Every AI security product eventually has to answer “which control does this satisfy?” Building that mapping means watching frameworks that move, extracting normative requirements, deciding honestly what can be detected and what can only be attested, and keeping all of it current. We do that work continuously and deliver it as a versioned pack. Use it inside Agenor Attest, or license it for your own product.
Counts, never percentages · unknowns render as unknown · every requirement anchored to a source hash
What is actually in it today
These are counts from the live corpus, not a roadmap. The frameworks below carry extracted requirement records; the 43 pins are what we watch for version movement, which is a larger set than what we have finished extracting.
The part nobody else does honestly
A mapping that claims coverage of everything is worthless, because the buyer discovers the gaps during the audit instead of during the sale. Each control carries a disposition and, where it is automatable, a scope. 18 controls are marked not-feasible, and they stay in the pack saying so.
A detector can, in principle, evidence this control from telemetry. Split further by scope: 257 in-scope for AI agents, 7 in-scope for cloud, 553 feasible but not yet in a shipped scope, and 18 ruled not-feasible.
No telemetry will ever evidence these. They need a human to assert something: a policy exists, a review happened, a person is accountable. Calling them automatable would be the easiest lie in this business.
Automatable in principle, and we judge that no detector we would build gives a useful signal on them. A scope ruling inside the 835, not a fourth bucket. It is the most consequential call in the schema, so it is flagged for human review every time regardless of the extractor's confidence.
Why mappings rot
A pinned file that returns 200 forever is not a current framework. It is an abandoned one.
We track what the publisher has shipped, what we captured, and what the product actually ships against. The gap between the last two is the one nobody else measures, and it is where a mapping quietly stops being true.
We pinned a framework artifact at a branch path. The publisher moved to tagged releases and abandoned it. The file kept returning 200, the hash stayed stable because the file was dead rather than because the standard was, and every check reported healthy. Unchanged bytes and current content are different claims.
A publisher's own landing page can advertise an old edition long after the current one ships elsewhere. A page that is maintained but no longer authoritative is more dangerous than one that 404s, because nothing ever breaks.
How it is delivered
You receive the frameworks you licensed, at a stated version, with a receipt you can verify offline. On renewal you can take a delta rather than the whole pack.
The requirement mapping and the detection rules that evidence it cost different things to produce, so they are entitled separately. A mapping-only tier is a real tier, not a crippled one.
Reaching a control is not satisfying it, and a single blended coverage number is the most quotable and least true output this pack could produce. You get counts, and unknowns that render as unknown rather than as zero.
Tell us which frameworks matter to you and we will show you what we hold for them, including what we have marked not-feasible.