Models: objective to compared candidate

00:00:00.000 — Start with the operating question
Predict shipment SLA breaches before 21:00. The objective defines the target, owner, and intended use.

00:00:02.000 — Define what the prediction may use
The decision contract uses only data known by 14:30 UTC and retains the hazardous-load restriction.

00:00:06.000 — Keep each candidate tied to its evaluation
Three imported candidates retain their own Evaluation references. Comparing them does not train or deploy a model.

00:00:09.000 — Compare the candidate evidence
Compare the baseline and two candidates against the same predictive evaluation. Aggregate quality alone does not establish readiness.

00:00:13.000 — Inspect the slice that fails
Trial 3 has hazardous-load recall of 0.84 against a minimum of 0.90. A stronger overall score does not remove this failed guardrail.

00:00:19.000 — Retain the next decision
Continue study preserves the decision and its reason. No new model release is selected by this comparison.

00:00:23.000 — Keep release separate from comparison
No new Model Release is linked. The recorded decision is to improve hazardous-load recall before proposing a release.