Source attribution per Fact, QA pipeline for imports, confidence scoring for AI-extracted Facts, conflict resolution when sources disagree.
Why this matters
OLN’s Fact graph is the substrate for HellaThis (G-029) and the source of AEO/AI licensing value. Data quality is no longer just internal hygiene — it’s a trust layer for a consumer product with revenue at stake.
Needs explicit treatment of:
- Source attribution (which Fact came from where)
- QA pipeline (Commons proposal model handles new entries; what handles imports?)
- Confidence scoring for AI-extracted Facts (Memory Alpha migration)
- Conflict resolution when sources disagree
First controls shipped
Several pieces of this now exist:
- Type-trust extraction gate (Entry 038) — extraction is fail-closed on entity-type confidence: facts are only auto-extracted when independent signals (Wikidata type + source infobox) corroborate the type; ambiguous or conflicting cases wait for human validation. This is the front-line confidence scoring.
- Attribute QA pass — a sweep that flags real-but-unregistered and erroneous fields per type, feeding a propose → approve → registry loop (Entry 035).
- Source Registry (G-066) — each source is a first-class entity carrying its license class, citation tier (Entry 020), and reliability — the spine of source attribution.
Still open: conflict resolution when sources disagree (the review-queue duels, G-060), and propagation/versioning semantics at scale.
Related
- Entry 008 — HellaThis as Sister Product
- Entry 009 — Memory Alpha Migration
- Entry 030 — Discoverability posture (the structured graph is the licensable asset)
- Entry 038 — Type-trust extraction gate (the fail-closed confidence control)
- G-029 — HellaThis ↔ OLN intercompany licensing
- G-030 — Memory Alpha migration architecture
- G-060 — Unified Fact review & confirmation queue (conflict resolution)
- G-066 — Community source nomination & the Source Registry
- G-058 — AI crawler & data-licensing posture (gates external access to this graph)