bajan v0.3.1
Spec-driven knowledge-graph pipeline CLI — ingest sources, extract claims, query the graph. Deterministic-first, spec-gated, human-approved.
What it does
- Ingest normalized episode streams — persisted verbatim, never rewritten, machine-readable outcome per record.
- Extract candidate claims — deterministic extractor by default; LLM passes only propose, humans approve.
- Query a ranked, lineage-traceable claim graph with bounded traversals and tag-scoped reads.
- Review entity-merge and contradiction candidates — the only human-in-the-loop surface; nothing adopts automatically.
- Maintain — tombstone retired episodes
(
bajan delete) and revive them (bajan revive); deleted ids reject re-ingest with a machine-readable reason until revived. - Evaluate extraction runs against a human-adjudicated golden corpus — deterministic precision/recall, hedge fidelity as a first-class number.
Specs are the code
Seven corpus specs (ingestion-contract, extraction-claims,
graph-model, query-tools, entity-review,
eval-claims, contradiction-review) deploy into a
drift-gated mirror, with 68 scenario-scoped property contracts executing
against the test suite — every commit is gated on
ah check --run-tests staying green.
Quickstart
# ingest a normalized episode stream (JSON array)
bajan ingest < episodes.json
# extract candidate claims (deterministic by default)
bajan extract
# adopt staged claims (human-in-the-loop)
bajan adopt 1 2 3
# query the graph
bajan query "profit" --scope workspace:dev
Links
- Repository — source, issues, full README
- CHANGELOG — release notes (v0.3.1: CI repair + idempotent releases; v0.3.0: shape rejection + evidence checks + coverage reads)
- Specs — the deployed spec mirror the code is gated against