// begin
One method.
Two ways to begin.
Both start the same way: one system, one claim, and the bar agreed in writing before anything is measured. Both end the same way, in a source-bound record and a verifier you run yourself. What differs is the instrument and what you hand over.
// § 01 · a gate review
One repository. What your merge gate actually covers.
The shorter path, and the one most teams should start with. It needs no access to your systems and no engineering time from you beyond choosing the repository.
01
You send
A workflow file, or a repository name if it is public. Send your required-check list as well if you have it. Nothing in a workflow file establishes whether a gate is actually enforced.
02
I return
The required-check closure, the jobs outside it labelled active or latent, exact ordering counts where the graph can be enumerated, a counterexample, and the minimal dependency change that closes the gap.
03
You check it
A verifier that recomputes every claim from your copy of the file, with no network access and none of my code in the loop. Findings you judge to be intentional policy are recorded as exactly that.
Static analysis only. Nothing is executed, no repository access is required, and no result is published without your approval. A first read is an afternoon, not a procurement cycle.
// § 02 · a model evidence pilot
One release. Three weeks. Bars in writing first.
The deeper engagement, for a model or agent release where the question is what the system is entitled to assert.
W1
Your run
You run the public kit against your endpoint, on your infrastructure. Nothing leaves your machines. We agree thresholds in writing: what counts as a pass, before any pass is possible.
W2
Scored runs
Frozen banks, including held-out items whose commitment hash was published in advance, scored with the strict public scorer.
W3
The record
An evidence report, or an issued certificate where the full signed profile is met. A miss is reported as a miss, with an autopsy. That is the point of setting the bar first.
// § 03 · the terms, both paths
A
Your approval controls publication
Nothing is published, named or referenced without your written agreement. A finding you decline to publish stays private permanently, not until it is convenient.
B
Bars before runs
Thresholds are fixed in writing before anything is measured. A threshold chosen after seeing the result is not a threshold, and I will not set one that way.
C
A miss is a result
If the system passes, you get a record that survives hostile reading. If it does not, you get a specific account of where and why, which is usually worth more.
// § 04 · poor fit, said plainly
A gate review is a poor fit if your CI policy is deliberately tiered and you already know which validators are advisory: the analysis will tell you what you already know. A model pilot is a poor fit if you need a number for a deadline rather than a number you can defend, because the bar is agreed first and it may not be met. Neither is a substitute for a security audit or a penetration test, and neither claims to be.