Relex Ontology

The Ontology Collaboration Loop

The ontology is the case's understanding; the graph is only its data — and it can be wrong or incomplete. Your job (as counsel, see relex-counsel) is to audit that understanding, repair it, and direct the Relex harness to acquire what is missing. The harness's next reasoning turn automatically reads what you fixed.

read → audit → repair/enrich → direct acquisition → agent re-reasons → re-read → converge

Everything below is PII-safe by construction: people appear only as [PARTY_NAME_n] label tokens; fact entities carry keys, never values. The server's PII gate rejects raw identifiers you might try to write — don't.

1 · Read

• execute GET /ontology/case/{caseId} — the case graph: real-world legal objects (parties, obligations, clauses, statutes, events, ISSUES) and the verbs binding them (cites, concerns, supports, contradicts, undercuts, party_to, raises, established_by).
• execute GET /ontology/firm (?scope=org&id= for an org) — the practice's abstract concept graph (doctrines, clause types, argument patterns), grown automatically from knowledge uploads and concluded cases. No case instances, no PII. ANY org member can read the org graph; only curation is admin-gated. If the graph is sparse but the practice HAS indexed knowledge, backfill it: POST /ontology/firm/reconcile (self / org-admin) — call repeatedly until the response's remaining is 0 (each call processes up to 25 sources).
• Read the case itself too (GET /cases?caseId={caseId}&full=true — caseId is a query param; full=true is REQUIRED to get the timeline/phases, a plain GET /cases returns a bounded summary): timeline, phases, locked issues, drafts. The graph must MATCH the case data — mismatch is a finding.

2 · Audit — formal conflicts first, then the gap taxonomy

The system audits with you. Every ontology read carries conflicts[] — formal integrity findings the platform derives fresh from the graph on every read (never stored, so never stale): unresolved contradicts/undercuts edges, claims resting on ALLEGED facts, ungrounded claims, deadlines without date or trigger, obligations without obligor, open issues with no acquisition plan, duplicates, low-confidence structure, entities flagged stale, and a whole-graph staleness flag when case data changed after the last ontology update. The digest view and your steering block (steering.conflicts) carry the top findings; ?view=full (or the plain GET) has the complete list with resolution hints. The user sees the SAME list in the case UI.

Work them symbiotically: resolve each conflict via §3 ops (settle/contest the issue, add the missing link/date/party, merge duplicates, re-verify and clear a stale flag) — and contribute conflicts the formal checks cannot see: when YOU spot a semantic contradiction, add the contradicts/undercuts edge; when something presupposes outdated law, set properties.stale=true on it. The moment you write it, the system formalizes your finding and shows it to the user and the internal agent too. Never leave a conflict silently unaddressed: resolve it, acquire what resolves it, or escalate both sides to the human.

The gap taxonomy

Walk the graph against the case facts. For each gap type there is one correct acquisition verb — never fill a gap from model memory:

3 · Repair / enrich

execute POST /ontology/case/{caseId}/agent with a natural-language instruction; the editor turns it into typed, transactional, PII-gated ops (upsert, merge, link, retype, set issue status):

execute({ method: "POST", path: "/ontology/case/abc123/agent", body: { message:
  "Add issue 'limitation period' status=contested; link it as undercutting the
   payment claim; merge the duplicate 'Delivery Contract' entities; the
   [PARTY_NAME_2] node is the addressee institution — retype it authority." } })

Rules:

• Write only what case data or verified sources support — never model memory.
• ISSUES carry status open | contested | settled. Mark settled only when grounded (cached verbatim text or a user-confirmed fact) — say the ground in the instruction so it lands in the graph.
• After big case-state changes, POST /ontology/case/{caseId}/reconcile rebuilds the deterministic backbone (additive; your edits survive), then re-read.
• Curate reusable patterns (clause types, argument shapes — nothing case-specific) into the firm graph: POST /ontology/firm/agent.

4 · Direct acquisition

You discover; the harness fetches-and-caches (division of labour — relex-research has the full doctrine):

• GET /research/sources — what the harness can fetch verbatim vs discovery-only.
• POST /research/scrape {jurisdiction, code, article, authorityType, sourceHint?, caseId?} — cache one authority's verbatim text. Poll GET /research/scrape/{jobId}.
• POST /agent {type:"case_req", caseId, payload:{prompt}} — steer the case agent: name the issues to (re)enumerate, the knowledge to search, the parties that matter. It reads the updated ontology automatically and emits its own scrape needs for anything still missing. These are steering-session turns on your private steering branch (relex-steering has the protocol); the reply's steering.missing_data maps directly onto the §2 gap taxonomy — treat each item as a gap with its acquisition verb (or its browser deep link).

5 · Converge — and say so

The loop is done when:

1. every locked/settled issue is verbatim-grounded,
2. every open issue has an acquisition plan (directive issued, question asked, or escalated),
3. no unresolved contradicts/undercuts edge remains without a human flag,
4. addressee and deadlines are locked in the case data.

Record the state as a short Votum (interim vote) on the case via a case_req message — e.g. "Ontology audit: 6 issues grounded, 1 contested (limitation — directive pending job 4f2c), no open conflicts; addressee locked. Next: re-reason after ingest." Interim Vota land on the steering branch; capture the convergence state in the session's conclude summary (relex-steering) so it propagates to the main thread. Report progress to the user with counts and labels only — never names.

Anti-patterns

• Filling a graph gap from memory ("I know Art. 823 says…") — that's a grounding gap, not knowledge. Issue the directive.
• Re-doing the harness's work: you never scrape, never RAG-search private corpora yourself, never re-implement its reasoning directive.
• Overwriting agent-locked issues without evidence — repair with grounds, or contest with an undercuts edge and acquire.
• Big-bang rewrites of the graph. Small typed ops, then re-read.