From raw extraction to trusted knowledge — the human-in-the-loop pipeline that turns staged content into production-quality graph data.
Builds on: Phase 1 — Ingestion & Pipelines
Phase 1 produces staged content with confidence scores. Phase 2 processes that staging queue at scale.
Three-Tier Confidence Routing
Every staged extraction is automatically routed based on confidence: auto-accept (above threshold), steward review (middle band), or auto-reject (below floor). The thresholds are calibrated from Phase 1's quality baselines — per topic, per source.
Domain Steward Curation Queue
Domain Stewards get a full curation queue — not just quality sampling. They approve, reject, or edit individual extractions, with full source evidence and confidence context. Approve-once patterns propagate to similar future extractions.
Entity Deduplication
Cross-source entity resolution at graph scale. When the same entity appears from multiple sources with different names, identifiers, or attributes, the system detects candidates and presents them for steward-assisted merge or split.
Conflict Resolution
When two sources disagree about the same entity — different addresses, different ownership, different status — the conflict is surfaced with both source evidence chains. The Domain Steward or Knowledge Engineer resolves with full audit trail.
Approve Once, Govern Millions
Pattern-based approval scaling. When a steward approves a type of extraction (e.g., "decision rules from BRMS exports"), the system learns the pattern and auto-applies it to similar future extractions — turning one human judgment into thousands of governed decisions.
Staging-to-Production Promotion
Reviewed and approved content is promoted from the staging graph to production named graphs with full provenance — who reviewed, when, what evidence supported the decision. Production content is immutable once promoted.