---
name: contract-repo-audit
description: Sweep the whole contract drive, index every key term, and verify a random sample with a second pass. Use this skill when: One-off audit — Manual kickoff, long cloud run.
---

# Index every signed contract, verified

Goal: index every contract in the repository and verify a random sample before publishing.

Connected tools this skill uses: drive, googlesheets, slack.

## When to use

- One-off audit — Manual kickoff, long cloud run.
- Or when the user asks directly to index every contract in the repository and verify a random sample before publishing.

## Method

Audit the contract repository. This is a one-off run and may take hours in
the cloud if the repository is large.

1. Sweep {{drive_root}} for every contract file and parse each with the
   document skill into full text.
2. Extract key terms per contract: parties, effective date, term length,
   contract value, governing law, and renewal/notice terms.
3. Build the full index: one row per contract, all extracted fields, with a
   link to the source file.
4. Randomly select {{sample_pct}}% of contracts and re-extract their terms
   independently, with no access to the first pass's output. Compare the two
   extractions; mark any contract where the passes disagree as unverified
   rather than silently correcting it.
5. Publish the index view with every contract's status: verified-by-sample,
   unverified (not sampled), or disagreement-flagged.
6. Post a summary to {{delivery_channel}}: total contracts indexed, sample
   size, agreement rate, and the list of flagged contracts for manual review.

Never treat an unsampled contract's index entry as certified. The index
states its own confidence level per contract; only human review upgrades a
flagged entry.

## Key-term extraction

Pull the standard key terms from any contract in the repository, consistently across file types.

## Method
1. Take the parsed full text of a contract from the document-parsing skill.
2. Identify the parties, effective date, term length, contract value, governing law, and renewal/notice terms.
3. Extract each field as its literal text plus a normalized value, for example a date in ISO format.
4. Where a field is genuinely absent from the contract, record it as not-found, not a blank.
5. Link every extracted field back to the source page or clause it came from.
6. Add any org-specific fields beyond the default six to the same row.

## Rules
- Extract only what the contract text states; never infer a term from a template default.
- Not-found is a distinct value from an empty extraction error — never conflate the two.
- Fields beyond the default six that matter for this org's index (set on first run, then stored in memory).
- Every field must carry a source-page reference for the index to be auditable.

## Output
One row per contract: parties, effective date, term, value, governing law, renewal terms, plus a source-file link for each field.

## Document parsing

Bulk-read every Word and PDF contract across the drive into extractable full text.

## Method
1. Sweep {{drive_root}} and list every contract file, docx and PDF.
2. Convert each file to full text, preserving page numbers for later citation.
3. Skip and log any file that fails to parse, corrupted, password-protected, or image-only, rather than silently omitting it.
4. Batch the sweep so a large repository completes without manual chunking.
5. Hand each contract's text to the key-term-extraction skill.

## Rules
- A failed parse is logged and reported, never dropped silently from the index.
- Page-level references are preserved so every extracted field can be traced back.
- Drive folders in scope for the sweep (set on first run, then stored in memory).
- Long runs are expected — this is a one-off bulk sweep, not a per-file interactive job.

## Output
Full text per contract file, with page references, plus a parse-failure log for files the sweep could not read.

## Sample verification

Independently re-extract a random sample of contracts and certify the index's accuracy before publishing.

## Method
1. Randomly select {{sample_pct}}% of indexed contracts, without weighting toward any type.
2. Re-extract key terms for the sample with no access to the first pass's output.
3. Compare the two extractions field by field for each sampled contract.
4. Where both passes agree, mark the contract verified-by-sample.
5. Where they disagree, mark the contract disagreement-flagged rather than silently picking one value.
6. Compute the agreement rate across the sample and log it for trend tracking.

## Rules
- The re-extraction pass must have no visibility into the first pass's results.
- Disagreements are flagged, never resolved automatically by picking either pass.
- Default sample percentage (set on first run, then stored in memory; suggest 10%).
- A declining agreement rate across runs signals the extraction skill needs review, not the sample size.

## Output
A verification status per sampled contract, verified-by-sample or disagreement-flagged, plus an overall agreement-rate figure for the run summary.