Methodology
How the Quiet AI Standard is scored, and how a review is run. The framework is published so anyone can judge our rigor, and every AI Trust File™ cites the designation of the standard it was issued under.
Designation and versions
The standard is designated QIA-1:2026. Material changes produce a new version on this page; a review is always read against the version in force on its issue date.
- QIA-1:2026, 5 August 2026. Formal designation adopted. Pillar names aligned across the site, the specimen and published material. No scoring changes.
- Standard v3.1, 10 June 2026. First published edition: four pillars, 100 points, evidence maturity scale, three-pass verification.
- Standard v3.0, spring 2026. Internal drafting editions.
The four pillars
The Quiet AI Standard evaluates AI governance across four pillars, each worth 25 points, for a total of 100.
- I · Data integrity. What goes into the systems: provenance and lineage, consent and lawful basis of inputs, representativeness of training data, minimization, defined use boundaries.
- II · Human infrastructure. Who answers for the systems: named ownership and roles, an oversight body, authority to intervene, competence and training, external assurance.
- III · Experience and transparency. What the people affected can know and do: disclosure of AI involvement, explanation of decisions, recourse and human review, communication clarity, affected-party impact.
- IV · Privacy and security. How information is held: protection in transit and at rest, access control, retention and deletion, breach response, third-party handling.
How findings are classified
Each criterion produces one finding, classified as a Strength (documented, operational, well evidenced), Partial (exists but incomplete or in draft), or Gap (material, recorded and scored conservatively). Findings cite their evidence, with dates; scores never exceed what the evidence supports.
Evidence maturity
Every finding is graded for the strength of its evidence: E0 unsupported · E1 self-attested · E2 public · E3 client document · E4 tested or corroborated · E5 monitored. A strong score resting on limited evidence is shown as such.
How a review is run
A review is produced by the Quiet AI diagnostic engine and is reviewed and signed by the Quiet AI Standards Committee before a file is issued; scoring is human-reviewed by design, and no result is issued by a fully automated decision. Each finding is drafted, challenged by a second pass, and confirmed by a third that checks every citation against its source. No score is awarded without a specific supporting source. Where evidence is ambiguous or incomplete, the lower score is recorded; the method is built to understate rather than overstate.
Score and level are separate
The score describes the strength of governance the evidence shows, out of 100. The review level describes how deeply we examined it: what you commission, never a grade. A shallow review can return a high score and a deep review can return a low one. Quiet AI does not pass or fail anyone, and reviewer confidence is reported separately from the score.
Review levels, by depth
- Listed: registry listing, declared posture, public verification page.
- Silver: review against the four pillars, reviewer-signed.
- Gold: adds independent verification against public sources and the Trust Centre.
- Platinum: adds document and evidence examination.
- Diamond: adds monthly monitoring, subsidiary coverage and red-flag alerts.
Review bases
A file is produced under one of three bases, recorded on its result page: an authorized paid review commissioned by the organization; a public-source specimen or research file built only from public evidence, without the organization’s participation; or a buyer-requested review commissioned by a third party assessing the organization.
The integrity rule
The fee buys a review, never a result. Nothing about how the work is priced or paid affects the score, the level or the findings, and nobody at Quiet AI is compensated on the outcome of a review. This rule is published so clients can hold us to it.
Corrections
Verified factual corrections to an issued file are made within ten business days and recorded in the file’s verification record. Material changes to this methodology produce a new version above, never a silent edit.