Assert the current run has not regressed against a committed baseline — the
CI gate (Phase C). A regression is an arm×metric that moved significantly in
the bad direction vs. baseline (Welch t-test, so sampling noise doesn't
trip it; see src/eval-baseline.ts). Higher is better by default; list
lowerIsBetter metrics (cost/latency) to flip them. Load the baseline with
readBaseline(path) and record a fresh one with writeBaseline(path, reports).
Assert the current run has not regressed against a committed baseline — the CI gate (Phase C). A regression is an arm×metric that moved significantly in the bad direction vs.
baseline(Welch t-test, so sampling noise doesn't trip it; seesrc/eval-baseline.ts). Higher is better by default; listlowerIsBettermetrics (cost/latency) to flip them. Load the baseline withreadBaseline(path)and record a fresh one withwriteBaseline(path, reports).