Package dev.dokimos.core.gate

Class RegressionGate

java.lang.Object

dev.dokimos.core.gate.RegressionGate

public final class RegressionGate
extends Object

Compares a candidate experiment result against a committed baseline and produces a GateVerdict. This is the comparison step only; reading the baseline file and turning a verdict into a test failure are handled by RegressionGateRunner and Assertions.

The gate fails if either of two independent guards fires:

• Guard 1 (broad) — the RunComparison engine reports a significant aggregate pass-rate drop or any significantly regressed evaluator (RunComparisonResult.hasRegressions()). It stays quiet on judge noise but cannot see a localized break on a small dataset.
• Guard 2 (localized-severe) — any shared item whose worst raw per-evaluator score drop falls below -severityMargin. This catches a single item that breaks hard even when the aggregate change is not significant.

Coverage-loss conditions fail the gate independently of failOnRegression, because guard 1 cannot see them: a removed evaluator (per onRemovedEvaluator), a removed item (only when failOnRemovedItems), and — under dataset_item_id pairing — a candidate item with no id, which pairs against nothing and would hide a regression on it (always a FAIL). A baseline id absent from the candidate warns; whether it also fails is governed by failOnRemovedItems. A threshold change between sides only warns.

Method Summary

Modifier and Type	Method	Description
static GateVerdict	evaluate(ExperimentResult candidate, BaselineFile baseline, GateConfig config)	Compares a candidate against a baseline and returns the gate verdict.

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

evaluate

public static GateVerdict evaluate(ExperimentResult candidate,
 BaselineFile baseline,
 GateConfig config)

Compares a candidate against a baseline and returns the gate verdict.

Parameters:

candidate - the candidate experiment result

baseline - the committed baseline to compare against

config - the gate configuration

Returns:

the verdict (status PASS or FAIL; never NO_BASELINE, which the assertion layer produces when no baseline file exists)