Setup

Install

Three ways in, least friction first.

Claude Code plugin

/plugin marketplace add 0trm/gallop
/plugin install gallop@gallop

All eight skills, invoked by name or picked up automatically when a question matches one. The bundled scripts call the gallop package, so add pip install gallop-pds if you want the checks to run rather than be described.

Any other agent, or none

A skill is a directory of markdown. Copy what you need:

git clone https://github.com/0trm/gallop
cp -r gallop/skills/reading-experiments .claude/skills/

Works with anything that reads Agent Skills, and reads fine as prose: each SKILL.md is the procedure, the reference/ files one level down are the depth.

The package

pip install gallop-pds

numpy, pandas, scipy, nothing else. Every function takes and returns arrays or DataFrames, never a database connection; SQL for the three queries every readout needs ships as string.Template files in sql/.

Prove it works with no configuration and no warehouse:

python3 -m gallop.examples.quickstart

That runs one simulated experiment through every check. This is what it prints:

gallop quickstart: one experiment through every check

1 · Size it before running it (gallop.power)
   at n=40,000 per arm on a 12.5% rate, the MDE is 0.66pp;
   detecting 0.35pp instead would need 140,159 per arm

2 · The trust gate (gallop.trust)
   SRM: chi2 0.16  p 0.689  -> pass
   exposure: pooled rate 97.00%  -> pass

3 · The effect, with CUPED (gallop.variance)
   raw    +0.292pp  se 0.234pp
   cuped  +0.306pp  se 0.227pp   variance reduction 6%

4 · An interval that survives peeking (gallop.sequential)
   always-valid 95% CI [-0.394pp, +1.005pp]   boundary |z| 3.08 (vs 1.96 fixed)
   significant under continuous monitoring: False

5 · Shrunk toward what this metric has done before (gallop.shrink + priors)
   prior from 8 readouts: mu +0.097pp   tau 0.065pp
   observed +0.306pp -> shrunk +0.113pp   (weight on data 0.08)

true simulated effect: +0.350pp. The shrunk estimate is the
one to write back to the store; the raw one is the winner's curse waiting.

The modules

Module	What it computes
`gallop.power`	MDE, sample size, duration; two-proportion and continuous
`gallop.trust`	Sample ratio mismatch; exposure versus eligibility
`gallop.variance`	CUPED against a pre-period covariate
`gallop.sequential`	Always-valid confidence sequences; O'Brien-Fleming bounds
`gallop.shrink`	Empirical Bayes shrinkage toward the prior store
`gallop.priors`	The prior store and metric registry on disk, validated JSONL
`gallop.explore`	Mix versus rate, funnel steps, segment scan, opportunity sizing
`gallop.validate`	Out-of-time validation: baseline lift, calibration, leakage, Qini, MASE

Each runs as a script too: python3 -m gallop.trust srm --counts counts.csv.