Trace design
Evaluation should inspect actions, committed effects, retry behaviour, and policy boundaries rather than score final text alone.
Fixture release statusEvaluation methods
Method notes,
not a live observatory.
This section explains the evaluation direction behind the public fixture work. It is static documentation; it has no live backend and no agent trigger.
Failure taxonomy
Useful distinctions include wrong-scope probes, expired evidence, hidden underlay risk, valid alternate routes, and no-safe-action cases.
MethodsPublic method artifacts
The public repository exposes the proposed trace fields, failure categories, and publication checks used to make method boundaries inspectable. These documents are versioned research material, not proof of a benchmark result.
Trace schema · Failure taxonomy · Publishing preflight · Validity registry