Separate model, backend, trace, and outcome signals

Rank conversational nodes with separate ledgers for model-routing defects, backend or tool outcomes, trace quality, and downstream call outcomes; do not collapse them into one failure rate.

Use one fixed half-open event-time window and completion cutoff. Count unknown and malformed traces separately, and link decisions to results only when a tool-call identifier has exactly one owner and one chronologically valid result. Compare traces only when they are governed by the same versioned contract or rubric. Keep recoverable business outcomes distinct from validation and infrastructure errors.

Prioritize avoidable model defects per eligible caller turn only after backend and trace tables show model work is the appropriate intervention. Abrupt endings and downstream outcomes remain associations rather than causal proof.