Relex Research

Legal Research: You Discover, the Harness Grounds

A Relex draft may only cite law whose verbatim text sits in the global cache — the server's verifier rejects anything else. Your web results and memory can never ground a citation; they only find what to ground. So:

discover (you: web + public legal MCPs)
  → directive (POST /research/scrape — the harness caches VERBATIM text)
    → ground (the case agent cites the cached text; verifier enforces)

Epistemic discipline while discovering is relex-citations — in short: no docket numbers, margin numbers or quotes from model memory, ever.

Step 1 — know what the harness can fetch

execute GET /research/sources — the live registry. enabled: true sources can be fetched-and-cached by directive. Everything else is discovery-only: you research it yourself and pass the best official URL as sourceHint, or the harness falls back to its generic fetch ladder.

Step 2 — discover (your job)

Use your own WebSearch/WebFetch and, when connected, public legal MCPs. Prefer official/free sources; identify the authority in anchor form — topic, court/ bench, date if known, official source to verify in — never a memorized citation.

Per-jurisdiction method, citation schema and source whitelists live in the ../jurisdictions/<XX>.md packs when installed.

Compliance limits you must respect during discovery (encode in your plan): UK Find Case Law bulk/computational use needs a licence — read individual decisions, let the harness handle caching lawfully; Spain CENDOJ prohibits bulk reuse (read single decisions only); CanLII full text is site-only per its terms; AustLII discourages scraping. When a channel is restricted, discovery stays manual-and-targeted and the sourceHint you pass is the official page.

Step 3 — directive (the harness's job)

POST /research/scrape grounds one authority — params via search (jurisdiction, code, article, authorityType = statute | case_law | regulator, sourceHint?, caseId? for provenance; results are a shared cache).

• Cache-first: status: "cached" means it's already grounded — done.
• status: "enqueued" → poll GET /research/scrape/{jobId} until ingested, then trigger a case-agent re-reason turn so it locks the issue to the text.
• status: "cooldown" → a recent fetch failed; give a better sourceHint later.
• Targeted only: directives are for the citations a draft actually needs (the case's pending_citations / your audit's grounding gaps). Never bulk. There's a modest daily cap; case law behind rate-limited APIs (CourtListener free tier) makes every directive count.
• US case law is now harness-groundable: us_courtlistener is a provisioned source, so POST /research/scrape with authorityType: "case_law" fetches the opinion's verbatim text directly (no HTML-ladder fallback). Always re-check GET /research/sources for the live enabled set rather than assuming.
• Shared quota, not per-user: the CourtListener token is ONE credential for the whole deployment — every directive across every user draws from the same free-tier bucket (as of the May 2026 policy change: 5/min, 50/hr, 125/day). This is a stronger reason to keep directives targeted (above): a burst of bulk requests from one case can throttle every other case's case-law grounding for the rest of the day.
• For case law, put court in code and docket/ECLI/reporter cite in article (e.g. code: "Cass. 1re civ.", article: "21-12.345"), authorityType: "case_law", and the decision URL you found as sourceHint.

LOCUS — US local ordinances (two vantage points, one corpus)

• Grounding path (backend): execute GET /research/locus?state=CA&city=…&query=… → discovery results with coverage: matched / covered_no_match (broaden terms) / not_in_locus (fall back to web; NEVER invent an ordinance). Then GET /research/locus?locusId=… fetches the section's verbatim public-domain text for citation-grade grounding.
• Analysis path (yours, optional): the LOCUS dataset (huggingface.co/datasets/LocalLaws/LOCUS-v1, paper "Freeing the Law with LOCUS") also carries analytical layers (topic, function, enforcement- discretion, opacity, paternalism scores) under CC-BY-NC-4.0. You may read those yourself (HF datasets-server /rows) to shape strategy — keep that analysis on your side (it is non-commercial-licensed; the Relex backend never serves it), attribute the corpus, and still ground any citation through the backend path above.
• The case agent runs its own LOCUS search on local-law questions; read its coverage verdicts from the case before duplicating a search.

Anti-patterns

• Citing from your web page reads ("the court held… [link]") in a draft — the verifier will reject it; issue the directive and cite the cached text.
• Directive-spamming a jurisdiction "to have it all" — targeted grounding only.
• Treating a refused/blocked source as an error — restricted sources are discovery-only by design; route around with sourceHint.