StaticMeasureC1 → S4.2: the stored decision on C1's reference split is keep-simple-baseline shipping the GLM (the reference keeps its GLM).
StaticMeasureC1 → S1.3: the unit's declared heuristic ("renewed last year") is scored on the same holdout as a named comparator
(its baseline row with both bounds, heuristic:<name>VsDummy and modelVs:heuristic:<name> stored), and the stored
decision agrees with the stored model-versus-rule comparison: when the model does not beat the rule by the margin the
outcome is not model and carries its reason (heuristic-carries-signal when the rule is the member kept). Measured
on the default year-cutoff holdout and on C1's reference split; each reports the outcome it saw. On both the model
beats the rule (measured), so this exercises the "beats it" side at runtime; the keep-the-rule branch is Core's
gate.test.ts › treats the unit incumbent heuristic as a simple member (heuristic-carries-signal).
StaticMeasureW7 → S4.2 #2: 0 of 3 seeds return model while the holdout AUC interval contains 0.5, and none ships anything while
the minority class is below the protocol minimum; every seed stores its guard inputs (a minority count below the
minimum on W7) and both bounds on every stored number.
StaticStoredAdditional coverage: every stored number has both bounds, slices cover the protocol, and precision at capacity is stored when the unit declares a capacity.
Gate outcome measurements read from the gate record stored on the training run.