[go-fan] Go Module Review: gonum.org/v1/gonum #72043

Go Fan Report: gonum.org/v1/gonum

Module: gonum.org/v1/gonum v0.17.0 (already latest release, published 2026-01-10)

Selected via round-robin plus recent-activity priority: the gonum repo was pushed most recently (2026-10-25) among all direct dependencies not reviewed in the last 7 days.

Current Usage in gh-aw

Only 1 file imports gonum, and only one subpackage: pkg/cli/experiments_grader_statistics.go imports gonum.org/v1/gonum/stat/distuv (nothing from stat, mat, optimize, etc.).

Key APIs used:

• distuv.StudentsT{Mu,Sigma,Nu}.CDF(...) for Welch t-test p-value
• distuv.UnitNormal.CDF(...) for Mann-Whitney U and two-proportion z-test p-values
• distuv.Beta{Alpha,Beta}.Prob/.CDF(...) for Bayesian A/B posterior-superiority probability (numerically integrated over a fixed 4096-point grid)

Everything else in the file, including meanValues, sampleVariance, the Mann-Whitney ranking and tie-correction logic, and sumValues, is hand-rolled, not from gonum.

Research Findings

The gonum stat package (not currently imported here) provides Mean, Variance, and notably MeanVariance, which computes the unbiased (Bessel-corrected) sample mean and variance together in a single pass. That is the exact statistic the file hand-rolls today via meanValues plus sampleVariance, but with a redundant pass: welchTTest calls meanValues directly, and sampleVariance recomputes the mean again internally.

gonum does not ship a built-in Mann-Whitney or rank-sum test, so the hand-rolled implementation in this file is a reasonable, justified choice, not a gap.

Recent Updates (v0.16.0 to v0.17.0)

• distuv: added a non-central t distribution
• stat: added Wasserstein distance
• stat/distuv: added a documentation example for computing a confidence interval, notable because MetricComparison in this repo currently reports only p-values and superiority probabilities, no intervals

Best Practices

Gonum documentation favors the single-pass MeanVariance over separate mean and variance calls, and recommends passing nil for unweighted inputs instead of building an all-ones weight slice.

Improvement Opportunities

Quick Wins

• Replace meanValues(values) plus sampleVariance(values) in welchTTest (pkg/cli/experiments_grader_statistics.go) with a single stat.MeanVariance(values, nil) call per group. This removes a redundant mean computation per group and drops the need for the hand-rolled sampleVariance helper.
• The meanValues helper used for the Delta field in computeGraderMetricComparisons could be replaced by stat.Mean(values, nil), removing another hand-rolled helper in favor of the already-vendored dependency.

Feature Opportunities

• Add confidence intervals to MetricComparison, for example a DeltaCI [2]float64 field, using distuv.StudentsT.Quantile and distuv.UnitNormal.Quantile, mirroring the confidence-interval example gonum added to its docs in v0.17.0. This is standard practice for A/B reporting and is currently missing from the grader output.
• betaBinomialProbability integrates over a fixed 4096-point grid to get a posterior-superiority probability; a credible interval for the same Bayesian path, derived from distuv.Beta quantile or CDF, could be reported alongside it for consistency with the frequentist paths above.

Best Practice Alignment

• The file currently imports only stat/distuv and never stat, despite stat providing exactly the descriptive-statistics helpers (Mean, MeanVariance) this file reimplements by hand. Importing stat would better align with the intended split in gonum (stat for descriptive statistics, stat/distuv for distributions) and reduce custom code surface.

General Improvements

• No version upgrade needed; go.mod already pins the latest gonum release (v0.17.0).
• No other files in the repo import gonum, so there is no cross-file consistency work needed beyond this one file.

Recommendations

1. Swap meanValues and sampleVariance for stat.MeanVariance in welchTTest (quick win, no behavior change, same unbiased-variance formula).
2. Consider adding confidence intervals to MetricComparison using existing distuv quantile methods, since the module already imported for this purpose supports it directly.

Next Steps

• If confidence intervals are added, update any downstream report rendering that consumes MetricComparison to display the new field.
• Re-review after any change to AnalysisType handling, since that is the main place gonum APIs are invoked.

Generated by Go Fan
Module summary saved to: scratchpad/mods/gonum.md
Run: §40289806152

Generated by 🐹 Go Fan · claude · agent · ~97 AIC · ⌖ ~1 AIC · ⊞ 4.4K · ◷

• expires on Oct 27, 2026, 12:14 AM UTC-08:00