Capture one answer shell.
Store application, environment, question/answer text or hashes, model-run commitment, policy flags, actor/session hashes, idempotency key, and status.
Why rawctx
Why rawctx?
When an AI answer is challenged, show the evidence trail.
rawctx is built for the moment after an agent answers: what did it say, which reviewed evidence did it use, which model run produced it, which source evidence supported it, and what changed afterward?
Answer evidence first
rawctx is the evidence layer around AI answers.
Catalogs, vector databases, and LLMOps tools each own part of the stack. rawctx owns the audit record that ties the final answer back to the reviewed evidence, model-run commitment, and later correction history.
Bind source evidence.
Attach source_refs for documents, registered audio/video evidence, answer segments, and purpose-bound media retrieval events.
Keep runtime lineage.
Link model provider metadata, config hashes, OpenTelemetry trace ids, conversation/message ids, and submitted trace bundles without replacing the agent runtime.
Seal the record state.
Expose correction, void, redaction, Merkle proof, KMS-signed STH, OpenTimestamps, Rekor witness receipt, and explicit proof status.
What rawctx adds
Evidence that survives review, export, and dispute.
The answer record stays reviewable even when text storage is hash-only, media downloads expire, or a correction arrives later.
Answer Audit is the system of record
Each completed answer gets an append-only audit shell with evidence refs, inference commitment, source refs, policy flags, and lifecycle events.
rawctx.log_answer(...)Inference commitment binds the model run
Provider-attested or weight-verified model refs, input/output hashes, and config hashes are bound to the same answer hash.
rawctx.build_inference_commitment(...)Registered media evidence
Audio and video evidence are registered in a private vault, cited by source_ref, and retrieved through purpose-bound short-lived URLs.
rawctx.register_media_evidence_asset(...)Trace bundles and source_refs
OTel GenAI bundles, external trace ids, model run ids, answer segments, and source refs become searchable evidence instead of scattered debugging artifacts.
rawctx.ingest_otel_trace_bundle(...)Trust proof states
Auditors can distinguish ANCHORED records from pending or local-only evidence candidates before making claims.
rawctx trust proof answer <id>Adjacent tools
rawctx does not replace the systems around it.
It gives them an evidence spine. Your catalog keeps metadata, your runtime executes business logic, your LLMOps stack observes model behavior, and rawctx binds the answer users saw to the model run and proof trail auditors need.
- Catalogs keep live metadata and ownership.
- Runtime systems execute definitions and SQL.
- LLMOps tools trace prompts, model calls, evals, and retrieval.
- rawctx records the answer evidence receipt and model-run commitment that connects them.
Pilot shape
Bring one answer that already needs evidence.
Start with one customer-facing AI workflow, one approved evidence source, and one model call path. The goal is to prove that your team can answer: what did the AI say, why, from which model run and evidence, and what changed afterward?