Observability & Auditability
Other tools show you what happened. TCP/AP shows you why it was admissible.
Every verdict binds content hashes (the Agentic Protocol, the Artifact, the generation metadata) and a trace id, and, when signed, a Conformance Event that re-verifies offline against the root.
Conventional observability gives you the calls, the logs, the traces. TCP/AP gives you the exact policy, artifact, and outcome a decision was evaluated under, cryptographically bound. Governance becomes verification, not trust.
The trust anchor
Trust is anchored by the Cognition Trust Layer, the root of trust for the protocol. Every conformant verdict is signed as a Conformance Event, and the signatures chain back to it: the event is signed by an issuing key, the issuing key is authorized by a certificate, and the certificate is signed by the pinned root.
Clients verify but MUST NOT sign: a client that could sign could forge verdicts. Verification is offline, with no call back to the Kernel.
{
"trace_id": "trc_4a1f8b2c",
"agentic_protocol_hash": "sha256:81b0…",
"artifact_hash": "sha256:9de2…",
"conformant": true, "label": "AUTO_APPROVE"
} # what + why
Conformance Event → signed by issuing key → authorized by certificate → signed by Cognition Trust Layer # verified offline
The decision stream
Because every decision is a typed, self-classifying event, the runtime is a live stream of reasons, not just a record of calls. Each line is one governed decision: the model proposes, the Kernel renders a verdict with a status code, and the Operator routes on it. A 422 drift halts the action and hands off to a human, in the same shape as a 200 that proceeds.
Every line is a classified decision with a code and a reason, never a bare success or failure.