AgentTriage required runtime evidence before calling an agent action completed

A model log saying that a tool call was requested cannot establish that a file was modified or an API transaction completed. AgentTriage reconciles runtime traces and effect evidence from heterogeneous environments under an explicit diagnosis contract. primary evidence turns that distinction into a measurable auditing method.

The reported diagnosis and evidence-coverage figures are benchmark results; they do not guarantee that unavailable or altered production logs can always reconstruct the action.

Bind tool start, completion, retries, external acknowledgement and final state to stable IDs. If one required part is missing, mark the outcome unknown instead of converting the request into proof of effect.

The evidence chain has three separate identities: the invocation, its returned result, and any durable external effect. AgentTriage normalized 159 applicable executions across five runtimes and judged 120 cases. Its reported grounded joint diagnosis was 0.858, required-evidence coverage 0.966 and unsupported-claim rate 0.036. These are scored diagnoses against an adjudicated set, not evidence that every production tool effect can be recovered from missing or tampered telemetry. An auditor should mark the side effect indeterminate when it lacks an authoritative result or external receipt, even if the assistant has already claimed success.