SDKs & spans

The mappers in both SDKs understand the Vercel AI SDK's ai.* spans (recording the provider-call child, skipping the orchestration parent that would double-count every model call), OTel GenAI semconv — both attribute generations, since field names renamed between versions and an audit schema cannot drift with them — and OpenInference. Chains, retrievers and embeddings are engineering telemetry and are deliberately not recorded: the examiner's question is what the agent did.

model_call    tokens observed (never priced as $0), model, prompt/output hashes
tool_call     tool name, input/output hashes, outcome — success and failure alike
delegation    agent → agent handoffs (the compounded-autonomy record)
lifecycle     session open/close + heartbeats, so silence ≠ a dead emitter

Every event a turn emits — the model calls, the tool executions, and any mutation they lead to — shares one sessionId, and cross-plane corroboration is computed from it: a trace observed on two planes is evidence, on one plane it is an assertion, and the coverage report says which is which. The manual API exists for hosts with no OpenTelemetry at all:

handle.record("send_invoice", input={"invoice": 42}, output=result)

auditant.record("send_invoice", { input: { invoice: 42 }, output: result });

Evidence writes never block your agent: events queue and flush in batches, spill to a local buffer on outage (Python) or retain their batch bounded with counted drops (TypeScript), and the policy check is the only synchronous call — by design, since it runs before the action it governs. Serverless hosts must await auditant.flush() before returning.