Frequently asked questions

Does my data leave my process when I use this?

No. 43 of the 43 modules run as a library inside your own process, on your own

compute. There is no ingestion endpoint, no telemetry on by default, and no licence heartbeat —

the licence token is verified offline for 90 days against a key compiled into the package. The

only outbound calls the toolkit can make are your own model calls, through GuardedClient, on

your key, to your provider — and those are optional, because every module has a deterministic

path that runs with no key and no network. The core imports 0 third-party

packages, and a CI job asserts it on every commit.

Is this a certification, or does it make us compliant?

Neither. The AtherosAI Compliance Kit produces assessments and evidence — it assesses and

evidences, and it does not certify. It is not a conformity assessment within the meaning of

Regulation (EU) 2024/1689, and it is not legal advice.

Regulatory obligations rest with you and no output of this toolkit transfers them. A lint over

every generated report and every marketing page — in English and Turkish — fails the build if the

words "certified", "fully compliant", "guaranteed compliant" or "no further action required"

appear anywhere.

What does "unmeasured" mean, and why does it fail my build?

A score of unmeasured means the measurement did not run or could not be computed — not that it

passed. It renders grey, never green, and it fails the CI gate. Treating it as a pass would mean

the gate goes green at exactly the moment it should not: when the measurement itself broke. This

is the single design decision the product rests on, and it is enforced in code rather than

documented as an intention.

Which module evidences which EU AI Act article?

Each obligation names the module that produces its evidence.

Which ISO/IEC 42001 clauses does this cover?

5, and the Kit is explicit about which. It produces evidence toward named clauses — it does not automate an internal audit of the standard, and nothing here should be read as covering it end to end.

Clause 6.1.2 is listed as an input, not a discharge. An EU AI Act risk classification is a regulatory categorisation; the clause asks for your own AI risk criteria, analysis and evaluation. atheros-kit iso export prints that distinction on the clause itself, so an auditor is never handed the more flattering reading.

Not covered:

• Clause 4 — context of the organisation and the scope of the AI management system
• Clause 5 — leadership, the AI policy, and assigned roles and responsibilities
• Clause 6.1.3, 6.1.4 and 6.2 — risk treatment, the AI system impact assessment, and the AI objectives
• Clause 7 — competence, awareness, communication and control of documented information
• Clause 9.2 — the internal audit programme
• Clause 9.3 — management review
• Clause 10 — nonconformity, corrective action and continual improvement
• The Statement of Applicability
• The Annex A controls, as controls: the Kit evidences the operation of a few of them, and selects, justifies and documents none

Those are management-system work that a tool running inside your CI cannot do for you, and a report that implied otherwise would be the failure this product exists to prevent.

Does it work without an API key or a network connection?

Yes, and that is the supported default rather than a degraded mode. Every module has a

deterministic path. atheros-kit doctor prints which providers you hold keys for and therefore

which path each check will take. Where a model was configured and did not answer, the score is

tagged degraded and the report says so — a fallback is never silent.

Is this a library or a governance platform?

A library and a CLI, plus a read-only console that renders the reports the library emitted. It

runs where your data is. A hosted platform structurally cannot make that claim, which is why the

console has no server, no database and no write path, and why there is no SaaS version planned.

What does it cost, and how is it metered?

It is not metered. The toolkit runs on your compute against your model key, so there is no

cost-per-run to pass on — and metering it would punish the behaviour the product wants, which is

running the gate on every commit.

Which vector stores are supported?

4 connectors ship: Chroma, pgvector, Pinecone and Milvus. For anything else, read

the corpus yourself and pass the chunks in — that path is first-class rather than a fallback, and

it is why the assessment code never touches a driver. Every connection is read-only by

construction; remediation emits recipes for a human to run and never mutates your corpus.

How can an auditor verify the audit chain independently?

Two independent implementations. atheros-kit audit verify recomputes every SHA-256 digest in

Python; the console recomputes the whole chain again in the browser using WebCrypto, sharing no

code with the writer — a verifier that shares code with the writer can only prove they agree with

each other. An intact chain makes deletion, reordering and editing detectable. It does not make

the contents true: the chain attests to what was recorded, not to whether the assessment behind

the record was correct.

Does the free tier expire?

No. It requires no activation, no key and no network call, it may be used commercially by any

number of people, and it does not expire. It covers the guardrail wrapper, EU AI Act risk

classification and the complete hash-chained audit ledger. Gating the ledger would have been easy

and would have removed the reason the free tier is evidence rather than a toy.

How do I know whether my system is high-risk under the EU AI Act?

Run atheros-kit euact classify with your sector and use cases. The engine evaluates in statutory

order — Art. 5 prohibitions first, then Annex I product safety and the

8 Annex III categories, then Art. 50 transparency, then minimal — against a

versioned vocabulary (EU-2024/1689:2024-07-12), and every classification records the version it ran

under. Where the inputs support more than one answer it reports a grey zone with the conflict

named rather than a confident tier, and a minimal verdict states that it rests on no indicator

matching, which is not the same as evidence of low risk.

How much of an Annex IV file can be generated automatically?

Part of it, and the report tells you which part. The generator produces all

9 sections and marks each covered, partial or missing. Machine evidence

makes a section partial, never covered: Annex IV asks for an account of the system, which

evidence supports and does not replace. Our own self-assessment publishes a 33% completeness score

for our own product, incomplete, because a generator that emitted plausible prose for all nine

sections would produce a document that looks finished and is not — one that survives an internal

review and fails an external one.

Last updated: 2026-09-07 · AtherosAI B.V.