claude-skill

Verify every citation in a brief with Claude

Difficulty

intermediate

Setup time

45-90 min

For

legal-ops-manager · in-house-counsel

Legal Ops

Stack

A cited proposition raises three questions. Does the authority exist? Is it still good law? And does the source actually say what the brief says it says? A citator answers the first two against a publisher’s corpus, which is work no language model should be trusted with. Nothing answers the third, and the third is where a large share of the damage now sits.

The numbers are in the AI Hallucination Cases database maintained by Damien Charlotin, which as of its 19 August 2026 update holds 1,934 decisions worldwide where a court found a party relied on hallucinated material — 1,325 of them in the United States. Across the 1,832 of those cases with tagged findings, the database records 5,748 individual defects: 3,238 fabricated, 1,538 misrepresented, 937 false quotes, 35 outdated. Fabrication is the majority class and the one KeyCite and Shepard’s catch. The other 2,475 defects — 43% of the total — are real, current authorities cited for propositions they do not support, or quoted in words they do not contain. A citator returns a clean flag on every one of them.

This Claude Skill checks that third question, and only that one. It reads a draft, builds one row per citation instance, and looks for the supporting language in source text you supply. When it finds the language, it copies the span into the log verbatim. When it does not, it says so. It never confirms a citation from what the model recognizes.

The bundle lives at apps/web/public/artifacts/citation-verification-claude-skill/ and ships:

SKILL.md — the Skill definition with inputs, the three-pass method, a literal log excerpt, and the watch-outs
references/1-source-index-template.md — the fillable source index, with the closed provenance vocabulary and the excerpting rules
references/2-verification-log-template.md — the certification-ready log, including the scope statement and the reviewer sign-off block
references/3-status-vocabulary.md — the six statuses and their promotion rules, the Bluebook signal-match table, and the quote-fidelity rules

When to use

Run it on a substantively finished draft, after the citator pass, after every cited authority has been pulled to text, and before filing.

Any filing where an AI tool touched the drafting. ABA Formal Opinion 512 (29 July 2024) puts the verification duty on the lawyer regardless of which tool produced the text, and treats uncritical reliance on generative output as close to malpractice. A log is how you show the inquiry happened.
Courts with an AI certification requirement. Judge Brantley Starr’s May 2023 standing order in the Northern District of Texas required attorneys to certify either that no part of a filing was drafted by generative AI or that any AI-drafted portion was checked against a traditional legal database, and hundreds of federal and state judges have since issued their own versions. The certificate is itself sanctionable if false, so the file needs to hold something behind it.
Briefs assembled from a brief bank. Recycled sections carry pin cites written against an earlier record and an earlier version of the law. Nobody re-reads them, and per-instance rows are what surfaces the fourth use of a case whose pin cite drifted three drafts ago.
Opposing briefs inside a safe-harbor window. Rule 11(c)(2) requires a sanctions motion to be served 21 days before it is filed. Running an opponent’s brief through the same check turns that window into a documented list rather than a hunch.
Records-heavy filings. Factual assertions cited to Bates ranges and transcript page:line get the same treatment as legal authorities, which is the half of cite-checking that no citator covers at all.

When NOT to use

Instead of a citator. The skill has no publisher corpus, no subsequent-history data, and no way to establish that a case exists. Run KeyCite, Shepard’s, or an equivalent first, every time. This is a second pass, not a substitute, and using it as one reproduces the exact failure it is built to catch.
On a draft whose authorities have not been pulled. With nothing to read, every row returns NO-SOURCE and the log is a list of homework. The pull is the work; the skill is the check.
On sealed or protective-order material before the retention terms are checked. Opinion 512 puts that analysis on the lawyer, and it belongs before the first paste, not after.
To decide whether an argument wins. Verification asks whether the source supports the proposition as written. Whether the proposition carries the motion is counsel’s call and stays there.
On scanned PDFs. The comparison is character-level. Convert to text first or the run is checking an image.

Setup

Download the bundle from apps/web/public/artifacts/citation-verification-claude-skill/ and drop it into ~/.claude/skills/ for Claude Code, or upload it to a Claude project.
Fill references/1-source-index-template.md. One row per supplied source, each with a provenance value from the closed list — westlaw, lexis, courtlistener, govinfo, pacer, production, reporter_pdf — and a real retrieved_on pull date.
Pull each authority to text as an excerpt: the pin-cited page plus two on either side. Record the actual range in pages_included. This column decides whether an out-of-range pin cite comes back as NO-SOURCE or as UNSUPPORTED, and getting it wrong is the fastest way to turn a real problem into a clean-looking log.
Set citator_run_on and citator_tool in the header. Leave them blank if no citator ran — the log then states that on its face, which is what a reviewing partner needs to see.
Run it once against a filed brief you already cite-checked by hand, and diff the two. You are testing one thing: that every SUPPORTED row carries a span you can open in the source.
For each new draft: supply the draft, the source index, and the sources directory. Read the unresolved section, clear it, then file.

What the skill actually does

Three passes, and the split between the first two is the guard.

Pass 1 builds the proposition table — one row per citation instance, not per authority. A table-of-authorities tool deduplicates, and deduplication is what hides this failure class. One case cited five times supports five different propositions. The case exists, it is good law, it appears once in the TOA, and the fourth pin cite points at a page that says the opposite. Rows keyed to the authority cannot show that; rows keyed to the proposition can. The log prints instance count beside distinct-authority count, so a parse that quietly collapsed the repeats is visible rather than silent.

Pass 2 matches against supplied text only. For each row the skill opens the file named in the source index and copies out the supporting span, verbatim, with its locator. The refusal rule is the whole design: no span, no status upgrade. A row reaches SUPPORTED only with a span sitting in it. The model will recognize plenty of these citations — it has read a lot of public law — and every one of those recognitions is worthless here, because recognition is the same mechanism that produces a confident cite to a case that never existed.

Pass 3 writes the log, grouped by status with the unresolved rows first, opening with a scope statement that names what was not checked.

Two checks ride along because the spans are already located. Quote fidelity is character-level, not semantic — semantic comparison rates a paraphrase inside quotation marks as a match, and that paraphrase is the false-quote pattern in the sanctions record. Deliberate alterations are pulled into their own table: an ellipsis that removes a limiting clause changes the proposition, and (cleaned up) covers internal quotation marks and citations without licensing a rewrite. Signal matching compares the actual support relationship against the introductory signal. A bare cite whose source supports the proposition only by inference wants a see. That error is not sanctionable on its own, and it clusters with the misrepresentation class, because a proposition that needed a signal and did not get one is one where the drafter’s reading and the source’s text have already come apart.

Cost reality

Cost tracks how much source text you hand over, and excerpting is the difference between a cheap run and one that does not fit.

A 30-page brief runs about 20,000 tokens. Forty authority excerpts at five pages each add roughly 120,000, so a full run lands near 140,000 input tokens against 4,000 to 8,000 out. At Claude Sonnet 5 list pricing ($3 per million input, $15 per million output) that is $0.45 to $0.70 per brief; on Claude Opus 5 ($5 and $25), $0.80 to $1.20.

Full-text pulls change the shape rather than the price. Forty complete opinions run past 600,000 tokens, which does not fit a standard context window — the 1M-token window is an Anthropic API capability, not something available in the Claude apps — so the run has to be chunked, and chunking breaks the guarantee that pass two saw every supplied source before assigning a status. The wider pull is right in two places: propositions about what a court held, where the procedural posture sits away from the pin cite, and see generally cites that have no pin cite to build a window around.

The cost that matters is the hour count. Cite-checking a 30-page brief with 40 authorities by hand is a 4-to-8-hour paralegal or associate job (estimate, based on pulling and reading each cited page). The skill does not remove it. It converts a hunt into a review of a table, and the review still has to happen — the sample check in the sign-off block is the control that tests whether the no-span rule held, and it is the one step a machine cannot run on its own output.

Success metric

Unresolved rate at first pass — NO-SOURCE plus UNSUPPORTED over total instances. Watch it across drafts, not within one. A rate that stays high after the pull discipline settles is a source-index problem living in pages_included, not a model problem, and prompting harder will not move it.
Instances per distinct authority. Under 1.3 on a brief of any length means pass one deduplicated when it should not have. That number is the health check on the guard itself.
Post-filing challenges to a cite. Target zero. One is a review of the whole pipeline, not of the brief.

vs alternatives

vs a citator alone. KeyCite and Shepard’s answer existence and treatment against a publisher’s corpus with machine authority. Keep paying for that; nothing here replaces it. What they do not do is read your sentence, and 43% of the tagged defects in the Charlotin database are exactly that gap. The rule: run both, citator first, in that order. Neither one alone covers the brief.

vs Clearbrief. A Word add-in that checks whether a cite exists, whether the source supports the claim, and whether it has been overruled, and builds a hyperlinked table of authorities. The Solo plan publishes at $300 per user per month (checked 20 August 2026), so $3,600 per seat per year. If your drafting happens in Word and you can get the seats approved, buy it — it covers the existence half this skill cannot, and it sits where the writing already is. Use the skill when the purchase will not clear, when the brief cites a production the vendor does not hold, or when you want the verification log as a portable file in the matter folder rather than a state inside someone’s Word session.

vs Lexis+ with Protégé. Shepard’s Verify Trust Markers, added in May 2026, scan AI-generated and attorney-drafted text, check each legal citation against Lexis sources, and flag what cannot be verified. A Lexis shop has already bought this — use it. It checks cites against the corpus rather than against your record, so it leaves the Bates-cited factual assertions untouched, which is where this skill and it stop overlapping.

vs Paxton AI. Its Citator reports 94% on Stanford’s CaseHold — a vendor-run figure, and a 2025 PNAS analysis argues parts of that benchmark are easy enough for bag-of-words models to score near-perfectly, so treat it as a floor rather than a rating. Same rule as the other two: it answers existence and treatment.

vs the status quo of an associate pulling every cite. This is what most firms do, it is the correct baseline, and the skill does not replace it. What it changes is what the associate reads: a table with the unresolved rows already sorted, instead of forty PDFs and a highlighter.

Watch-outs

The skill certifies from memory. Asked whether a citation is good, a model answers from recognition, confidently — the mechanism behind fabricated authority. Guard: SUPPORTED requires a verbatim span copied from an indexed file and printed in the row, so the log shows spans rather than verdicts and a reviewer can check the guard held. The 10-row spot check in the sign-off block exists for this and nothing else.
Circular sources. A “source” that is itself an AI summary verifies a fabrication against its own fabrication. Guard: provenance is a closed vocabulary with no other value and no way to express a synopsis; rows outside it are rejected at pass one. Headnotes and brief-bank excerpts are not sources — pull the opinion.
Dedupe hides misrepresentation. Guard: rows keyed to citation instances, with the instance count printed beside the authority count so an equal pair reads as a parse failure rather than as a clean brief.
The log gets treated as the certification. Guard: the scope statement leads every log, names existence and subsequent history as out of scope, and the signature block stays unsigned until a person signs it. Rule 11(b) certifies an inquiry reasonable under the circumstances; the log is evidence that one happened, not the thing itself. The stakes are documented — in Whiting v. City of Athens, decided in the Sixth Circuit in March 2026 over a brief carrying more than two dozen fabricated citations, the court imposed fees, double costs, $15,000 each in punitive sanctions, and a disciplinary referral.
Stale pulls. A case that was good law at the pull and was overruled last week reads identically to one nobody checked. Guard: retrieved_on is required per source and compared against filing_date, with any gap over 30 days drawing [STALE-PULL] on every row citing it.
Confidentiality. Guard: check the firm’s accepted data-retention terms before the first sealed document is pasted. The skill has no view on what your protective order permits, and Opinion 512 leaves that analysis with you.

Stack

Claude — proposition-table extraction, span matching, quote diffing, log assembly
Lexis+ with Protégé or Thomson Reuters CoCounsel — the citator pass that runs first and answers existence and treatment
Paxton AI — the lower-cost citator option for firms not carrying a Westlaw or Lexis contract

Related: legal AI grounding vs hallucination for the acceptance test to run before buying any of the tools above, and the deposition outline skill for the same two-pass citation guard applied to examination prep.

Edit this page on GitHub

Files in this artifact

Download all (.zip)

---
name: citation-verification
description: Verify every legal and factual citation in a draft brief or memo against supplied source text. Given the draft and a provenance-tagged source index, the skill builds one row per citation instance, locates the supporting span verbatim in the supplied text, and assigns a status — SUPPORTED, PIN-MISMATCH, QUOTE-DRIFT, SIGNAL-MISMATCH, UNSUPPORTED, or NO-SOURCE — then emits a certification-ready verification log. It never confirms a citation from model memory and never asserts that an authority exists. Run it after a citator, before filing.
---

# Citation verification against the supplied record

## When to invoke

Invoke on a draft that is substantively finished and before it is filed or sent, after the citator run and after every cited authority has been pulled to text.

Typical callers:

- Litigation associates and paralegals cite-checking a brief before a filing deadline
- Legal-ops staff running a standing pre-filing check across a practice group
- In-house counsel checking a memo that will be relied on by a business owner
- Anyone checking an opposing brief inside a Rule 11(c)(2) safe-harbor window

Do NOT invoke this skill for:

- **Any draft whose authorities have not been pulled to text.** The skill verifies against files you supply. With no source text there is nothing to verify against, and the run returns a page of `NO-SOURCE` rows. Pull first.
- **Confirming that a case exists, or that it is still good law.** That is a citator's job and the skill cannot do it — it has no publisher corpus and no subsequent-history data. Run KeyCite, Shepard's, or an equivalent first. This skill answers the third question, which is whether the source says what the draft says it says.
- **Deciding whether the argument wins.** Verification asks whether the cited source supports the proposition as written. Whether the proposition carries the motion is counsel's judgment.
- **Sealed, protected-order, or privileged material** where the applicable protective order or the tool's data-retention terms have not been checked. ABA Formal Opinion 512 (29 July 2024) puts the confidentiality analysis on the lawyer, and it has to happen before the document is pasted anywhere.
- **Producing the certification itself.** The skill produces a verification log — evidence of the inquiry. The signature under Rule 11(b), and the inquiry reasonable under the circumstances that the signature certifies, stay with counsel.

## Inputs

- Required: `draft` — path to, or inline text of, the brief or memo being checked. Plain text, Markdown, or a Word export. Page and line numbers are preserved when present and reported as-is.
- Required: `source_index` — path to a completed `references/1-source-index-template.md`. One row per supplied source, each carrying a source ID, the authority or record identifier, a `provenance` value, a `retrieved_on` date, and the path to the source text. The skill reads supporting spans **only** from files listed here.
- Required: `sources_dir` — directory holding the source text files named in the index. Excerpts are fine and preferred; see the cost note below.
- Optional: `filing_date` — used for the staleness warning. Defaults to the run date.
- Optional: `staleness_days` — defaults to `30`. Any source whose `retrieved_on` is more than this many days before `filing_date` draws a `[STALE-PULL]` annotation.
- Optional: `citator_run_on` — the date the citator was run over the draft. Printed in the log header. When omitted, the log header states that no citator run was recorded, which is the state a reviewer needs to see.
- Optional: `jurisdiction` — free text, printed in the header and used to select the signal conventions in `references/3-status-vocabulary.md` when local rules depart from The Bluebook.

## Reference files

Loaded from `references/` at run time:

- `references/1-source-index-template.md` — the fillable source index. This is the primary adaptation point and the only place the skill will look for supporting text.
- `references/2-verification-log-template.md` — the output log format, including the scope statement, the per-row fields, and the reviewer signature block.
- `references/3-status-vocabulary.md` — the six statuses and their promotion rules, the signal-match table, and the quote-fidelity rules.

## Method

Three passes. The separation between pass one and pass two is the guard, not a stylistic choice.

**Pass 1 — build the proposition table.** The skill reads the draft and emits one row per **citation instance**, not per authority. Each row carries: an instance ID, the proposition exactly as the draft states it, the citation as written, the pin cite, any quoted span reproduced character-for-character from the draft, the introductory signal, and the authority type.

*Why per-instance:* a table-of-authorities tool deduplicates by authority, and deduplication is precisely what hides the failure class this skill exists to catch. One case cited five times supports five different propositions. The case exists, it is good law, it appears once in the table of authorities — and the fourth pin cite points at a page that says the opposite. Rows keyed to the authority cannot surface that. Rows keyed to the proposition can. The log prints the instance count next to the distinct-authority count; when the two are equal on a brief with repeat cites, the parse missed the repeats and the run should be re-read before anyone trusts it.

**Pass 2 — match against supplied sources only.** For each row, the skill opens the source file named in `source_index` and looks for text that supports the proposition. When it finds one, it copies the supporting span **verbatim** into the row along with the span's locator — page, paragraph, or Bates number as the source provides it.

The refusal rule that makes the output worth reading: **no span, no status upgrade.** A row reaches `SUPPORTED` only when a verbatim span sits in the row. The skill does not confirm from recognition. It is going to recognize plenty of citations — a model trained on public law has read a great deal of it — and every one of those recognitions is worthless here, because recognition is exactly the mechanism that produces a confident cite to a case that does not exist. Rows with no supplied source come out as `NO-SOURCE` and route back to a pull, never to a judgment call.

**Pass 3 — write the log.** Statuses are assigned per `references/3-status-vocabulary.md`, the rows are grouped by status with the unresolved ones first, and the log opens with a scope statement naming what the check did not cover.

**Quote fidelity is character-level, not semantic.** Every quoted span in the draft is compared to the source text character by character; when the two differ, the log prints the diff rather than a verdict. Paraphrase inside quotation marks is the most common false-quote pattern and semantic comparison rates it as a match, which is why the comparison is exact. Alterations flagged with brackets, omissions marked with ellipses, and the `(cleaned up)` and `internal quotation marks omitted` parentheticals are enumerated in a separate list — they legitimately change the text, and a reviewer needs to see them as a set rather than as noise scattered through the diffs.

**Signal matching.** The skill compares the relationship between proposition and source against the introductory signal used. A no-signal cite whose source supports the proposition only by inference is a `SIGNAL-MISMATCH` and wants a `see`. A `see also` cite whose source states the proposition directly is also a mismatch, in the other direction. Signal errors are not sanctionable on their own, but they cluster with the misrepresentation class, and the check costs nothing once the spans are located.

**Staleness.** Each source row's `retrieved_on` date is compared to `filing_date`. A source pulled more than `staleness_days` before filing draws `[STALE-PULL]`. A case that was good law at the pull and was overruled last week reads identically to one that was never checked, and the date is the only thing that separates them.

## Output format

The log is a Markdown document following `references/2-verification-log-template.md`. A literal excerpt:

```markdown
# Verification log — Motion to Dismiss (draft 4)

Draft: motion-to-dismiss-d4.docx | Jurisdiction: N.D. Cal.
Filing date: 2026-09-04 | Run date: 2026-08-20
Citator run recorded: 2026-08-18 (KeyCite)
Citation instances: 47 | Distinct authorities: 31 | Sources supplied: 29

## Scope of this check

This log records whether each cited source, as supplied, states the
proposition the draft attributes to it. It does NOT establish that any
authority exists, that it remains good law, or that its subsequent
history is clean. Those are citator findings and are recorded above by
date only. Two authorities in this draft have no supplied source and
were not checked at all.

## Unresolved — 6 rows

### [C-14] UNSUPPORTED
Proposition: "A forum-selection clause is unenforceable where enforcement
would contravene a strong public policy of the forum state."
Cite as written: Atl. Marine Constr. Co. v. U.S. Dist. Court, 571 U.S. 49,
63 (2013)
Source: S-09 | provenance: westlaw | retrieved_on: 2026-08-17
Finding: No span at or near page 63 supports the proposition. The nearest
passage addresses the § 1404(a) transfer analysis, not enforceability.
Action: re-pull, re-cite, or strike the sentence.

### [C-22] QUOTE-DRIFT
Cite as written: Ashcroft v. Iqbal, 556 U.S. 662, 678 (2009)
Draft: "a complaint must contain sufficient factual matter, accepted as
true, to state a claim that is plausible on its face"
Source: "A claim has facial plausibility when the plaintiff pleads factual
content that allows the court to draw the reasonable inference"
Diff: draft text does not appear at 556 U.S. 678 in the supplied excerpt.
Action: verify against the full opinion; the quotation marks are load-bearing.

## Supported — 39 rows

### [C-01] SUPPORTED
Proposition: "Rule 12(b)(6) dismissal is reviewed de novo."
Cite as written: Dougherty v. City of Covina, 654 F.3d 892, 897 (9th Cir. 2011)
Supporting span (verbatim, S-03 at 897): "We review de novo the district
court's grant of a motion to dismiss under Rule 12(b)(6)."

## Reviewer sign-off

Citator run by: ______________________ Date: __________
Spans spot-checked (sample of 10): ______________________ Date: __________
Unresolved rows cleared by: ______________________ Date: __________
```

## Watch-outs

- **The skill certifies from memory.** A model asked whether a citation is good will answer from recognition, confidently, and recognition is the exact mechanism behind fabricated authority. Guard: `SUPPORTED` requires a verbatim span copied from a file listed in `source_index` and printed in the row. No span, no upgrade — and the log prints spans rather than verdicts so a reviewer can check the guard held.
- **Circular sources.** A "source" that is itself an AI-generated summary of a case verifies a fabrication against its own fabrication. Guard: `provenance` is a required column with a closed vocabulary — `westlaw | lexis | courtlistener | govinfo | pacer | production | reporter_pdf` — and any row outside it is rejected at pass one rather than flagged at pass three. A summary written by an assistant is not a source.
- **Dedupe hides misrepresentation.** Guard: rows are keyed to citation instances, and the log prints instance count beside distinct-authority count so an equal pair is visible as a parse failure.
- **The log gets treated as the certification.** Guard: the scope statement is the first section of every log, it names existence and subsequent history as out of scope, and the signature block stays unsigned until a person signs it. The log is evidence of the inquiry, not the inquiry.
- **Stale pulls.** Guard: `retrieved_on` is required per source, compared against `filing_date`, and any gap over `staleness_days` draws `[STALE-PULL]` in the row rather than a silent pass.
- **Confidentiality.** Guard: the run happens under whatever data-retention terms the firm has accepted, checked before the first sealed document is pasted, per ABA Formal Opinion 512. The skill has no view on what your protective order permits.

# Source index

The only place the skill looks for supporting text. A citation whose authority is not listed here comes out as `NO-SOURCE` — it is not checked, and the log says so.

Fill one row per supplied source. Replace the sample rows; they are there to show the shape and the rejection rules, not to be kept.

## Header

```yaml
matter: "Ramirez v. Vantage Logistics, No. 3:26-cv-01184 (N.D. Cal.)"
draft: "motion-to-dismiss-d4.docx"
filing_date: "2026-09-04"
citator_run_on: "2026-08-18"
citator_tool: "KeyCite"
jurisdiction: "N.D. Cal."
sources_dir: "./sources/"
staleness_days: 30
```

`citator_run_on` and `citator_tool` are printed in the log header verbatim. Leave them blank if no citator was run — the log will then state that no citator run was recorded, which is the state a reviewing partner needs to see rather than a silent gap.

## Rows

| source_id | authority_or_record_id | type | provenance | retrieved_on | file | pages_included | notes |
|---|---|---|---|---|---|---|---|
| S-01 | Ashcroft v. Iqbal, 556 U.S. 662 (2009) | case | westlaw | 2026-08-17 | sources/S-01-iqbal.txt | 674-687 | excerpt, majority only |
| S-02 | 15 U.S.C. § 1681e(b) | statute | govinfo | 2026-08-17 | sources/S-02-fcra-1681e.txt | full | current through P.L. 119-xx |
| S-03 | Dougherty v. City of Covina, 654 F.3d 892 (9th Cir. 2011) | case | courtlistener | 2026-08-17 | sources/S-03-dougherty.txt | full | — |
| S-04 | RAMIREZ_0004412-0004418 | record | production | 2026-08-11 | sources/S-04-onboarding-packet.txt | full | Bates range as produced |
| S-05 | Ramirez Dep. 118:4-121:19 (2026-07-22) | transcript | production | 2026-08-11 | sources/S-05-ramirez-dep.txt | 112-126 | rough transcript |

### Column rules

- **`source_id`** — stable, referenced from every log row. Do not renumber between runs on the same draft; a reviewer comparing draft 4 to draft 5 needs the IDs to hold.
- **`authority_or_record_id`** — the citation in full form for authorities, the Bates range or transcript page:line for record material. Copy it from the source, not from the draft. Copying it from the draft imports the draft's error into the thing that is supposed to catch the draft's error.
- **`type`** — `case | statute | regulation | rule | record | transcript | secondary`.
- **`provenance`** — closed vocabulary, and the skill rejects anything outside it at pass one:

  ```
  westlaw | lexis | courtlistener | govinfo | pacer | production | reporter_pdf
  ```

  There is no `other`, and there is deliberately no value for a summary or a synopsis. A source generated or paraphrased by an assistant verifies a fabrication against itself, so the index has no way to express one. If the text you have is a headnote or an editorial summary rather than the opinion, do not index it — pull the opinion.
- **`retrieved_on`** — the date this text was pulled, not the date of the decision. Compared against `filing_date`; a gap over `staleness_days` draws `[STALE-PULL]` on every row citing this source.
- **`file`** — path relative to `sources_dir`. Plain text or Markdown. PDFs need to be converted first; the skill compares characters and cannot do that against a scanned image.
- **`pages_included`** — the reporter pages, paragraph numbers, or Bates range actually present in the file. Load-bearing: when a pin cite falls outside this range, the row comes out `NO-SOURCE` rather than `UNSUPPORTED`, because the absence of support has not been established. Getting this column wrong is the fastest way to turn a real problem into a clean-looking log.
- **`notes`** — free text. Note excerpts, rough transcripts, dissent-only pulls, and translations.

## Excerpting

Excerpts are preferred over full text, and not only for cost. Pull the cited page plus two pages on either side. A five-page window around a pin cite is enough to find the supporting span or to establish that it is not there, and it keeps the run inside a standard context window.

Full-text pulls across a 40-authority brief run past 600,000 tokens, which does not fit a standard context window — the 1M-token window is available through the Anthropic API, not in the Claude apps — so a full-text run has to be chunked or split across sessions. That is a worse check as well as a more expensive one: chunking breaks the guarantee that pass two saw every supplied source before assigning a status.

Two cases where the wider pull is right:

- The proposition is about what a court **held** rather than what it said in passing. Holdings need the procedural posture, and the posture is rarely on the pin-cited page.
- The draft's cite carries a `see generally` signal or points at an opinion as a whole. There is no pin cite to build a window around.

## Before the run

1. Every authority cited in the draft has a row here, or you accept a `NO-SOURCE` row for it in the log.
2. Every `file` path resolves and opens as text.
3. Every `retrieved_on` is a real pull date, not the date you filled the spreadsheet.
4. `pages_included` covers the pin cites in the draft. Check this against the draft's table of authorities before running, not after.
5. No row's text came from an AI summary, a headnote, a brief bank, or another party's brief.

# Verification log template

The output artifact. It records the inquiry: what was checked, against what text, with what result, and what was not checked at all. Keep it with the matter file.

Adapt the header fields and the sign-off block to your firm's practice. Do not delete the scope statement — it is the section that stops the log from being read as a certification.

---

# Verification log — {{DRAFT_TITLE}}

Draft: `{{DRAFT_FILE}}` | Jurisdiction: {{JURISDICTION}}
Filing date: {{FILING_DATE}} | Run date: {{RUN_DATE}}
Citator run recorded: {{CITATOR_RUN_ON}} ({{CITATOR_TOOL}})
Citation instances: {{N_INSTANCES}} | Distinct authorities: {{N_AUTHORITIES}} | Sources supplied: {{N_SOURCES}}

## Scope of this check

This log records whether each cited source, **as supplied**, states the proposition the draft attributes to it.

It does **not** establish that any authority exists, that it remains good law, or that its subsequent history is clean. Those are citator findings, recorded above by date only. {{N_NO_SOURCE}} authorities in this draft had no supplied source and were not checked.

The signature under Rule 11(b) and the inquiry reasonable under the circumstances that the signature certifies remain with counsel. This log is evidence that an inquiry happened. It is not the inquiry, and it is not the certification.

## Summary

| Status | Rows |
|---|---|
| SUPPORTED | {{N_SUPPORTED}} |
| PIN-MISMATCH | {{N_PIN}} |
| QUOTE-DRIFT | {{N_QUOTE}} |
| SIGNAL-MISMATCH | {{N_SIGNAL}} |
| UNSUPPORTED | {{N_UNSUPPORTED}} |
| NO-SOURCE | {{N_NO_SOURCE}} |

Annotations present: {{STALE_PULL_COUNT}} `[STALE-PULL]`.

## Unresolved

Everything that is not `SUPPORTED`, most severe first: `UNSUPPORTED`, then `QUOTE-DRIFT`, then `PIN-MISMATCH`, then `SIGNAL-MISMATCH`, then `NO-SOURCE`. Read this section before the supported rows; it is the only part with work in it.

### [{{INSTANCE_ID}}] {{STATUS}}

Proposition: "{{PROPOSITION_AS_DRAFTED}}"
Cite as written: {{CITE_AS_WRITTEN}}
Source: {{SOURCE_ID}} | provenance: {{PROVENANCE}} | retrieved_on: {{RETRIEVED_ON}}
Finding: {{WHAT_THE_SOURCE_DOES_AND_DOES_NOT_SAY}}
Action: {{RE_PULL | RE_CITE | STRIKE | RE_SIGNAL | ATTORNEY_CALL}}

For `QUOTE-DRIFT`, include both texts and the diff:

```
Draft:  "{{QUOTED_SPAN_FROM_DRAFT}}"
Source: "{{SPAN_FROM_SOURCE}}"
Diff:   {{CHARACTER_LEVEL_DIFFERENCE}}
```

## Supported

One row each. The verbatim span is the point of the row — a status with no span next to it is the failure this log exists to make visible.

### [{{INSTANCE_ID}}] SUPPORTED

Proposition: "{{PROPOSITION_AS_DRAFTED}}"
Cite as written: {{CITE_AS_WRITTEN}}
Supporting span (verbatim, {{SOURCE_ID}} at {{LOCATOR}}): "{{SPAN}}"

## Alterations and omissions

Quoted material the draft changed on purpose. Listed as a set so a reviewer reads them together rather than meeting them one at a time between diffs.

| Instance | Device | Draft text | What was removed or changed |
|---|---|---|---|
| {{ID}} | ellipsis | {{TEXT}} | {{WHAT_WAS_DROPPED}} |
| {{ID}} | bracket alteration | {{TEXT}} | {{ORIGINAL_WORD}} |
| {{ID}} | (cleaned up) | {{TEXT}} | {{WHAT_THE_PARENTHETICAL_COVERS}} |

An ellipsis that removes a limiting clause changes the proposition. Check every row in this table against the surrounding source text, not against the quoted fragment.

## Not checked

| Authority | Why | Who owns the follow-up |
|---|---|---|
| {{CITE}} | no supplied source | {{NAME}} |
| {{CITE}} | pin cite outside `pages_included` | {{NAME}} |
| {{CITE}} | source is a scanned image, not text | {{NAME}} |

An empty table here means every cited authority had text behind it. That is the state to file in.

## Reviewer sign-off

| Step | Who | Date |
|---|---|---|
| Citator run over every authority | ______________ | ________ |
| Sample of 10 supporting spans opened and confirmed in the source | ______________ | ________ |
| Every unresolved row cleared, re-cited, or struck | ______________ | ________ |
| Alterations table reviewed against surrounding text | ______________ | ________ |
| Draft approved for filing | ______________ | ________ |

The sample check on the third row is not ceremonial. It is the control that tests whether the no-span-no-upgrade rule held on this run, and it is the only step here that a machine cannot perform on its own output.

# Status vocabulary, signal rules, and quote fidelity

The rules pass three applies. Edit the signal table for your jurisdiction and the quote rules for your style manual; leave the promotion rules alone unless you are prepared to explain the change to a reviewing partner.

## The six statuses

| Status | Means | Requires | Who resolves it |
|---|---|---|---|
| `SUPPORTED` | A verbatim span in the supplied source states the proposition, at or within the pin cite | The span, copied into the row, with its locator | Nobody — but 10 rows get spot-checked |
| `PIN-MISMATCH` | The source supports the proposition, at a different page than the pin cite | The span plus both locators | Drafter corrects the pin cite |
| `QUOTE-DRIFT` | Quoted text in the draft does not match the source character for character | Both texts plus the diff | Drafter re-quotes or removes the quotation marks |
| `SIGNAL-MISMATCH` | The support relationship does not match the introductory signal | The span plus the signal actually warranted | Drafter changes the signal |
| `UNSUPPORTED` | Source supplied, pin cite inside `pages_included`, no span supports the proposition | A statement of what the source does say | Attorney call: re-pull, re-cite, or strike |
| `NO-SOURCE` | No source text supplied, or the pin cite falls outside `pages_included` | The reason | Pull the source and re-run |

### Promotion rules

1. **No span, no `SUPPORTED`.** A row reaches `SUPPORTED` only when a verbatim span copied from a file listed in the source index sits in the row. Recognition is not verification, and a model that "knows" the case is producing exactly the signal that fabrication produces.
2. **`NO-SOURCE` never becomes `UNSUPPORTED`.** Absence of supplied text is not absence of support. Collapsing the two turns a gap in the pull into a finding against the drafter, and reviewers stop trusting the log.
3. **`UNSUPPORTED` never becomes `SUPPORTED` on a second look.** If the first pass found no span and a later pass finds one, the finding is that `pages_included` was wrong. Fix the index and re-run the whole draft; do not patch the row.
4. **A row can carry one status and any number of annotations.** `[STALE-PULL]` and `[PRIVILEGE-CHECK]` are annotations, not statuses.
5. **Record cites follow the same rules as authorities.** A Bates range that does not contain the asserted fact is `UNSUPPORTED`, and a produced document nobody indexed is `NO-SOURCE`.

## Signal matching

Compare the relationship the span actually has to the proposition against the signal the draft used. Bluebook conventions; adjust for local rules and for jurisdictions that use a different citation manual.

| Signal | The source must | Common mismatch |
|---|---|---|
| *(none)* | State the proposition directly, or identify the source of a quotation | Used where the proposition follows only by inference — wants `see` |
| `see` | Support the proposition by an inferential step the reader makes without help | Used where the source states the proposition outright — drop the signal |
| `see also` | Provide additional support beyond authority already cited | Used as the first and only cite for a proposition |
| `cf.` | Support a different proposition analogous enough to lend support | Used as a synonym for `see`, which it is not |
| `compare … with …` | Both sources present, and the comparison stated in a parenthetical | Parenthetical missing, so the reader has to reconstruct the point |
| `but see` | Contradict the proposition | Used where the source merely distinguishes on facts |
| `see generally` | Provide background; no pin cite expected | Used to avoid finding a pin cite for a specific proposition |
| `e.g.` | Be one of several sources stating the proposition | Used where only one source exists |

Signal errors are not sanctionable on their own. They are worth flagging because they cluster with the misrepresentation class — a proposition that needed a `see` and got a bare cite is one where the drafter's reading and the source's text have already drifted apart.

## Quote fidelity

The comparison is character-level. Semantic comparison rates a paraphrase inside quotation marks as a match, and that paraphrase is the most common false-quote pattern in the sanctions record.

Exact match required on: every character between the quotation marks, internal punctuation, capitalization at the start of the quoted span unless bracketed, and any citation embedded in the quoted text.

Flagged as `QUOTE-DRIFT` and shown as a diff:

- Any word substituted, dropped, or added without brackets or an ellipsis
- Case changed at the start of the span without a bracketed letter
- Internal citations silently deleted with no `(citation omitted)` parenthetical
- A quotation reproduced from a second source's quotation of the first, where the intermediate source altered it

Enumerated in the alterations table rather than flagged as drift — these change the text on purpose:

| Device | What to check |
|---|---|
| Ellipsis | Whether the removed text was a limiting clause. "The rule applies … in every case" is a different sentence from the original if the removed span was "absent exigent circumstances" |
| Bracketed alteration | Whether the bracket changes the grammatical subject or the tense in a way that changes who the sentence is about |
| `(cleaned up)` | What the parenthetical is standing in for. The convention covers internal quotation marks, citations, brackets, and ellipses — it does not cover substantive edits, and it does not license paraphrase |
| `(emphasis added)` | That the emphasis is in the draft and not in the original, and that the reverse case carries `(emphasis in original)` |
| `(citation omitted)` | That what was removed was a citation and not text |

## Reading the output

Read the unresolved section first, and read it in this order: `UNSUPPORTED`, `QUOTE-DRIFT`, `PIN-MISMATCH`, `SIGNAL-MISMATCH`, `NO-SOURCE`.

Two counts are worth watching across runs rather than within one:

- **Unresolved rate at first pass** — `NO-SOURCE` plus `UNSUPPORTED` over total instances. A rate that stays high after the pull discipline settles is a source-index problem, not a drafting problem, and the fix is in `pages_included`.
- **Instances per distinct authority** — under 1.3 on a brief of any length means the parse is deduplicating when it should not be. Re-read pass one before trusting the run.