An open, versioned standard for recording, computing and verifying an assessment of an organization's AI controls against the published Apeiris corpus. Two recorded sub-states per control — design and operating — compute a rung on a closed ladder through a published nine-cell table and a numbered fail-closed cap catalog. Failures are first-class tokens that dominate every rollup, every population names its enumeration basis, and any conforming verifier recomputes every verdict from the recorded inputs.
The specification and its annexes. Written to be implementable by someone who has never spoken to us.
The method itself: sub-states, the ladder, the nine-cell table, the cap catalog, dominance, rollups, rendering rules and the assurance-basis axis.
Sampling minimums with their assumptions stated, population rules, selection methods, the assessment window (W1/W2) and freshness.
The exception-class taxonomy and the procedure enum — what counts as having performed a procedure, and what an exception is.
Ordinal scoping inputs — consequence class, autonomy, reversibility, population exposed, regulatory classification. Judgment made visible and challengeable, not removed.
The conflicts declaration that must travel with any identity binding. Verified identity without disclosed conflicts is worse than neither.
What ACAM is, what it explicitly is not, the release unit, and the versioning rules.
The same rules as data, so a tool can apply them without transcribing them. Our own /assess/ reads these at page load and hardcodes nothing — which is the point: it cannot drift from the standard, and neither can yours.
Ladder, sub-states, the nine-cell design × operating table, the numbered cap catalog, dominance and rollup and scorecard rules, assurance basis, verifier states, prohibited terms.
Minimum-n table with its confidence parameters and tolerable deviation rates, population rules, window rules, freshness, deviation handling.
The closed procedure enum and the machine-attested execution-log alternative, so automated controls are first-class rather than forced through human tickmarks.
Exception classes as data, counted rather than summarised into a single adjective.
The entity-risk axes that feed recommendation weights and sampling minimums.
The conflicts-declaration schema.
The prohibited-terms list, machine-enforced. Words we will not use about our own output, and neither should a conforming implementation.
What ACAM is not. It is not an audit standard, an attestation standard, a certification scheme, or a
conformity-assessment scheme. Its quality intent is stated plainly: designed to exceed conventional assessment
deliverables in transparency, machine readability, provenance and repeatability, while explicitly not replacing
independent evidence inspection or professional attestation. The gap that self-recording leaves — interviews,
evidence inspection, sampling verification, professional accountability — is declared in the assurance-basis tiers
rather than papered over.
Build your own. The corpus, this method, the dossier schema and the verification tooling are complete and
free, and are meant to be sufficient for someone else to build a competing tool. That is deliberate. A standard that
only its author can implement is not a standard, and an ecosystem of implementations is worth more to us than a
moat around one.
Attribution. CC BY 4.0 — adopt it, extend it, ship it in your product; cite ACAM and its version.
Versioning rules and the release unit are in the README.