LegalDocument Review

Proposition Checking

Use when users say "does this authority support the point", "check propositions", "fact-check this argument", "verify record support", or need cited cases, statutes, exhibits, transcripts, emails, or…

LLegalQuants·Legal·Apache-2.0

Library skill — the default version is maintained in GitHub; edits you make live in your own clone.

Use this skillDownload .zip
How does this work?
  • ChatGPT opens a new chat with the skill loaded. If it's too long for a link, it's copied to your clipboard — just paste.
  • Claude works the same way. To install it permanently, download the .zip and upload it under Claude → Settings → Capabilities → Skills (Pro/Team/Enterprise).
  • Copy prompt copies the skill so you can paste it into any assistant, including Grok.

proposition-checking

When to Use

This skill checks support. It does not predict who wins and does not replace legal judgment.

Audience and Work Shape

Audience: litigators, legal researchers, trainees, and paralegals checking draft arguments or evidence summaries for lawyer review.

Work shape: pattern-matched source checking with accretive judgment. The skill maps propositions to sources and grades support; it does not make the legal argument for the user.

Legal Failure Modes

Access Modes

This skill works in three modes:

  1. Live source mode - use browser, web search, MCP, API, or other configured access to retrieve authorities, statutes, rules, or public records.
  2. User-supplied source mode - use uploaded or pasted judgments, statutes, exhibits, pleadings, transcripts, correspondence, witness statements, or record bundles.
  3. No-source mode - extract propositions and citations, then prepare a verification queue. Do not classify support as supported, unsupported, or contradicted.

If the source cannot be retrieved or supplied, mark the proposition unverified-source-unavailable. Do not assess source support from model memory.

How It Works

1. Separate propositions from citations

For each paragraph, extract:

Do not treat a string citation as support until the proposition it supposedly supports is identified.

2. Retrieve or read the source

Use supplied source text first:

If the source cannot be retrieved, classify the proposition as unverified, not unsupported.

For every source used, record the source type, URL or document identifier, access date, and pinpoint. Do not assess support from model memory.

Confidence Bands

3. Compare proposition to source

Classify each proposition:

For legal propositions, distinguish:

For factual propositions, distinguish:

4. Report with traceability

Output:

Field Meaning
paragraph source paragraph or location
proposition claim being checked
citation authority or record cite
status support classification
source_location page, paragraph, Bates, section, exhibit, or other pinpoint
source_excerpt exact supporting or contradicting text
problem overstatement, mismatch, missing source, etc.
suggested_fix narrow, replace cite, add source, or remove

Use short excerpts. The goal is fast lawyer verification, not a replacement brief.

5. Identify dependency risk

After checking individual propositions, summarize:

Escalation

Stop and route to the responsible lawyer when a key proposition is contradicted, an authority appears fabricated, a quote is materially inaccurate, admissibility/privilege controls the answer, or the user asks for merits advice rather than source support.

Example

Check whether each cited authority in this draft actually supports the proposition it is cited for. Separate citation existence problems from misuse problems.

For a compact output pattern, see examples/output.md. For proposition types, support states, and argument dependency mapping, see references/proposition-checking-model.md.

Limitations