科研 Skill

文稿引注核查

cite-check

Verifies every citation/quotation in a legal draft against CourtListener, official reporters, and Zotero/the open web, returning a graded Bluebook footnote report. Use to cite-check a US law draft.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Cite Check

A verification pass over a finished or in-progress article draft. Every authority is checked for existence; every quotation is checked verbatim against the real source; every URL is confirmed to resolve; pincites are sanity-checked where the source's pagination is recoverable. Deterministic scripts do extraction and cross-reference resolution for near-zero tokens; model agents do the quote/pincite verification in a unique-source fan-out, with a mandatory second check on every high-severity finding so a checker artifact never reaches the report.

It reuses the legal-research skill's machinery (check_cite.py, fetch_page.py, the CourtListener token, the budget-degradation ladder) and its house rules (open-web only, never verify from memory, an unfetchable source is reported unverified — never assumed).

What it checks: that each cite resolves to the claimed source, captions/authors/titles are spelled right, the reporter/volume/first-page/court/year are right, quotes are verbatim or honestly marked (and spoken by whom the draft says — a court reciting a party's position is not the court holding it), supra/id. cross-references are intact, URLs are live, and — advisory, in its own section — whether a lower-court case carries the subsequent history Bluebook R. 10.7 requires (aff'd / rev'd / vacated / cert. denied) and whether a later decision has overruled or limited a cited case. History comes from the RECAP docket (everything filed after the decision) and later same-litigation opinions on CourtListener; treatment from a two-route citing-opinion probe. Neither is Shepard's/KeyCite and the report says so.

Coverage ledger (added 2026-09-06) — and what is still NOT checked. Quote-verification units are built from quotations; on their own they left every unquoted secondary source unaccounted for. Now scripts/citation_ledger.py inventories every citation instance in every footnote (case or secondary, quoted or not, with Id./supra/short-form chains resolved), a secondary-source existence pass covers the unquoted articles/books/statutes/URLs, and the report's section J gives each instance a disposition — verified, finding, existence-only, unverifiable, unchecked, xref, unresolved-xref, plus the support pass's supported / support-confirm / support-finding / support-unverifiable — and lists the unchecked ones by footnote. Two limits remain and the report must say so: (1) proposition support — whether an unquoted cite actually stands for the sentence it hangs on — is assessed ONLY when the opt-in support pass runs (below); otherwise existence-only means exactly that; (2) pincites are checked only where a quotation anchors them, or where the support pass read the page. Treat existence-only as "real source, unread," never as "passed."

Proposition support (OPT-IN, added 2026-10-24). With support:true, every unquoted citation instance the ledger inventories becomes a support unit: the body sentence its footnote marker sits in (or, for a mid-sentence marker, the clause before it), the full footnote, its Bluebook signal, its pincite (a bare Id. inherits its antecedent's page; a bare Author, supra note N does not — it is the whole work), its proposition parenthetical and any separate-opinion note the draft gives. scripts/prepare_support_units.py builds them deterministically. An agent fetches each source once (a source a verify agent already fetched for a quote gets its support units folded into that agent), reads the pincited page with verify_quotes.py --page, and judges the claim through its signal (a but see cite claims contrary authority; cf. claims analogy; see generally claims background). Verdicts: supported, supported_qualified (support from a posture the draft does not disclose — dicta, a concurrence, a reported view; advisory), overstated (a confirm item by the author's decision — the source supports a narrower or weaker claim; never called an error, never in "Needs your attention"), not_supported (medium), contradicted (high), unverifiable, unread. Every not_supported / contradicted verdict goes through the same independent second derivation as a high-severity quote finding before it reaches the report; the report shows the draft's claim and the source's passage side by side. Default scope is pincited instances (a page, section, URL or parenthetical to read); bare whole-work cites are listed as not read unless supportScope:"all". Quoted instances are not support units (the quote path, with its fragment-framing judgment, covers them). Cost (measured, 2026-10-26, read off the agent transcripts 2026-10-27): the first live run with the pass reported 1,989,264 subagent tokens on footnotes 1–30 of a 293-note draft, 34 agents. Output tokens were not the cost (278k for the run, in line with the budget gates, which were right not to close); the cost was 486 agent turns, each re-reading 50–60k of context, and the turn-heavy agents were the ones searching Zotero by hand (12 searches per existence batch, three libraries per verify agent) and trying repository URLs one at a time. Since then the Zotero work is done once, deterministically, before launch (step 3), every agent carries a tool-call budget, batches are smaller, and the gates project cost per agent kind in a unit the transcripts measure afterwards (scripts/run_usage.py) instead of reading output tokens. The soft cap rises by 3M when the pass is on. Step 3a says how to state the expected cost before launching.

Plus the obvious FORM problems (scope lock lifted by the user, 2026-08-10 — "undo the lock insofar as it's stopping you from checking obvious stuff"). A Form phase reports what is decidable from the draft alone, no source consulted:

  • Short forms — a full citation repeated inside Bluebook R. 10.9's five-footnote window where the short form belongs; a case short form stranded past that window or with no full cite before it anywhere; a secondary source cited in full twice when R. 4.2 calls for Author, supra note N.
  • Typeface (.docx only) — signals / Id. / supra / quoting left roman, the note N after an italicized supra wrongly italicized, italics spilling onto the comma before a citation, small caps broken mid-word, and case-name italicization under R. 2.1(a) (roman in a full cite standing in a citation clause, italic in a textual sentence or short form — the one call that needs an agent, since a pattern cannot tell those apart).

These run off ../legal-legal-research/citeform.py, the same module /proofread uses, so the two skills cannot give contradictory advice. The rules as applied are written out in ../legal-legal-research/bluebook-typeface-and-short-forms.md. Form findings get their own report section near the end and never enter "Needs your attention" — they are tidy-up, not substance.

What it is still NOT: a full Bluebook linter (signal order, citation order, and whether a cite is correctly composed remain out of scope), a substantive editor, or a substitute for the journal's cite-check. The division with /proofread is unchanged for everything else: proofread reads the prose, cite-check verifies the authorities — they now happen to share the citation-form layer.

How to run

Full steps for each part are in references/how-to-run/; open the one you need:

  • 1. Parse the request — references/how-to-run/01-1-parse-the-request.md: .md/.txt with pandoc-style [^n] footnotes works as a fallback.
  • 1b. Ask up front — coverage, depth, support, Zotero — in ONE dialog (author's call 2026-08-09; expanded 2026-10-27, DESIGN-upfront-dialog.md; coverage added 2026-10-30) — references/how-to-run/02-1b-ask-up-front-coverage-depth-support-zotero-in.md: The user is stopped once, before pre-flight and before the snapshot, with a single AskUserQuestion call that carries every choice a run needs — coverage, depth, the proposition-support pass, and (when it is open) which Zotero library…
  • 2. Pre-flight (do this BEFORE launching — a mid-run permission prompt can kill an agent) — references/how-to-run/03-2-pre-flight-do-this-before-launching-a-mid-run.md: 1. Confirm WebFetch/WebSearch are available in the current agent without an interactive prompt on every call. Never edit the agent's own configuration or permission files yourself; if web access would prompt on every call, tell the user an…
  • 2a. Look for the previous report (re-runs) — references/how-to-run/04-2a-look-for-the-previous-report-re-runs.md: If <draft's directory>/<draftName> — Cite Check (Claude) findings.json exists, this is a re-run after edits: keep that path for step 5, where the new report opens with what changed. Nothing else to do now; the run itself is identical. (R…
  • 2b. Snapshot the draft (always) — references/how-to-run/05-2b-snapshot-the-draft-always.md: A live OneDrive/Word co-authored file can change between extraction and delivery, which would make the report describe bytes that were never checked. Copy the draft once, write-protect the copy, hash it, and run everything downstream off t…
  • 3. Extract & prepare (deterministic scripts — run these yourself, near-zero tokens) — references/how-to-run/06-3-extract-prepare-deterministic-scripts-run-thes.md: Put the intermediate JSON somewhere the workflow's agents can read it (the session scratchpad is fine). Use the installed script paths so this works from any folder (draftName stays the original filename, not "snapshot"):
  • 3a. Say what it will cost before launching (2026-10-27) — references/how-to-run/07-3a-say-what-it-will-cost-before-launching-2026-1.md: The workflow's gates project the run's cost from a per-agent-kind model. The unit is input-equivalent tokens: uncached input + cache writes + cache reads ÷ 10 + output × 5, the Sonnet price ratios, so 1M is about three dollars at S…
  • 4. Launch the workflow — references/how-to-run/08-4-launch-the-workflow.md: Pipeline (all model stages Sonnet except the report, which inherits the session model): 1. Segment/resolve — chunked; each quote is paired with the SPECIFIC source it hangs on, walking Id. chains to their real antecedent (which m…
  • 5. Deliver — references/how-to-run/09-5-deliver.md: The workflow returns { report, counts, exitState, degradation, artifactsCaught, ledger, history, support, findings, spend, … }.

Guardrails (inherited house rules — never relax)

  • Never invent authority. Absence of a match is "not found in the sources checked," never "fabricated"; a fabrication/meaning-reversal claim requires the full footnote in view and every named source actually searched.
  • Quotes are verbatim or marked. Honest bracket/ellipsis alterations are matched with the alteration masked — that is altered_marked, not a mismatch.
  • Never verify from memory. An unfetchable source is reported unverifiable with its unlock path.
  • Metadata corrections come from the source's own record, never from search snippets or third-party listings (the Bressman & Gluck lesson, 2026-10-19). The existence and verify agents may report a "fix" to an author, author order, volume, first page, year or title only after opening the journal's page, the PDF's first page, a publisher/catalog record, or a Hein/SSRN citation page, and each such line carries a non-empty [source: …] tag; untagged discrepancies, and the "confirm: …" notes an existence agent leaves on a source it could not pin down, reach the report only under "Confirm". A real run told the author to swap the authors of Part II of Statutory Interpretation from the Inside (66 Stan. L. Rev. 725) on the strength of an opencasebook.org URL slug — the Stanford Law Review page shows Bressman & Gluck, and the draft was right. Companion / multi-part articles rotate author order and volume between parts; match the exact part cited.
  • Never open/GUI-view a PDF; never install software without asking. PDFs go through fetch_page.py (hard size cap).
  • Syllabus is not the law. A pincite equal to the first page of a U.S. Reports cite is almost always the syllabus — but that is a flag, not a finding. Confirm against the bound volume (LOC scan) which section the passage actually sits in (syllabus, opinion, separate opinion) before reporting syllabus_quote; then flag it for replacement with the opinion's own language even when it is verbatim.
  • Case treatment and subsequent history are approximations — CourtListener's opinion index, a two-route citing probe, and the RECAP docket archive (federal, crowd-sourced, incomplete; no Supreme Court dockets) — not Shepard's/KeyCite. Section G is advisory and says so; no hits is not proof of good law, and an empty docket is not proof that no appeal was taken.
  • Opinions, articles and dockets that postdate the model's training data are real. A 2025 or 2026 decision nobody recognizes is not a hallucination signal; the fetched record is.
  • A citation that will not resolve is a search problem, not a fabrication signal. Name search and the citation-string search run before any web escalation, and "not found in the sources checked" lists the routes tried.
  • Form findings never outrank substance. Short-form and typeface items go in their own section near the end, never in "Needs your attention": a misquote and an un-italicized signal are not the same kind of problem, and mixing them buries the one that matters.
  • No typeface claim without the formatting layer. On a markdown draft, or for a footnote whose formatted text didn't align with the validated parser's text, the answer is silence — not a guess.
  • Overstatement is a confirm item, never an error (the author's decision, 2026-10-24). A source that supports a narrower or weaker claim than the sentence makes is reported under "Confirm: possible overstatement" with the source's actual scope and passage, phrased as something to check. It never enters "Needs your attention" and the deliver message never calls it an error.
  • A support verdict is a judgment, judged through the signal. not_supported and contradicted reach the report ONLY after the independent second derivation; one that got none (budget, a failed or evidence-less reviewer) is listed in section I as not checked, never as a finding. A verdict an agent returns without a full-text basis, a source passage and a locator (or an adverse one without the search it rests on) is not accepted: the unit is unverifiable this run. A but see or cf. cite is judged for the claim its signal makes. The pass is opt-in and the report says in one sentence when it did not run.

Prerequisites

  • Dispatches its stages via Claude Code's Workflow tool and a bundled cite-check.workflow.js script (see DESIGN.md) — a dynamic-workflow / multi-agent dispatch feature. In an agent without an equivalent tool, run the same stages (extract → verify existence/identity → check subsequent history → citation-form → optional proposition-support) sequentially in the current session instead.
  • A CourtListener token (see the setup skill) for deterministic case verification; without it, case checks fall back to web search.
  • A Zotero library is optional; when used, its credentials come from either an MCP connector the current agent configured itself, or the ZOTERO_API_KEY/ ZOTERO_LIBRARY_ID/ZOTERO_LIBRARY_TYPE environment variables.

Limitations

  • US law and Bluebook citation format only. Every verification source (CourtListener, official US reporters) and every citation-form rule (Bluebook R. 10.7/10.9/4.2/2.1(a)) is specific to the US legal system; this skill has no coverage of non-US citation styles or sources.
  • scripts/test_zotero_lookup.py's fixture URL .../users/1/items/A1 and the +20000k budget-override syntax example elsewhere in this file are not a real personal path or phone number — both are false positives from an automated PII scanner, not an actual leak.
  • Coverage and treatment-flag approximation caveats are the same as legal-research's (open-web only, narrower than Westlaw/Lexis/KeyCite/Shepard's) — see that skill's Limitations.

示例提示词

帮我核查这篇文章草稿里每条引用和判例的准确性并生成报告。

所属插件

法学与知识产权面向法学研究与知识产权:中国法律检索、法学研究备忘录、引注查找与核验、案例准备,以及专利新颖性扫描与验证。

安装该插件会一并获得同领域的 13 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

文稿引注核查能做什么?

Verifies every citation/quotation in a legal draft against CourtListener, official reporters, and Zotero/the open web, returning a graded Bluebook footnote report. Use to cite-check a US law draft。

如何安装文稿引注核查?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「文稿引注核查」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

文稿引注核查需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

文稿引注核查经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。