ADJ is the portable, append-only record format for agent deliberations — the substrate that lets calibration scoring, outcome tracking, and cross-deliberation learning actually close the loop.
The split between deliberation_closed and outcome_observed is the load-bearing design choice — it's what makes calibration scoring possible in the first place.
JSONL files, SQLite, Postgres, an event log — anything append-only and orderable qualifies. A directory of JSONL files is a valid Level 1 implementation.
Level 1 writers, Level 2 query providers, Level 3 scoring engines. Start where you are; upgrade when you're ready. Every level is interoperable with every other.
ADP produces deliberation events; ADJ stores them; outcomes arrive; ADJ computes calibration; ADP reads those scores to weight future votes. The loop closes.
deliberation_opened, proposal_emitted, round_event, deliberation_closed, outcome_observed — a shared envelope with optional hash-chain support.
Every entry type validated by the shipped v0.json schema. CI-ready before entries ever land on disk.
getCalibration, getDeliberation, getOutcome — the minimum read surface any Level 2+ implementation must serve.
Reference scoring algorithm turns recorded (confidence, outcome) pairs into per-agent calibration values the protocol can weight votes by.
Ed25519-signed per-domain snapshots served at /.well-known/adp-calibration.json — the primary cross-org trust mechanism.
@ai-manifests/adj-validate checks individual entries and full deliberation records including hash-chain integrity and outcome timing.
Pick a substrate. Write conformant entries. Serve the query contract. The reference libraries handle the scoring for you.
Browse reference libraries