abkit

Sample-ratio, peeking, multiple-testing and winner's-curse checks for experiment readouts.

An experiment readout can say "B wins, p = 0.03, ship it" while the traffic split was broken, three whale customers carried the whole effect, the analyst stopped at the first significant refresh, the winning metric was one of forty tried, or the experiment never had the power to estimate what it now claims. None of that shows up in the p-value. abkit reads the per-unit data and the intended design, measures each of these failure modes, and flags whatever leaves the range a healthy experiment could occupy.

The audit in one table

Eight synthetic experiments, seven implanted defects. Bold values are flags; each lands on the row where its defect was implanted, and the clean row carries none.

The row to sit with is peeker: a true null effect, analyzed after every batch of 250 units and stopped the moment z crossed 1.96. Its final readout is indistinguishable from a real win — p < 0.05, decent sample size — but the schedule it followed has a 19% false-positive rate, not 5%. The audit flags the process, which is the only place that defect lives.

make demo regenerates the table, the full report and the figures from fixed seeds; CI rebuilds them from pinned dependencies and fails if a committed number differs from what the code produces.

4,873 real experiments, audited

examples/upworthy/ runs abkit's aggregate checks over the exploratory split of the Upworthy Research Archive — 4,873 real headline A/B tests from 2013–2015. Committed results: examples/upworthy/results/upworthy.md.

• 16% of real tests fail the sample-ratio check at p < 0.001 against the platform's equal-allocation design — and not because packages were added mid-test; the rate is, if anything, higher in tests whose variants all launched together. Those readouts carried unknown exposure bias.
• 10.9% of tests have a "significant winner" (best headline beats the runner-up at raw p < 0.05), but Benjamini–Hochberg across all 4,873 winner comparisons keeps only 2.7% of them.
• The median significant winning lift (0.85 CTR points, power 0.62) is expected to overstate its true effect by 1.27x even when the winner is real — the winner's curse, measured on real experiments.

Install

pip install abkit

Python 3.11+. Runtime dependencies: numpy, pydantic, matplotlib.

Quickstart

Log one JSON object per randomized unit, and the design the experiment was supposed to follow:

{"unit_id": "u1", "arm": "treatment", "metrics": {"revenue": 12.4, "conversion": 1.0}, "pre": {"revenue": 9.1}, "t": 1718040000}

{
  "split": {"control": 0.5, "treatment": 0.5},
  "control": "control",
  "treatment": "treatment",
  "primary_metric": "revenue",
  "alpha": 0.05,
  "mde": 0.5,
  "looks": [{"n": 20000, "z": 1.31}, {"n": 40000, "z": 2.11}]
}

from abkit import load_units, load_design, run_audit

audit = run_audit(load_units("units.jsonl"), load_design("design.json"))
for flag in audit.flags:
    print(flag.name, "--", flag.detail)

Or from the shell, with an exit code CI can gate on:

abkit report units.jsonl --design design.json --out audit --fail-on-flags

Alongside report.md the audit writes report.json for pipelines that gate on specific numbers. Any check that could not run is listed with the exact fields to log to enable it — a skip describes the log file, not the experiment.

What it checks

How the numbers are defended

• Cross-checked implementations. The chi-square survival function matches scipy to 1e-10 across df 1–10, Welch's statistic matches scipy.stats.ttest_ind, the two-proportion z matches statsmodels proportions_ztest, and Benjamini–Hochberg reproduces statsmodels multipletests decisions exactly. The library itself depends on none of them.
• Known closed-form results. The Gelman–Carlin power / Type-S / exaggeration triple is checked against numerical integration to 1e-6, and the peeking simulator reproduces the classical repeated-significance rates (0.083 at 2 looks, 0.142 at 5, 0.193 at 10) within Monte-Carlo tolerance.
• Validation by implantation. The synthetic experiments have defect dials, and the tests require each probe to recover its dial — an implanted 0.3 covariate shift must read as SMD 0.3 — to stay monotone in it, and to stay silent on the clean experiment.
• Drift-checked results. The demo table, report and README numbers are regenerated by CI from pinned dependencies and diffed against the committed copy.

Design notes

• Innocent bands, not point nulls. With enough data an SMD of 0.02 excludes zero while meaning nothing. Band probes declare the interval a healthy experiment could plausibly occupy — SMD and novelty gap within ±0.10 — and flag only when the whole 95% bootstrap CI leaves it. Exact procedures use the field's conventional thresholds: SRM at p < 0.001, contamination at zero tolerance. All constants live in one place (probes.py) and are easy to disagree with.
• The stopping rule is audited, not the final p. A peeked experiment's final test statistic looks ordinary; what is broken is the procedure that produced it. The peeking probe Monte-Carlos the analyst's actual look schedule under the null and reports the false-positive rate that schedule had.
• Fragility is counted in units, not sigmas. "Removing two customers overturns the decision" is a statement a stakeholder can act on; a leave-one-out influence statistic is not.
• Exaggeration needs a basis effect. With design.mde the winner's-curse probe reports power and exaggeration at the effect the experiment was designed for; without it, the observed effect is used and labeled an optimistic bound, since conditioning on significance inflates it.
• Stratified bootstrap. Units are resampled within their arm, never across, so group sizes — which determine every statistic's sampling distribution — are preserved.
• Deterministic to the digit. Seeded PCG64 everywhere, no BLAS in any statistic's path, and a pinned drift environment: the same data gives the same report on any machine.

Limitations

• An audit is bounded by what was logged. No pre-period covariates, no balance or CUPED checks; no look history, no peeking check; no per-unit data at all, and only the aggregate checks (abkit.stats) apply.
• The innocent bands are judgment calls, not derivations.
• The peeking probe assumes the recorded looks are all the looks there were.
• Fragility uses a greedy removal path; it upper-bounds, and in adversarial cases may not find, the true minimal flip set.
• Implantation shows the probes detect the mechanisms simulated; real experiments can fail in ways not simulated here.
• The Upworthy analysis treats the archive's documented equal allocation as the intended split; if some tests intentionally used unequal splits, their SRM flags are misattributed.

Related work

The failure modes measured here are the standard ones from the online experimentation literature: sample-ratio mismatch and its taxonomy (Fabijan et al., 2019), trustworthy experimentation practice (Kohavi, Tang & Xu, 2020), peeking and always-valid inference (Johari et al., 2017), variance reduction with pre-experiment data — CUPED (Deng et al., 2013), Type-M/Type-S errors (Gelman & Carlin, 2014), the winner's curse in A/B testing (Lee & Shen, 2018), and experimentation culture at scale (Kaufman, Pitchforth & Vermeer, 2017). The real-data case study uses the Upworthy Research Archive (Matias, Munger & Watts, 2021). abkit's contribution is packaging the checks as one auditable tool with uncertainty on every number and validation by implantation.

Companion projects: judgekit applies the same audit-before-you-trust pattern to LLM judges; raterkit applies it to the human labels underneath both; trajectory-judge measures what outcome-only judges miss on agent trajectories; arenakit applies it to pairwise model leaderboards.

Citation

Releases are archived on Zenodo. Cite the concept DOI 10.5281/zenodo.21809478, which always resolves to the latest version; structured metadata is in CITATION.cff.

@software{mohammadi_abkit,
  author  = {Mohammadi, Hadi},
  title   = {abkit: audit an A/B-test readout before you ship the decision},
  url     = {https://github.com/mohammadi-hadi/abkit},
  doi     = {10.5281/zenodo.21809478},
  version = {0.1.0},
  year    = {2026}
}

License

MIT