Docs / Concepts
Evidence
Evidence is the ordered protocol record that makes accepted, denied, failed, and replayed invocations inspectable.
Plain English
Evidence is the structured record of what happened at the capability boundary.
Why it exists
Logs are not enough when independent agents, hosts, applications, and infrastructure providers need to reconstruct outcomes.
Formal definition
Evidence is ordered invocation metadata emitted by a host and queryable by correlation ID for audit, debugging, verification, and export.
Concrete example
Ground the concept before the schema.
A denied schedule_technician invocation emits execution_denied with capability ID, version, host, sequence, and correlation ID.
Finds an available qualified technician and reserves a service window.
Invocation trace
- 01 Actor
Planning Agent
- 02 Capability
schedule_technician
- 03 Host
ServiceOpsHost
- 04 Policy
manager_approval
- 05 Context
job_context
- 06 Result
Confirmed Appointment
{ "event_id": "evt_8f3a1c", "event_type": "execution_denied", "invocation_id": "inv_session_abc_001", "capability_id": "schedule_technician", "capability_version": "1.0.0", "host_id": "service-ops-host", "correlation": { "correlation_id": "session-abc" }, "sequence": 2, "timestamp": "2026-06-16T15:14:22.104Z", "outcome": "denied", "payload": { "policy": "manager_approval" }, "redacted": true, "assurance": { "level": "S1" }}Relationships
Where this sits in the protocol.
Each concept should explain its neighbors so implementation teams can preserve the boundary across manifests, invocation, evidence, and tests.
Evidence is emitted after validation, denial, failure, or execution.
Replay uses evidence to reconstruct the ordered protocol trail.
Conformance tests should prove evidence shape and ordering.
Visual model
- 01Invocation starts with a correlation ID.
- 02Host emits typed evidence events in order.
- 03Replay returns the evidence stream for inspection.
Implementation notes
- Use stable event types for started, completed, failed, denied, and replayed outcomes.
- Redact sensitive payload fields while preserving decision metadata.
- Make evidence queryable without requiring private host logs.
Common mistakes
- Calling arbitrary logs evidence.
- Recording only successful executions.
- Dropping sequence and correlation fields.
Related concepts