-room, offline (no network) referee-review report with Markdown + strict JSON output (generic profile).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Correctness and legibility checklist for data/results figures in any field: charts, parameter scans, spectra, distributions, fit comparisons, constraint contours, heatmaps, and low-dimensional projection scatters. Use when an agent plots computed or measured data, revises a results figure for a manuscript, or audits whether a data figure is publication-ready. Covers data fidelity (excluded rows never enter plotted summaries, connected-series evaluator fingerprints are homogeneous, a rendered continuous field is continuity-scanned before delivery, claim-titles true against every axis category, one canonical value per claim), label economy floor and ceiling, colour threading with a CVD-safe palette, role-mapped typography, chart choice by data shape, a render-then-verify QA loop (bbox overlap check plus per-panel perceptual crops), a figure-reproduction provenance bundle, and a fail-closed display-acceptance gate that refuses durable or outward-facing use until every plotted quantity is bound to a verification-
Use when writing Julia numerical or scientific code, to apply always-on performance guardrails, gate an accelerated implementation as an identity against its verbatim reference over the full domain of intended use, and escalate to reproducible benchmark gating for speedup or regression claims. Works standalone by default and can optionally emit ecosystem artifacts.
Convergence and reliability gate for NUMERICAL results in any field, including fits, optimizations, integrals, eigenvalues, roots, poles, zeros, ODE/PDE solutions, Monte-Carlo estimates, and downstream feature extraction. Use before trusting, comparing, publishing, or folding a computed number into durable research artifacts. Requires resolution convergence, independent-method checks, regression anchors, post-correction downstream recomputation, method-precondition checks, configuration-threading audits, gate-discrimination (negative-control) audits of purpose-built validation chains, accelerated/heuristic fast-path scoping (no false guarantee, unconditional escape hatch, production-validated precondition), and honest uncertainty/reporting. Emits an auditable reliability matrix. Sibling to derivation-verify for symbolic claims and julia-perf for speed claims.
AI failure-mode checklist (M1-M8) for research agents. Generic across domains. Walk M1-M7 at the moments work becomes durable — you just finished a derivation or computation and are about to write conclusions, fold a number or claim into the plan or contract, check off a task-board item, make a milestone closeout commit, hand off, assemble submission material, or claim a result is final — and, in projects using the engine's approval flow, before requesting an A1-A5 approval gate. Walk M8 before starting each heavy or production computation or hand-building a reusable computational method.
Use when a research milestone (theory or computation) needs executed, reproducible work whose results are independently reviewed before being trusted. Milestone-based research-team workflow for theory+computation projects with reproducible artifacts, independent parallel workstreams (default: host-native subagents; configurable), and a strict convergence gate.
Use when a result, manuscript, derivation, or diff needs independent review. Runs clean-room multi-agent review loops across Claude/Gemini/Codex/OpenCode/Kimi with strict review-contract checks, fallback policy, and convergence gates.
| name | referee-review |
| description | Use when a manuscript or draft needs an independent referee-style assessment before submission, revision, or handoff. Generates a clean-room, offline (no network) referee-review report with Markdown + strict JSON output (generic profile). |
This skill defines a generic (non-venue-specific) peer-review output contract that is designed to be ingested by an orchestrator as artifacts.
review.md: first line must be exactly VERDICT: READY|NOT_READYreview.json: must validate against schemas/review.schema.json (fail-fast)The author agent should walk the M1-M7 pre-approval ritual from
skills/research-integrity/ before posting a draft to this reviewer.
Findings the author could have caught (hallucinated citation, hallucinated
measurement, methodology-not-in-artifacts, frame-lock, etc.) should not be
the reviewer's BLOCKING items; this reviewer is for adjudication against
the packet, not for catching omissions the author skipped.
Recording the pre-flight. The M1-M7 walk is always required before posting a draft to this reviewer. How the walk is recorded depends on whether an approval gate is open:
nullius integrity-record --approval-id <id> --modes <Mx,...> --notes "..."
against that gate's approval_id. The receipt is the canonical
machine record of the pre-flight and is fail-closed-enforced by
nullius approve (missing receipt → INTEGRITY_RECEIPT_REQUIRED,
see packages/shared/src/integrity-receipt.ts).
This matches the gate-coupled "machine record" path documented in
skills/research-integrity/SKILL.md
§"Recording the check".skills/research-integrity/ §"Recording the check" is the audit
trail here. Do not invent a synthetic approval_id to feed the
CLI: nullius integrity-record is parameterized on
approval_id precisely because the receipt is the
gate-coupling key, and a junk identifier pollutes
.nullius/integrity_log.jsonl without producing any
enforcement (this reviewer's pipeline does not consume the log —
scripts/run_referee_review.py does not read it).This reviewer remains generic and non-venue-specific; the pre-flight is the author-side discipline, not a venue requirement.
evidence_requests with structured queries (to be executed by an external stage), and keep verdict=NOT_READY when blocking.python3 scripts/run_referee_review.py --profile generic --packet fixtures/packet.md --out-dir /tmp/review_out
This writes:
/tmp/review_out/review.md/tmp/review_out/review.jsonThe packet is a Markdown file. Recommended sections:
## Paper diff## References (short bibliography or notes)## Artifact pointers (one per line, e.g. - hep://...)