The StackAttest Standard: how we decide whether AI-built software is production ready

Software is being written faster than trust in it can keep up. Roughly 84% of developers now use AI coding tools, yet close to 46% say they do not trust the output. Independent research keeps finding that around 45% of AI-generated code ships with a known security weakness. That gap is the whole reason StackAttest exists, and the StackAttest Standard is how we close it.

This page is the canonical explanation of the Standard. If you want to know exactly what a StackAttest score means, how it is produced, and why we report four numbers instead of one, this is the reference.

Why one number is not enough

Most trust products give you a single grade and ask you to believe it. A single grade hides the two things that actually matter: how much evidence sat behind the verdict, and how confident the reviewers were. A high score built on thin evidence is not the same as a high score built on deep evidence, and pretending they are the same is how trust products lose the trust.

So the Standard reports four separate signals, and each one answers a different question.

1. The Score

The headline result. It is produced by a deterministic engine, which means the same inputs always give the same score, with no model opinion mixed in. Run it twice on the same commit and you get the same number every time. A single critical failure, like a missing production gate, cannot be quietly averaged away by strong results elsewhere.

2. Evidence Coverage

How much of the Standard we could actually check for this product. A control with no evidence is marked as uncovered, not silently passed. Coverage tells you how complete the picture is, so you never mistake an absence of evidence for a clean bill of health.

3. Validation Confidence

How strong the evidence behind the covered controls is. Live, observed evidence counts for more than a founder declaration. Confidence separates a result we watched happen from one we were told about.

4. AI Consensus

How much a 6-model council agreed when it examined the evidence. The Passport shows how many configured reviewers answered, whether a critical objection remains, and the minority opinions behind a split result.

Evidence over opinion

The active control pack maps technical checks to OWASP ASVS 5, NIST SSDF, CWE, OWASP API Security Top 10, WSTG, SLSA. URL, source, dependency, and runtime checks gather evidence directly. Where a fact lives outside those surfaces, a typed attestation can record the claim and its artifact without pretending it was independently observed.

Every finding points back at the evidence that produced it. You can ask why a control passed, failed, or stayed uncovered and get the evidence grade and rationale, not a vibe.

Deterministic first, AI second

This is the part people most often get wrong about StackAttest, so it is worth being precise. The Score is never decided by an AI model. It is computed by deterministic rules over the evidence, which is what makes it replayable and hard to argue with.

After the Score is set, a 6-model council examines the structured evidence and challenges the engine's verdicts. Reviewers can agree, disagree, ask for stronger evidence, and record minority opinions. Their agreement becomes AI Consensus. It never changes the deterministic Score.

The council challenges the evidence. It never changes the Score.

The verification ladder

A Passport starts at L0 Registered and can progress through externally verified, code verified, runtime verified, production validated, and continuously validated levels. Each rung records a stronger class of evidence; no level is awarded by marketing copy or model opinion.

The Passport you own

The output of a validation is a Passport. You decide whether to publish its trust page, which due-diligence sections to grant, when access expires, and when to revoke it. Public readers see verified results and approved evidence summaries, never the repository behind them.

A public Passport can also carry a signed attestation. Anyone can check that signature against our published key, without an account and without seeing any evidence, and confirm the result has not been tampered with.