Paper review
Strict, source-traced scientific paper review. Runs all A-R dimensions under the SCIPAPER_STANDARD feedback contract: mathematics, physics, logic and statistics, language and de-AI, document structure and narrative spine (contribution graph plus a seven-question cold read), citation existence and relevance, data and figures, interfaces, redundancy, reproducibility, modern-physics checks, cross-section consistency, adversarial derivation verification (three passes plus twelve-framing escalation to CONFIRMED / REFUTED / MARGINAL), staleness and drift, process artifacts, internal draft language, citation precision, and glossary alignment. Scientific-integrity blockers must be resolved, L0 targets must reach zero, and every strong advisory needs an explicit disposition; ordinary editorial advisories are never disguised as a universal paper PASS/FAIL. Use for: pre-submission audit, full manuscript review, source tracing, mathematics and physics re-derivation, 投稿前 audit, 完整论文审查, 严格迭代修订.From its SKILL.md
npx -y skills add skymanbp/sci-paper --skill paper-reviewAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 6 stars6 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
24.0 KB, as published. Nobody here has run it
paper-review — typed, source-traced scientific review
Normative authority:
docs/SCIPAPER_STANDARD.md./sci-paper:papersupplies writing guidance; field dossiers, baselines and learned models supply evidence. None defines an independent paper verdict. Funding proposals are not papers: review them under the/sci-paper:proposal-polishregister (standard §7) — the L0 policy and §6 invariants carry over, the significance-trimming rules of paper mode do not.
本 skill 不是快速 lint。它完整读取论文与源材料,逐项验证科学声明,并把结果写成
sci-paper.feedback.v1 finding。审查循环是:
measure → type → rank → edit → re-measure → disposition
终点是 disposition-complete review state,不是“0 个审美意见”或通用 PASS。
0. Non-negotiable rules
- 当轮重新取证。 数字、公式、引用、图表、时间和版本必须在本轮从源文件、 DOI/arXiv、数据或脚本输出重新读取。记忆和旧报告只能帮助定位,不能证明事实。
- grep 只定位。 每个命中或未命中都要通过上下文 Read、替代拼写和必要的全文 检查确认;不能把 grep 当作语义证据。
- 禁止猜测。 无法验证时写明
unmeasured、NEEDS SOURCE或NEEDS PHYSICAL ACCESS,不能制造近似答案。 - 每个数值都有 provenance。 合法来源是:具体数据/脚本输出、具体 DOI/arXiv
的表/图/公式,或由本文已编号公式直接算得。缺失来源是
integrity_blocker。 - 图表看实际内容。 每个 figure 必须查看实际 PDF/PNG;每个 table 必须与上游
数据逐 cell 对位。caption、正文和制品矛盾是
integrity_blocker。 - 科学优先。 数学、物理、统计、数据或引用错误不能靠 softening、disclaimer、 “schematic” 标签或风格改写掩盖。修复根因。
- 推导三重独立验证。 每个 displayed derivation、关键物理链和数值结论都执行 M.pass-1/2/3;任一命中后修复并重跑全部三 pass。
- 后果类别不能混用。
integrity_blocker: 科学、来源、引用、构建或必需制品错误;不可由风格偏好豁免。l0_target: Tier A、em-dash、同 section 同 Tier B 词的第 2 次及以后。advisory: editorial、narrative、结构、corpus distance、learned field-similarity。
- 缺失测量不等于 0。 所有 axes 显式标
measured、degraded、unmeasured或not_applicable。 - 用户保留主观裁决权。 strong advisory 可以
acted、accepted、rejected_as_false_positive或带理由pending。普通 advisory 可以保留,但必须报告。
1. Preparation
每轮均重新执行:
- Read
docs/SCIPAPER_STANDARD.md、/sci-paper:paper和当前 SKILL.md。 - 解析 field:单 profile 可自动选;多个 profile 必须
--field;无 profile 时继续 model-free 审查并标相关 axesunmeasured。 - 从头到尾 Read 论文;分块也必须覆盖全文。
- Read bibliography、项目事实源、数据/结果索引、生成脚本和相关 companion 当前版本。
- 建立五张清单:
- claims:claim、scope、stance、evidence relation、位置;
- numbers:值、单位、位置、source trace、核验状态;
- equations:定义、量纲、假设、边界、实现位置;
- figures/tables:制品、caption claims、生成脚本、上游数据;
- interfaces:论文符号 ↔ code variable ↔ data column。
- 对 LaTeX 论文执行项目规定的完整 build;记录 errors、undefined refs、duplicate
labels、missing assets 和 warnings。项目若使用其他 build system,执行其权威命令。
6b. 长度基线快照(标准 §5.3):首次 edit 前把目标文件复制到
<scratch>/length-baseline.tex(或记下干净的 git ref)。没有诚实的 基线,收尾的 length gate 无从执行。 - 运行共享 de-AI report:
python tools/ai_ism_lint.py <file> --field <field> \
--structure --distribution --document-structure --oracle --voice \
--format json --output <scratch>/paper-review-feedback.json
解释 exit status:0 = 无 L0 target,advisories 可存在;1 = 有 L0 target;2 = 输入、 配置或执行失败。不得从打印文本反解析 JSON。
同轮补充 claim-anchoring 质量带(写作质量轴,非 AI 判别轴——EVALUATION.md §9.6 记录了被证伪的假设;发现只作为"主张缺锚"质量问题处理):
python tools/deai_anchoring.py <file> --field <field>
Cooperative-layer tools (opt-in; each is honestly unmeasured without the
author's own material, never an AI verdict). When the author supplies an AI-draft
ancestor, report which spans are still the AI draft; when they point at their own
prior papers, report whether this draft is under-varied against their own baseline
(a confound-free same-author reference — EVALUATION.md §9.9):
python tools/deai_provenance.py <file> --ai-ancestor <earlier-draft> # or --git-ai-ref <commit>
python tools/deai_personal.py <file> --prior-papers-dir <author-own-papers>
2. A–R review dimensions
每个 finding 必须给出:kind、layer、rule、scope/location、当前证据、source
trace、measurement status、priority、recommended action、disposition。确认某条 critique
存在,不自动决定它是 blocker;后果类别单独判断。
A. Mathematics
- 逐式检查量纲、代数、定义、符号一致性、近似条件、边界和极限。
- 手算或用独立脚本复现关键化简与数值例;脚本结果必须保留 provenance。
- 所有
\eqref{}、labels 和符号定义闭合。 - 数学错误、未声明的必要假设或结论无法推出均为
integrity_blocker。
B. Physics
- 核心物理量、算子、滤波器、信号/噪声模式和 sign convention 与代码一致。
- 守恒律、对称性、宇称、渐近行为和适用域成立。
- 区分相关与因果;因果声明要有机制或设计支持。
- 物理错误、适用域矛盾或 claim/evidence 冲突为
integrity_blocker。
C. Logic and statistics
- claim graph 无循环论证、断链、偷换条件、充分/必要条件错误或未声明假设。
- 样本、split、CV/grouping、泄漏防护、metric、uncertainty、多重比较和 prior 可复现。
- 无效统计、泄漏、错误外推或 unsupported conclusion 为
integrity_blocker。 - Claim–evidence discipline(
/sci-paper:paper"声明-证据纪律" 的审查端): 每个经验声明有正文内数字/图/表/引用支撑;动词强度不超过证据强度;模糊量级 写成有归属的数字或区间;significantly无伴随检验或数字时是 claim 缺陷。 无支撑或超强度的声明按 claim-evidence defect 处理(integrity_blocker); 措辞层面的弱化建议为 advisory。 - 合法但表达不清的逻辑连接通常是 advisory。
D. Language and de-AI
- 术语、缩写、时态、句法、信息密度和段间连接支持准确阅读。
- Tier A、em-dash、Tier B cap 由共享 linter产生
l0_target。 - announced enumeration、parallel-modal runs、setup/list/wrap-up symmetry、重复段落/章节
几何、burstiness、UID、document shape、learned field-similarity、-ing 分词尾巴与
阐释式冒号(linter
ing-tail/colon-elaboration)是 advisory。 - learned signals 只表示 field-similarity/triage,不证明作者身份;没有 calibrated
operating point 时必须
degraded。 - 结构 tell 审计(每轮必跑;
/sci-paper:de-ai--audit-only模式,其 Pass 2): 对全文执行 humanizer Layers 1--2 结构清单——2.11 长句堆叠(约 30 词以上、 3 层以上从句链或双层括号嵌套)、negative parallelism("X, not Y" 模板密度)、 elegant variation(同一对象换名)、rule-of-three padding、公式化开场白、 connective 连排。结构命中为 advisory(同类超密度升 strong advisory);该 skill 的 Layer-4 claim--evidence 命中并入上方 §C 纪律(integrity_blocker)。其 Corpus overrides 优先于上游词表:landscape / demonstrate / significantly 不做词表级标记,只走证据条件。 - 修复 de-AI finding 时遵守
/sci-paper:paper的 Preserve List:不得把证据绑定的 hedging 改强,不得为避词改动数字、引用或 stance——制造 over-claim 的"修复" 本身构成 claim-evidence defect。 - 不能要求所有结构 advisory 归零。strong advisory 要 disposition,普通 residual 要报告。
E. Document structure and narrative spine
- 检查 contribution graph、依赖顺序、section function、abstract coverage、intro promise、 results/discussion/conclusion closure 和图表首次引用位置。
- 叙事主线协议(cold-read):先建 paper-level purpose record——root question、
contributions[]、每个贡献的 method_or_argument、key_evidence[]、take-home + scope,
全部带 file:line;再建 contribution graph(nodes = claims/definitions/methods/
evidence/conclusions;有向边 = prerequisite/supports/qualifies/contrasts/
derives-from)。
然后按 cold reader 回答七问:root question 是什么?贡献有哪些?贡献之间关系?
各自的方法/论证?各自的关键证据?take-home 与 scope?读者在哪里需要回溯、
补隐藏上下文或在竞争解读间抉择?每个困惑按后果分类:矛盾或缺必要支撑 =
integrity_blocker;高曝光可复现困惑 + 具体修复 = strong advisory;局部措辞 = ordinary advisory。 - motivation → method → validation 是常见默认,不是所有论文必须套用的三幕模板。
- 多贡献论文可以有多个明确相关分支;问题是关系是否解释清楚,而非图是否恰好只有 一个 component。disconnected node 是待查证据,不是自动 fatal;用连接词桥接 逻辑断链(Furthermore/Moreover 替代缺失前提)是禁手。
- 缺失支撑关键 claim 的必要论证可升为
integrity_blocker;其余 narrative findings 按证据强度分为 strong 或 ordinary advisory。
F. Citation existence and relevance
- 每个 bibliography item 的作者、年份、标题和唯一标识从 DOI/arXiv/出版页核验。
- 每个 citation 在当前句中确实支持所述 claim;查不到则标明访问缺口。
- fabricated、misused 或 claim-critical unverified citation 是
integrity_blocker。 - 边缘但可能更好的引用通常是 advisory,不自动阻塞。
G. Data, results, figures and tables
- 每个数字定位到 source;当轮重算能重算的结果。
- abstract、正文、table、caption、figure、conclusion 同名数字与单位一致。
- table 逐 cell 对位;figure 检查曲线、点、误差、轴、单位、legend 和 caption。
- baseline 比较使用相同 protocol;uncertainty 的定义与样本量明确。
- mismatch、stale result、missing required artifact 或不可复现 claim 为
integrity_blocker。
H. Interfaces
- 论文公式 ↔ 实现函数、符号 ↔ code variable ↔ data column、pipeline 叙述 ↔ 执行顺序 逐项映射。
- 因子、单位、cut、mask、preprocessing、randomization 和 aggregation 不得漂移。
- 接口矛盾为
integrity_blocker。
I. Redundancy
- 找重复 claim、无信息增量段、死定义、死 figure/table 和重复解释。
- 若冗余造成两个互相冲突的版本,则为
integrity_blocker;纯压缩与可读性通常 advisory。 - 本维度只产 finding;冗余修复的动作路由到
/sci-paper:condense执行 (one-canonical-home 去重 + length gate 收口)。
J. Reproducibility
- 数据来源、selection、preprocessing、software、hardware、seed、hyperparameters、split、 metric 和 availability 足以复现声明。
- 缺失使核心结果不可复现的必要信息是
integrity_blocker;非核心补充信息可为 advisory。
K. Modern-physics checks
对适用内容执行:
- dimensional analysis;
- asymptotic limits;
- symmetry/parity/conservation;
- statistical and probability assumptions;
- algebraic derivation;
- numerical traceability;
- foundational citation verification;
- build and cross-reference integrity;
- de-AI review 不得覆盖物理判断。
不适用项标 not_applicable,不能虚构检查结果。
L. Systemic consistency
- 同一符号、定义、常数、样本规模、split、cut 和假设跨章节一致。
- introduced concepts 被使用;used concepts 已定义;forward references 兑现。
- method → result → discussion → conclusion 和 abstract → intro → conclusion 闭环。
- companion claims 从当前版本重新核验。
- 跨章节矛盾或未兑现的核心 claim 是
integrity_blocker。
M. Adversarial verification
M.1 — Three-pass derivation verification. 对每个关键推导块和数值结论独立执行:
- pass-1: 量纲、代数、近似、边界、数值复算。
- pass-2: 以对手 reviewer 视角提出最强的具体反驳,并用当前证据回答。
- pass-3: 守恒、对称、渐近、概率与统计前提。
规则:
- 三 pass 不共享“已经 OK”的结论;都从源重新开始。
- 任一命中 → 修复 → 三 pass 全部重跑。
- 只有三 pass 均无未解决 scientific issue 才标
VERIFIED;否则UNDER_SCRUTINY,并产生integrity_blocker。 - 报告保留每个 pass 的方法和证据,不允许只写“checked”。
M.2 — 12-framing radial escalation. 当某 claim / 推导 / 数值三 pass 后仍可疑, 或需要 checklist 之外的开放式批判时,对该节点做辐射式 critique 升级:
- 12 个 framing 各产 ≥ 1 条 critique:first-principles 剥假设、inversion 反演、 cross-disciplinary reviewer、adversarial red-team、constraint relax/tighten、 scale extrapolation、substitution/ablation、need-question 六问、contrarian 共识可错、failure-driven、asymmetric-payoff fatal、meta 自审。
- 每条 critique 完整溯源:paper_position(file:line + 当轮 Read 的引用片段)、 evidence 论证链、全文 Grep+Read 找 paper_defense、CONFIRMED 时必带具体 proposed_fix。按 Severity/Specificity/Reproducibility/Fixability/fan-out (S/P/R/F/B) 评分。
- 每条裁决到 CONFIRMED / REFUTED / MARGINAL 之一——禁止 NEEDS-MORE-INFO 滞留、禁止 defer / future-work / “待确认”;缺 position+evidence+fix 的 CONFIRMED 自动降 MARGINAL。verdict 只表证据状态;consequence class 按 §C claim-evidence 影响单独赋予,CONFIRMED 的纯 editorial critique 仍是 advisory。
- M.2 在进程内执行、不 spawn 子代理;需要大宽度并行探索时移交用户级
cc-tree:attack(同一方法学的通用引擎),结论以 finding 形式回流本审查。
N. Staleness and drift
完整检查:
- 数字、叙述、公式依赖、结论、figure、table、labels、references、markup;
- source 脚本或数据变化后,下游全部同步;
- TODO/FIXME/placeholder/commented-out deprecated blocks 为 0;
- dead artifacts 删除,必要内容更新为当前真值,不堆叠旧版。
stale scientific content、冲突副本或 required artifact drift 是 integrity_blocker。
纯 dead prose 的删除建议通常是 advisory,除非造成歧义。
O. Process-artifact removal
- 删除“initially/originally/previous draft/we tried then changed”等本文内部研究旅程。
- 只保留当前科学状态。
- 外部 published baseline 的正式 head-to-head 比较可以保留,前提是有 citation、protocol 和 numerical comparison。
- 过程叙述若使当前 method/claim 不清或引入 stale content,按 blocker;否则按 advisory。
P. Internal and draft language
- 清理 TODO、内部路径、run/version 名、debug 名、skill/tool 名、commit-message 口吻、 实验日志语言和未定义内部缩写。
- required placeholder、内部版本残留或暴露错误来源为 blocker;普通口语和 polish 为 advisory。
Q. Reference completeness and precision
- 从论文 claim 和 field 重新识别应引用的 foundational/direct prior work。
- 对每个
\cite{}比对原文实际支持范围,分类为 CORRECT、WEAK、MISUSED、UNVERIFIABLE。 - MISUSED、fabricated、遗漏导致 novelty/attribution 错误的关键文献为 blocker。
- WEAK 或边缘 missing reference 为 advisory;不要把“可能还能多引一篇”强制成 blocker。
- paywall 不等于已验证;记录 access limitation,必要时请求用户获取原文。
R. Terminology and glossary alignment
- 优先读取项目权威 glossary/FACTS/notation source;不存在则以论文 definitions 做自洽检查,
并将外部对齐标
unmeasured。 - 关键 noun phrase、acronym、symbol 首次出现定义明确,全文单义且与权威定义一致。
- semantic conflict 为 blocker;可理解但不一致的 alias 通常 advisory。
3. Cross-validation matrix
必须完成四类对位:
- architecture in paper ↔ code/config;
- numerical claim ↔ current data/output;
- formula ↔ implementation, including corner cases;
- pipeline narrative ↔ execution order。
每个结论写 source trace,不能用“看起来一致”。
4. Fix and re-measure loop
默认修复,--no-fix 仅输出报告。
每轮:
- 按 unified priority 排序 findings;科学 blockers 先于 L0,L0 先于 advisories。
- 对每个 blocker 修根因;对每个 L0 做最小有效修改。修复遵守标准 §5.3 长度纪律:默认删减或精简,被修段落净增长必须有科学必要性理由, 报告给出每段字数差。解释性补丁(对被标记文本追加说明而非重写)是缺陷。
- 对 strong advisory:行动、接受、验证为 false positive,或带原因 pending。
- ordinary advisory 不要求消失;保留并报告即可。
- 每个 edit 后重新 Read 相关上下文。
- 改公式/数字/物理链后重跑 M 三 pass;改 figure/table 后重新查看制品与来源; 改 citation 后重新读原文;改 prose 后重新跑 shared linter。
- 重跑权威 build 和所有受影响测试/脚本。
- 收尾前跑 length gate(标准 §5.3 机械执行):
python tools/length_gate.py <file> --before <scratch>/length-baseline.tex(或--git-ref <ref>)。exit 1 = 存在无理由净增长——精简回预算内, 或用--allow "<section>=<理由>"记录作者批准的理由后重跑至 exit 0。length-growthfinding 未 disposition 时循环不得收尾。
伪代码:
for iter in 1..max_iter:
report = full_A_to_R_review_from_current_sources()
if no integrity_blocker \
and no l0_target \
and no UNDER_SCRUTINY derivation \
and every strong advisory has a disposition:
stop with DISPOSITION_COMPLETE
if no_fix:
stop with REVIEW_ONLY
apply_ranked_root_cause_fixes(report)
re_measure_affected_axes()
if budget exhausted:
return BREAK_WITH_USER_DECISION with unresolved findings
不得自动增加 --max-iter。预算耗尽不是 paper failure verdict,而是 review workflow
尚未完成;把 pending blockers、L0 和 strong advisories交给用户决定下一步。
5. Isolated modern-physics verification
单独运行 paper-review 时,达到进程内 disposition-complete 状态后,默认由主代理启动 isolated worktree agent,重新读取 host-level modern-physics-review skill 并 cold-read 目标论文。它不能继承当前审查的“已经验证”结论。
--skip-final-mpr: 仅用户显式选择时跳过,报告unmeasured/SKIPPED_BY_USER。--no-isolated-mpr: 仅供 final-review 等父 orchestrator 避免 nested agent;报告SKIPPED_FOR_ORCHESTRATOR,由父级独立执行。- isolated review 产生的 critique 先判断 evidentiary verdict,再单独赋 consequence class。 CONFIRMED editorial critique 不自动成为 blocker。
- 新 blocker/L0/strong advisory 回注主循环;agent failure 必须显式报告,不能伪装通过。
6. Report contract
# Paper Review — Typed Feedback Report
Target: <file> | Iterations: K | Field: <field or none>
Workflow state: DISPOSITION_COMPLETE | REVIEW_ONLY | BREAK_WITH_USER_DECISION | EXECUTION_FAILURE
## Measurement state
| axis | status | provenance / limitation |
## Summary
- integrity_blocker: total / resolved / pending
- l0_target: total / resolved / pending
- strong advisory: acted / accepted / false-positive / pending
- ordinary advisory: total / reported
- M derivations: VERIFIED / UNDER_SCRUTINY
- build: measured status and result
- length budget (§5.3): gate exit / total words before -> after / justified sections with reasons
- isolated MPR: measured / unmeasured / orchestrator-owned / failed
## Ranked findings
| id | kind | layer | rule | location | evidence | source trace | action | disposition |
## A–R coverage
For every dimension: measurement status, methods used, and finding IDs.
Do not print PASS/FAIL rows.
## Numerical anchors
| quantity | paper claim | source | verification |
## M record
| derivation | pass-1 evidence | pass-2 objections/answers | pass-3 invariants | M.2 escalations (framing/verdict) | status |
## Residual feedback
- pending strong advisories with reasons
- ordinary advisories
- degraded/unmeasured axes
- author decisions required
The human-readable report and JSON must come from the same structured findings. Totals must not change when detail is truncated.
7. Stopping semantics
Review may stop as disposition-complete only when:
- all integrity blockers are resolved or verified false positives;
- applicable L0 targets are zero;
- all critical derivations are VERIFIED;
- required build and artifacts are valid;
- the length gate (standard §5.3) exits 0 against the loop's pre-edit baseline, every recorded growth justification included in the report;
- every strong advisory has an explicit disposition or a stated pending reason;
- ordinary advisories and unavailable axes are reported.
This does not assert that the paper is universally perfect, human-authored, accepted by a journal, or free of subjective editorial alternatives.
8. Anti-patterns
- Declaring “grep found nothing, therefore correct”.
- Reusing numbers or citations from memory.
- Looking only at captions instead of figures.
- Softening a wrong claim instead of correcting its source, math, data or scope.
- Calling a learned score proof of authorship.
- Treating a missing baseline as zero findings.
- Requiring every yellow/editorial suggestion to disappear.
- Converting every confirmed critique into a blocker.
- Demanding exactly one narrative thread when multiple related contributions are legitimate.
- Calling advisory-only exit 0 a scientific or submission PASS.
- Skipping M pass-2/pass-3 because pass-1 looked correct.
- Hiding unresolved items after iteration budget exhaustion.
- Starting a nested sub-agent when
--no-isolated-mprdelegates MPR to the parent orchestrator.