Interpreting Agent Evidence

An agent result is a workflow record, not a claim that a model was correct. Interpret final content together with the authorized roles, lifecycle, provider calls, convergence decision, terminal status, and versioned trace.

flowchart LR
    request["input + task goal"] --> definition["PipelineDefinition"]
    definition --> calls["ordered role calls"]
    calls --> decisions["veto + convergence + termination"]
    decisions --> result["PipelineResult"]
    calls --> trace["RunTrace"]
    decisions --> trace
    result --> review["bounded workflow claim"]
    trace --> review

Read One Workflow Result

Review question	Evidence to inspect	What remains unproven
Which roles were allowed to act?	pipeline definition, resolved configuration, role order and fingerprint	whether the selected roles are sufficient
What did each role receive and return?	typed call record, input/output identity, provider/model metadata, failure	content truth or provider stability
Who controlled progression?	lifecycle transitions and controller decisions	that elapsed completion order defined authority
Why did work stop?	convergence, oscillation, veto, maximum-iteration, interruption and terminal records	that stopping implies correctness
What does the result classify?	status, decision, epistemic verdict, confidence, stop reason and issues	calibrated probability unless separately evaluated
Is the history complete?	trace header, ordered mandatory entries, schema version and completeness validation	events never exposed by the provider or host
What can replay establish?	retained inputs, deterministic fields, trace reconstruction and result comparison	historical provider serving or external tool state

Bounded Agent Vocabulary

Claim	Required evidence	Bound on the claim
contract-valid role call	strict input, output or failure, metadata, and version	does not establish content correctness
governed lifecycle	declared transitions, passive roles, controller decisions, and terminal state	applies to the canonical graph or an equivalently declared custom graph
converged run	named strategy, window, observations, snapshot, hash, and typed reason	stable agreement can still be wrong
successful outcome	accepted terminal status, decision, validation, and termination reason	cannot conceal failed shards or vetoes
complete trace	mandatory header and ordered entries sufficient to reconstruct the outcome	cannot recover unrecorded provider or host events
replayable trace	complete replay metadata, deterministic fields, retained inputs, and zero temperature	does not reproduce historical provider serving
provider connectivity	named provider and model, configuration, live response, usage, and failure behavior	proves neither truthfulness nor future availability
CLI and HTTP parity	matching outcomes and trace semantics for their shared contract	HTTP v1 currently uses a narrower fixed offline pipeline

Keep Output, Trace, And Acceptance Separate

Final content answers what the workflow produced. RunTrace answers how the authorized roles, calls, transitions, vetoes, and convergence decisions produced it. Runtime acceptance answers whether that traced workflow was admitted under run policy. Store and display their identities together; none can be reconstructed safely from another.

Useful text can accompany an aborted, vetoed, exhausted, interrupted, partial, or non-converged outcome. Preserve that classification. A confidence number is an agent-produced field until a versioned evaluation demonstrates calibration for the intended domain.

Provider Evidence Is External Evidence

A provider response is untrusted input even when it satisfies the adapter schema. Retain provider and model identity, parameters, prompt and input hashes, adapter configuration, usage, observed failure, and relevant tool identity. Keep credentials outside configuration, traces, logs, artifacts, snapshots, and committed examples.

Continue with invariants for enforced orchestration laws, known limitations for provider, replay, credential, and hosting bounds, and the risk register for operational failure signals.