Acquire authoritative reference documents — statutes, regulations, official policies, terms of service, agency guidance — into the case folder with provenance and plaintext extraction. Triggers when the user names a specific law/regulation/policy/ToS the case will cite, asks "do we have a copy of <X>" or "where do I find the official text of <Y>", supplies a copy and asks "can we use this", or when packet-builder needs a reference appendix that isn't on disk.
Acquire authoritative reference documents — statutes, regulations, official policies, terms of service, agency guidance — into the case folder with provenance and plaintext extraction. Triggers when the user names a specific law/regulation/policy/ToS the case will cite, asks "do we have a copy of <X>" or "where do I find the official text of <Y>", supplies a copy and asks "can we use this", or when packet-builder needs a reference appendix that isn't on disk.
Three independent paths to the same authoritative text:
Path A — User-supplied copy. The user already has the doc. Ingest it, then assess what's visible (truncation flags, missing effective dates, watermarks). The user makes the call on whether to trust it.
Path B — Project-known trusted source. Look up the curated source directory in data/reference_sources.yaml for the doc's (kind, jurisdiction) pair (e.g., (statute, MD) → Maryland General Assembly).
Path C — Constrained web search. Allowlist-only fetch (*.gov, *.us, NAIC, courts.gov, Cornell LII, etc.) with explicit denylist (Wikipedia, paywalled aggregators, marketing legal sites).
When two or more paths produce a copy, report both side-by-side. Hash-equal copies are consolidated; differences get a compare report and the user decides. Never silently prefer one path over another.
When this skill fires
The user names a specific statute, regulation, policy, or ToS the case will cite.
The user says "I have a copy of the policy / TOS / statute, can we use it?"
The user asks "do we have a copy of <citation>?" or "where do I find the official text of <X>?"
packet-builder is about to compile a reference appendix whose source isn't already under references/.
authorities-finder or authorities-reconcile has just identified a regulator and the next step is pulling the rules they enforce.
citation (e.g., "Md. Code Ins. § 27-303", "15 USC § 45", "Acme TOS as of 2026-04-15")
jurisdiction (2-letter US state, "federal", or "*" for cross-jurisdiction docs like ToS)
title (optional but useful for the index)
If the user is vague ("the statute about car insurance claims in Maryland"), translate to a concrete cite before proceeding. Don't ingest something the user can't name.
Ask about Path A. "Do you already have a copy?" If yes:
Locate the file (the user may name a path or hand you one).
Run the ingester with --source-origin user-supplied:
uv run python -m scripts.references.ingest \
--file <path> \
--kind <kind> --citation "<citation>" --jurisdiction <juris> \
--source-label "<where they got it, in their words>"
Read the assessment block in the resulting sidecar. Report flags to the user verbatim with the canonical phrasing: "You make the call on whether this copy is good enough; I can flag what I see but I can't certify authority."
Look up Path B. Read data/reference_sources.yaml for (kind, jurisdiction). If a curated entry exists, present each source to the user with name, URL, and how: instructions. Then ask: "Want me to fetch this for you (recommended for first cross-check), or would you rather download it yourself and hand me the file?"
Manual-download path: print the URL and the how: text, wait for the user to drop the file into references/raw/ (or anywhere), then ingest it as in Path A but with --source-origin manual-download.
Optional Path C — constrained web search. When Path B is sparse or a cross-check is wanted, search via the same allowlist as authorities-web-research. Forbidden: Wikipedia (for cites), paywalled aggregators, AI answer panels, marketing-legal sites. The fetcher will refuse non-allowlisted hosts; if the user is confident in a specific host that isn't listed, you may pass --allow-unknownonly after explicit user confirmation and only for primary publishers (.gov, .us, official agency domains).
Hash-equal: byte-identical copies; just note that the two paths agreed.
Hash-differ but text-equal: the rendering differs (PDF vs HTML) but the substance is the same; safe.
Text-differs: show the diff. Likely causes: the user's copy is older, an excerpt, or from a non-authoritative source. Ask the user to read the diff and decide.
Re-hash the references tree. The ingester does this automatically (it refreshes <case>/.references-manifest.sha256 on each call). If the user manually moves files under references/, run:
uv run python -m scripts.references.ingest --help # confirm path
# the manifest also gets refreshed by every ingest; if you need a manual refresh:
# python -c "from scripts.references._manifest import refresh_sha256_manifest; from pathlib import Path; refresh_sha256_manifest(Path('references'), Path('.references-manifest.sha256'))"
Hand back to caller. Typically that's pat-workflow (continuing to Phase 4) or packet-builder (which can now reference the doc from packet-manifest.yaml > reference_appendices).
Path A — assessing user-supplied copies
The agent does not certify authority. It surfaces what's visible and the user decides. Verbatim phrasing:
"You make the call on whether this copy is good enough; I can flag what I see but I can't certify authority."
The ingester runs heuristics in scripts/references/assess.py that produce flags:
Code
Meaning
truncation-suspected
Text ends mid-sentence or without final punctuation.
short-for-kind
Text is below a typical floor for this kind (e.g., a statute under 400 chars).
has-truncation-marker
Contains literal [truncated], (continued), * * *, etc.
looks-like-excerpt
First 500 chars contain "EXCERPT", "PARTIAL", "SAMPLE".
has-watermark
Contains "DRAFT", "CONFIDENTIAL", etc.
no-effective-date
No "as of", "effective", "last updated", or version marker found.
no-section-numbers
For statute / regulation, no §, "Section N", or (a)(b) markers.
encoding-issues
Replacement characters / mojibake.
Report every warn-level flag to the user. info-level flags are background; mention them only if relevant.
If the assessment looks bad and the user wants the doc anyway, that's fine — note the flags in the sidecar and proceed. The user owns the trust call.
Path B — project-known trusted sources
data/reference_sources.yaml is the curated map of where to get authoritative text. Two top sections:
Allow/denylist — domain glob patterns with trust levels (primary, secondary-trusted, secondary-confirm). Used by the fetcher.
Source directory — keyed by (kind, jurisdiction), lists 1-N curated sources with name, url, how:, and trust:. Falls back to (kind, "*") when the specific jurisdiction isn't listed (e.g., ToS is jurisdiction-agnostic).
Present sources to the user verbatim — don't paraphrase the how: line. The point of the curated directory is that the exact path-to-text instructions have been written down.
If the directory has nothing for (kind, jurisdiction), say so plainly and offer Path C. Don't invent a source.
Path C — constrained web search
Same allow/deny policy as authorities-web-research. Trust hierarchy:
Primary: official publisher of the text (mgaleg.maryland.gov, ecfr.gov, govinfo.gov, the agency's own .gov site, the platform's own /legal page).
Secondary-trusted: well-known free legal-information providers (law.cornell.edu, courtlistener.com). Allowed; sidecar records the trust level so the user knows it isn't first-party.
Secondary-confirm: archives, mirrors, secondary republishers. Refused unless the user explicitly types --allow-unknown after you've explained why.
Forbidden by default: Wikipedia (as a cite), paywalled aggregators (westlaw.com, lexisnexis.com, casetext.com), AI answer panels, marketing legal sites (findlaw.com, justia.com, etc. — fine for orientation, never as a cite source).
Cross-source comparison
When ≥2 paths produced a copy, always compare. The compare tool emits markdown by default; drop it under notes/references/<date>_<slug>_compare.md so the dual-process result is preserved alongside the case-intake notes.
raw_sha256_equal: true — byte-identical bytes. Consolidate to one record (the duplicate ingest can be removed by hand if desired).
raw_sha256_equal: false, readable_text_equal: true — different containers (PDF vs HTML), same substance. Keep both; note in the user-facing summary.
readable_text_equal: false — the rendered text differs. Show the diff to the user. Likely causes: one is older, one is an excerpt, one is from a non-authoritative republisher. The user decides which to rely on.
Disclaimers (verbatim — do not paraphrase or strip)
Stamp into every sidecar JSON: "This is reference information, not legal advice."
On any cite extracted from a fetched doc and quoted in a draft: [VERIFY WITH COUNSEL].
For Path A user-supplied copies: "You make the call on whether this copy is good enough; I can flag what I see but I can't certify authority."
These strings are embedded by scripts/references/ingest.py automatically; do not strip or rephrase them when reporting to the user.
Definition of done
For each doc the user wants:
references/raw/<slug>.<ext> exists (byte-identical copy of the source).
references/readable/<slug>.txt exists (extracted plaintext; may be empty if the format has no extractor — that's a flagged outcome, not a failure).
references/structured/<slug>.json exists with full sidecar including disclaimer, assessment, extraction.method, source_origin, and (if fetched) fetch.host + fetch.trust.
references/.references-manifest.yaml has an entry for source_id.
<case>/.references-manifest.sha256 is refreshed.
If 2+ paths produced copies, a comparison report exists under notes/references/.
Hand back to the caller (pat-workflow or packet-builder).
Synthetic example (Maryland-Mustang)
Maryland-Mustang's complaint cites Md. Code Ins. § 27-303 (unfair claim settlement practices) and COMAR 31.15.07 (motor-vehicle total-loss claim regulations). Pulling them via Path B (project-known trusted source):
After ingest, the packet manifest can reference them:
reference_appendices:-name:"md-statutory-reference"title:"Maryland Statutes and Regulations Cited"sources:-"../references/raw/md-code-ins-27-303.html"-"../references/raw/comar-31-15-07.html"note:|
Public-domain Maryland statutes and regulations cited in the
complaint. Verify against the publisher's site before relying
on a specific provision.
Do not
Do not invent statutes, regulations, or citations that aren't in the user's facts, in data/authorities.yaml, or surfaced by an actual fetch. If the user can't name the cite, ask — don't guess.
Do not strip the disclaimer ("This is reference information, not legal advice.") from sidecars, summaries, or downstream drafts.
Do not auto-fetch from a non-allowlisted domain. The fetcher refuses by design; --allow-unknown is for after-the-fact user confirmation, not for batch convenience.
Do not collapse a dual-source report into "the answer is X." If the user asked for a cross-check, present both.
Do not treat a single source as authoritative when others were unreachable. Say "Path B was unavailable" plainly.
Do not mutate references/raw/ after ingest. If a doc needs replacing (e.g., the agency posted a new version), re-ingest it; the ingester writes a new slug or, with --force, overwrites the manifest entry while preserving sha256 history.
Do not use the case-folder evidence pre-commit hook as a reason to put reference docs under evidence/. They belong under references/ precisely because they are reproducible third-party text.
