This repo is the methodology we actually audit against, not a summary of it: 86 controls, 355 test attributes with pass criteria, evidence standards, and the Type II testing method including how we call deviations, with worked examples. It's generated from the same JSON that drives our audit tooling, so it can't drift from what we actually do.
What "for your AI" means concretely: the framework is JSON rather than prose, so the controls, the attributes, their pass criteria and the evidence map are all machine-readable, and there's a file listing every tool our server exposes. It's CC BY 4.0 — point your own model at it and run your own readiness against the same bar the examination applies. The part that makes that work is 498 calibration examples: each one records a judgment call, the verdict an AI reached, the verdict that was correct, and why. Without those a model grades itself generously.
Why publish it: the audit criteria (AICPA's Trust Services Criteria) are public, but every firm's actual testing layer is a black box. A buyer holding two SOC 2 reports can't tell whether one auditor inspected evidence and the other just collected screenshots — the reports look identical. This year's Delve episode (hundreds of near-identical reports, procedures allegedly drafted before client evidence arrived — allegations Delve disputes) made that opacity harder to defend. Standards bodies publish their standards; we think the testing layer should be public too.
The thing that surprised us building it: sampling mostly shouldn't exist at small-company scale. Sampling exists in audit because looking at everything used to be expensive. At the company sizes we serve, populations are tiny — for most of our 79 sample-typed attributes, a proper sample would have been most of the population anyway — and AI collapsed the cost of looking at the rest. So the method defaults to testing complete populations, and sampling survives only as a disclosed fallback with hash-seeded selection that nobody (client, us, or either side's software) can steer.
Limitations, stated plainly: this is a methodology, not a track record. The firm is new — one Type I issued, no Type II yet. We published the method before the first Type II run on purpose, so it can't be quietly fitted to results afterward. Peer review: enrolled, first review due 2027.
Happy to answer anything about how SOC 2 audits actually work from the inside.