evalgate

Decide whether an eval delta is a real regression or just sampling noise, and fail CI only when it is real.

A model eval that drops from 90.0% to 89.4% on a 1,000-example set looks like a regression, but on that sample size it is noise. Gating CI on the raw number makes the build flap; ignoring it lets real regressions through. evalgate runs the appropriate significance test and fails only when the candidate is significantly worse.

$ evalgate proportions \
    --baseline-score 0.900 --baseline-n 1000 \
    --candidate-score 0.894 --candidate-n 1000
verdict     worse, but within noise
difference  -0.0060
p-value     0.6232
alpha       0.05
# exit code 0 -> build passes

Install

$ pip install evalgate-cli                 # from PyPI, once released
$ pip install git+https://github.com/jmweb-org/evalgate   # latest, available now

Pure standard library plus typer and rich. No heavy dependencies.

Usage

From two aggregate accuracies

$ evalgate proportions \
    --baseline-score 0.90 --baseline-n 2000 \
    --candidate-score 0.87 --candidate-n 2000 \
    --alpha 0.05

Uses a two-proportion z-test on the accuracies and their sample sizes.

From paired per-example results

When both models were evaluated on the same examples, a paired test is more powerful. Give a CSV with per-example correctness (or predictions plus a truth column):

$ evalgate paired results.csv --baseline base_correct --candidate cand_correct
$ evalgate paired results.csv --baseline pred_a --candidate pred_b --truth label

Uses McNemar's test: an exact binomial test on the discordant pairs, or the continuity-corrected chi-squared approximation for large samples.

In CI

- run: evalgate proportions --baseline-score 0.90 --baseline-n 2000
        --candidate-score "$SCORE" --candidate-n 2000

Verdicts and exit codes

A bad invocation (scores out of range, missing column, unreadable file) exits 2.

JSON output

When using the --json flag, evalgate returns a JSON object with the following fields:

What it does and does not do

It answers one question: is this difference larger than sampling variation? It does not correct for multiple comparisons across many evals, and a noise verdict means "not proven", not "proven equal". For small evals, collect more examples rather than trusting a borderline p-value.

License

MIT. See LICENSE.