Have extraction populate the accountability spine (owner / reports_to / members / decision_rights)

task-extraction-accountability-spine-rba

task confidence inferred status done 2026-06-19 owner extraction-engineer
source log-auditor — surfaced from the FDE QA pass on the 33-page RBA Firecrawl tenant (DEC-0055); the 0/494 owner count is from the read-only @dossier/okf-parsed audit; closed from the reference-tenant QA pass (tenant commit `8229530`)

Have extraction populate the accountability spine

The RBA Firecrawl run (First live FirecrawlConnector run against a real client source — field evidence for the reserved web seam) emitted 494 atoms with provenance and security that were excellent — but the accountability layer is entirely absent:

  • owner in 0/494 atoms.
  • reports_to / members / decision_rights also 0.
  • 18 role atoms own nothing; 95 process atoms have no owner.

The Dossier — The Knowledge Model (v0) makes accountability first-class precisely because it is the IP that walks out the door: role carries responsibilities/decision_rights/reports_to/members; process carries an owner; the owner/owned_by edge is core vocabulary. The extractor is capturing the facts (roles and processes exist) but not the spine that says who is accountable for what — so the "who's accountable" graph layer the model is built around is missing.

Shape

Teach the extraction prompt/schema to populate owner edges wherever the source grounds them — a service/process page that names its owning team or lead, an org/leadership/about page that grounds reports_to / members. The hard constraint: never fabricate accountability. An owner edge is emitted only where the source states it; an ungrounded one stays absent. This is the provenance discipline DEC-0001 and knowledge-model principle 8 require — wrongly-attributed accountability is worse than none.

Why a task, not a fix-in-place

This is an extraction-capability change (prompt + schema + grounding discipline) owned by the Knowledge-Extraction & GraphRAG Engineer, measured by source-grounded coverage — not a hygiene edit. Scoped to packages/extraction (durable) + the RBA tenant OKF (clients/rba/tenants-firecrawl/rba-consulting, a gitignored sandbox per Fix git-per-tenant isolation when a tenant root is nested inside another repo) for the re-extraction. Filed by the log-auditor from the QA pass; confidence: inferred.

Resolution (2026-06-19, tenant commit 8229530)

DONE via grounded data surgery (no LLM re-extraction). Closed backlog → done:

  • owner populated on 72/103 processes, discipline-matched to 9 RBA delivery roles. 31 cross-disciplinary processes intentionally left unowned — not fabricated, because the source does not ground a single owning role (provenance discipline, Adopt OKF as Dossier's canonical knowledge format / knowledge-model principle 8). Partial + grounded > complete + fabricated.
  • delivered_by UNCHANGED at 8/187 — this is a structural taxonomy gap (capabilities modeled as relates_to practices, not the delivery workflows/processes the DXA delivered_by edge expects), not a surgery omission; grounding it needs LLM-grade inference. Carved out and tracked as the honest remaining blocker: Close the RBA delivery-modeling taxonomy gap — capability→delivery `delivered_by` grounding + two persona-role cleanups.
  • okf tests 170/170 green; every emitted accountability edge resolves to a real atom (0 new dangling). The discipline-matched (never blanket) owner grounding is the curation lesson recorded in Dossier — Decision & Audit Log under DEC-0056's frame.