# RegCorpus — evaluation sample pack

This is a real slice of the live RegCorpus corpus, exported through the same public API a customer receives. Every record here can be checked against [regcorpus.com](https://regcorpus.com) and against the official state source it came from.

## What RegCorpus is

A versioned library of US state insurance regulation: department bulletins, adopted regulations, and enforcement actions, collected straight from state insurance departments, refreshed daily, with every version kept and hashed.

Three things make it different from a scrape:

1. **Version history.** Every document is stored as a chain of content-hashed versions with the date we fetched each one. History cannot be backfilled later — it only accumulates from the day collection starts.
2. **Provenance on every record.** Each document carries its official source URL, the date it was first seen, the date it was last checked, and a SHA-256 of the exact bytes.
3. **Clean ingestion rights.** The material is public law and public records. State-adopted regulation text is uncopyrightable under *Georgia v. Public.Resource.Org* (2020). We license it for ingestion into models and products — see LICENSE-SUMMARY.md.

## Files

| File | What's in it |
|---|---|
| `sample-documents.ndjson` | Up to 12 documents per state (regulations and bulletins), each with its full version history and full text where rights permit |
| `sample-enforcement.ndjson` | 200 enforcement records: respondent, case number, order date, official link |
| `sample-daily-changes.ndjson` | The 100 most recent corpus changes — the shape of the daily delta feed |
| `manifest.json` | Corpus totals, per-state coverage, and generation timestamp |

One JSON object per line. Field names are stable.

## Reading a document record

```json
{
  "slug": "md-comar-31-02-03",
  "jurisdiction": "MD",
  "agency": "MIA",
  "doc_type": "regulation",
  "citation": "COMAR 31.02.03",
  "title": "Chapter 03 Hearings Arising from Cancellation, Nonrenewal...",
  "official_source_url": "https://regs.maryland.gov/us/md/exec/comar/31.02.03/",
  "access_class": "open-license",
  "access_notes": "Official bulk channel; CC0 public-domain text",
  "first_seen_at": "...", "last_checked_at": "...",
  "versions": [
    { "version_no": 1, "fetched_at": "...", "effective_date": "...",
      "content_sha256_short": "...", "text_length": 2371 }
  ],
  "text": "...",
  "text_length": 2371
}
```

**`text` is null on some records, and that is deliberate.** Where a source's own terms cap reuse, we store no text at all and serve facts plus the official link instead. New York's circular letters are the clearest example: the department's site grants reuse for non-commercial purposes only, so we hold the citation, date, subject, and official URL, and nothing more, while we ask for written permission. You are never handed text we do not have the right to hand you. `access_class` and `access_notes` tell you the posture for every record.

Some records have `text: null` for a duller reason: the source document is a scanned image with no text layer. Those are queued for OCR. `text_length: 0` with an `access_class` that permits text means exactly that.

## Coverage

See `manifest.json` for the live numbers. Coverage grows on a schedule: the collectors run daily, and new states are added as their source rights are cleared. Each state page on regcorpus.com lists its sources, the official URL for each, and how we are permitted to show it.

## What full delivery looks like

The sample is a snapshot. Delivery is a bulk export of the full corpus plus a daily delta feed of what changed, over REST, with a read-only MCP server for agent-native consumption. Format and cadence flex to what your pipeline wants.

## Verify us

Pick any record. Open its `official_source_url`. Compare the text. Then open the same document on regcorpus.com and check the version history and the hash. Nothing here requires trust.

Questions: venu@aijourneymates.com