METRIC -> MECHANISM -> ARTIFACT -> OWNER -> DECISION
Context: workload slice + revision tags
Model: quality + calibration + convergence behavior
Hardware: utilization + bandwidth + precision impact
Workflow: signoff-correlation + confidence gating
Deployment: compatibility + monitoring + rollbackWhich metric and workload slice regressed first?
Is root cause in data, model, runtime, or hardware mapping?
Which artifact proves mechanism rather than assumption?
Who owns the smallest reversible fix?
What validation + rollback gates protect release quality?