Naevum’s EU AI Act architecture is a set of built-in records and oversight points for general counsel, so that an auditor’s question gets entries, proofs and timestamps as its answer.
Fix selected material and its derivation before constructing an external view.
Approved selected bytes and source ancestry.
Fix selected material and its derivation before constructing an external view.
Selected bytes, origin and holder.
Derived inputs and exact source closure.
The proposed view retains its exact source and holder basis.
Prepare an inspectable bounded projection under its supported profile.
Fixed source basis and exclusions.
Prepare an inspectable bounded projection under its supported profile.
Relevant local material and exclusions.
Exact proposed task view; larger source stays local where supported.
The supported profile determines what can remain local and what may be presented.
Compare new disclosure with prior and reserved exposure to the same observer.
Exact bounded task view and destination.
Prior and still-reserved units for the canonical observer.
Fixed combination rule and observer identity basis.
Exposure and uncertainty stay in observer history.
Compare new disclosure with prior and reserved exposure to the same observer.
Aliases share observer identity.
Prior/reserved units and prohibited combinations.
A new disclosure is evaluated together with prior and reserved exposure to the same observer.
Bind the admitted request before transport under all applicable limits.
Exact candidate with cumulative exposure result.
Matching approval, scope and expiry.
Actual supported route, stop state and cumulative limits.
Bind the admitted request before transport under all applicable limits.
Request, model, policy, expiry and stop state.
Call/byte bounds and approval; no grant inferred from available budget.
Available budget does not supply approval; all binding limits still apply.
Carry the reserved work through the specifically qualified route.
One admitted reserved binding.
Carry the reserved work through the specifically qualified route.
Declared client/tool/descendant scope.
The original effect identity.
The transport is limited to its specifically qualified route and original effect identity.
Retain the returned artifact, source lineage and uncertain effects.
Original issued identity.
Actual receipt/artifact or uncertainty.
Retain the returned artifact, source lineage and uncertain effects.
Received result and exact source/task binding.
Independent receipt or unresolved uncertainty without exposure refund.
Uncertainty does not refund possible exposure or authorize a replacement effect.
Approved selected bytes and source ancestry.
Fix selected material and its derivation before constructing an external view.
Selected bytes, origin and holder.
Derived inputs and exact source closure.
The proposed view retains its exact source and holder basis.
Fixed source basis and exclusions.
Prepare an inspectable bounded projection under its supported profile.
Relevant local material and exclusions.
Exact proposed task view; larger source stays local where supported.
The supported profile determines what can remain local and what may be presented.
Exact bounded task view and destination.
Prior and still-reserved units for the canonical observer.
Fixed combination rule and observer identity basis.
Exposure and uncertainty stay in observer history.
Compare new disclosure with prior and reserved exposure to the same observer.
Aliases share observer identity.
Prior/reserved units and prohibited combinations.
A new disclosure is evaluated together with prior and reserved exposure to the same observer.
Exact candidate with cumulative exposure result.
Matching approval, scope and expiry.
Actual supported route, stop state and cumulative limits.
Bind the admitted request before transport under all applicable limits.
Request, model, policy, expiry and stop state.
Call/byte bounds and approval; no grant inferred from available budget.
Available budget does not supply approval; all binding limits still apply.
One admitted reserved binding.
Carry the reserved work through the specifically qualified route.
Declared client/tool/descendant scope.
The original effect identity.
The transport is limited to its specifically qualified route and original effect identity.
Original issued identity.
Actual receipt/artifact or uncertainty.
Retain the returned artifact, source lineage and uncertain effects.
Received result and exact source/task binding.
Independent receipt or unresolved uncertainty without exposure refund.
Uncertainty does not refund possible exposure or authorize a replacement effect.
An illustrative method for comparing a selected view with observer history before an external task.
Additional mechanisms are described in the technical reference.
What the research says
Each card quotes the regulation, then names what the stack has built for it and its state.
Fines for prohibited practices reach €35M or 7% of worldwide turnover, whichever is higher; for SMEs, the lower of the two applies. Those maximums cover prohibited practices, not every gap in a log. For most deployers the practical test is an auditor’s question, so the stack writes the answer as it runs: who acted, under whose approval, with which version, and when.56
Regulation (EU) 2024/1689.
The digital omnibus moves the high-risk dates.
Marking of AI-generated content is postponed to 2 Dec 2026 for systems placed on the market before 2 Aug 2026.
The postponed marking duty also applies from this date to systems already on the market.
Record-keeping, log retention and human oversight apply to these systems.
The general counsel decides what to disclose; the record supplies entries, proofs and times. Illustrative.
01
Each entry carries the SHA-256 hash of the entry before it, so anyone who recomputes the chain sees a changed or missing entry.
02
An RFC 6962 Merkle tree gives each entry an inclusion proof, and consistency proofs show the record only grew between two checkpoints.
03
An Ed25519 signature over the chain head shows continuity. Proving who holds the signing key waits for external key custody, which is in development.
04
RFC 3161 timestamps fix when a chain head existed, so an auditor can check a date outside the system.
05
Records stay at least 180 days and at most 400 days: the AI Act’s six-month floor and the GDPR’s storage limit in one policy.
06
Personal data is encrypted under its own key. Destroying the key erases the data, and a tombstone keeps the chain verifiable.
| Obligation | Mechanism | Evidence kept |
|---|---|---|
| Record events automatically over the system’s lifetime (Art. 12)1 | Hash-chained, Merkle-anchored entries for every consequential action, with RFC 3161 timestamps | Entries, inclusion and consistency proofs, signed chain heads |
| Keep logs for at least six months (Art. 19, Art. 26(6))2 | Retention policy with a 180-day floor and a 400-day ceiling | Each record, held inside that window |
| Let people oversee, override and stop the system (Art. 14)3 | Approval before consequential actions; autonomy ceiling by risk class | Approvals and refusals with the approver; Art. 14 is not yet in the record’s framework map |
| Tell people they are interacting with an AI system (Art. 50(1)) | Disclosure in chat; disclosure by agents acting elsewhere is in development | Chat disclosure mapped to Art. 50(1) in the record |
| Erase personal data on request (GDPR) | Erasure by key destruction, with a tombstone | A chain that still verifies after the erasure |
| Explain the AI system’s role in a decision (Art. 86)4 | The record holds what an explanation draws on; producing it for the affected person is being built | Agent version, sources, approver and time for each entry |
Illustrative situations, connected to the mechanisms that make them possible.
Institutions and R&D teams
A useful model needs context, while the organization holds a much larger sensitive source.
Industrial edition
An operational process must react to observations while respecting equipment and responsibility boundaries.
Organizations commissioning custom AI
A difficult brief spans data, software, interfaces and operating responsibilities.
A regulated workflow needs to retain who can approve an effect and what the system was allowed to do. Manna’s envelopes and gates, Synapse’s source/effect admission and Agency Engine’s sealed governance set give those decisions explicit places in the architecture.
Human approval, stop and recovery are operating paths with recorded state. They are distinct from a policy sentence added to a model prompt.
Immutable source identity, effect identity, policy/runtime binding and event records connect a model task to the configuration that produced it. Black Box’s delivery design carries independent acceptance and the artifact into handover.
The record supports reconstructing the basis of a decision, an interruption and the next operating action. Technical documentation can refer to those actual controls and their qualified deployment scope.
Synapse accounts for cumulative observer exposure before transport. Agency Engine preserves approval and connector/store-custody boundaries. Metahive defines compute placement and operating responsibility.
The commissioned architecture maps intended purpose, data sources, human roles, monitoring and security controls to the particular deployment. The legal classification and applicable obligations are determined for that use case.
Naevum brings authority, oversight, traceability and risk boundaries into selected system designs from the outset. Each engagement has an explicit technical and operational responsibility map.
The phrase does not imply automatic conformity assessment or blanket certification. Applicable obligations and evidence are assessed with the responsible organization against the current EU AI Act.
The technical choices that give Built for the EU AI Act its particular behavior.
Read the legal framework at EUR-Lex: Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744.
If an auditor asked tomorrow what one of your agents did on 3 March, what could your general counsel hand over?
Discuss your AI Act obligations