Derived artifacts

Derived artifacts record what the AI decided (the generation receipt in RAG-style workflows) and the lineage that produced it. They must point to existing query artifacts and include hashes of the inputs used.

Envelope

• artifact_id – deterministic hash of canonicalized payload
• artifact_type – "derived"
• input_query_ids – required, sorted/deduped query artifact IDs
• input_hashes – required, sorted/deduped hashes of inputs
• policy_version – e.g. stub@0.0.0
• output_type – e.g. dr_recommendations
• output_data – structured decision payload
• Optional: assumptions, confidence (0–1), explanation
• created_at, created_by, correlation_id
• content_hash

Implementation: core/src/artifacts/derived.ts

Example (TypeScript)

import { Artifacts } from "core";
 
const derived = Artifacts.buildDerivedArtifact({
  input_query_ids: ["qry-123"],
  input_hashes: ["qry-hash-123"],
  policy_version: "stub@0.0.0",
  output_type: "dr_recommendations",
  output_data: { recommendations: ["DR-010", "DR-011"] },
  created_at: "2024-01-04T00:00:00Z",
  created_by: { actor_id: "tester", actor_type: "human" },
  correlation_id: "corr-derived",
});

Storage + validation

JSONL store enforces lineage: JsonlArtifactStore.putDerivedArtifact rejects writes unless every input_query_id already exists.

Reference test: core/tests/artifacts.test.ts

Notes

• Arrays are normalized before hashing to avoid nondeterministic IDs.
• Store is idempotent: re-appending the same artifact is a no-op.