Attribution
The 5×4 taxonomy and the deterministic rules behind every cell.
Attribution answers where a failure entered the agent configuration and what kind of mistake it was. A trace shows the hand-offs; attribution names the layer and the mode, then attaches evidence from the recorded steps.
Layers
- Spec — contradictory, ambiguous, or underspecified task
- Model — wrong reasoning, ignored instruction, hallucinated API
- Harness — swallowed error, no retry, bad stopping condition, dropped state
- Tool — wrong result or the tool's own failure
- Environment — network, timeout, quota, or external nondeterminism
Modes
- Decision — wrong choice among viable options
- Execution — correct choice, failed to execute
- Recovery — should have recovered from a fault, did not
- Detection — should have noticed an inconsistency, did not
Twenty cells. Seventeen rules ship today, including patch_apply_failed_unhandled (harness × recovery), test_failed_after_patch (model × decision), test_run_skipped (harness × execution), hallucinated_api (model × decision), and misleading_spec_followed (spec × detection). A rule that matches without a counterfactual is medium confidence and is never presented as proven.