The compliance gap is structural, not documentary
The EU AI Act assigns obligations to high-risk AI systems that assume the system has parts. Article 13 requires transparency sufficient for a deployer to interpret a system's output and use it appropriately. Article 14 requires that natural persons can oversee the system, understand its capacities and limitations, correctly interpret its output, and intervene or interrupt it. Article 15 requires accuracy, robustness, and cybersecurity across the lifecycle. Article 12 requires automatic logging of events over the system's operation. Article 9 requires a continuous, iterative risk management process running throughout the lifecycle.
Every one of these obligations presupposes an internal referent: a decision to interpret, a state to oversee, an event to log, a risk posture to manage. A model that maps a prompt to a probability distribution over tokens has none of these referents. There is no place in its forward pass where a deviation is recorded, no field an overseer can read to know whether the system currently considers itself competent to act, and no commit point at which an admissibility decision is made and stored. Behavioral alignment techniques shape the output distribution, but they do not add the missing referents. They make the system more likely to behave well without making its behavior inspectable, governable, or independently auditable. The result is compliance by documentation: a paper trail asserting properties that the architecture cannot exhibit on demand.
Closing this gap requires moving the compliance-relevant facts out of the documentation and into the running system. That is what a structural cognitive architecture does.
What the architecture adds: inspectable cognitive state, a governance gate, and a lineage record
The Human-Relatable Intelligence architecture disclosed in United States Patent Application 19/647,395 represents an agent as a persistent object built from a fixed schema of cognitive domain fields rather than as a stateless inference call. The base schema comprises an intent field, a context block, a memory field, a policy reference field, a mutation descriptor field, and a lineage field; the disclosure adds further structural fields including an affective state field that encodes valence-weighted feedback from prior outcomes, an integrity field that tracks whether the agent's actions remain aligned with its declared values, a confidence field, and a capability field. Each named field occupies a defined position and, per the specification, is independently readable, writable, and auditable.
Three structural facts about this representation do the regulatory work:
- Cognitive state is inspectable. Because each domain field is a fixed-schema, independently readable structure persisted with the agent across execution cycles, delegation events, and substrate migrations, an external party can read the agent's current risk sensitivity, confidence, and integrity trajectory at any moment. There is a state to oversee.
- Every action passes a governance gate before commitment. The disclosure defines a governance gate as a composite admissibility evaluation that integrates signals from a plurality of cognitive domain fields, including the integrity field, the confidence field, the affective state field, the capability field, and the personality field, to produce an admissibility determination for each proposed mutation before it is committed. There is a decision process to interpret and a point at which a human can intervene.
- The lineage field is a native audit trail. The lineage field records the complete history of the agent's state evolution, including every mutation, delegation event, and governance decision, and integrity evaluations are themselves recorded back into the lineage. The specification describes this as a self-reinforcing auditability structure. There is an event sequence to log, produced as a byproduct of execution rather than bolted on afterward.
Critically, the specification enforces that affective modulation is not an input to the governance gate: even an agent whose affective state produces maximal confidence disposition and minimal risk sensitivity is still independently constrained by the gate. Disposition cannot override governance. That separation is exactly the property an oversight regime needs.
Mapping AI Act obligations to structural primitives
The architecture is organized, in Chapter 14 of the specification, around ten conditions for human-relatable behavior, each implemented by a disclosed primitive. Several of these conditions map almost one-to-one onto AI Act obligations. The table below is an enabling map, not a legal opinion; it shows which primitive supplies the structural referent each obligation presupposes.
| AI Act obligation | Structural primitive (per 19/647,395) | What it makes verifiable |
|---|---|---|
| Article 14, human oversight and the ability to interrupt | Confidence-governed execution: execution is treated as a revocable permission that is continuously re-evaluated and withdrawn when assessed sufficiency degrades | An overseer can read the confidence field and the system can itself withhold action; there is a documented basis for interruption |
| Article 13, transparency and interpretability of output | Inference-time governance: each candidate inference transition is evaluated for semantic admissibility against persistent cognitive state before commitment | Every committed output traces to an admissibility decision against a readable state, rather than to an opaque sampling step |
| Article 9, continuous risk management | Integrity tracking and the deviation function: the agent records deviations from declared values and propagates corrective pressure | Risk is a maintained state with a recorded trajectory, evaluated continuously rather than at a single conformity checkpoint |
| Article 12, record-keeping and automatic logging | Lineage field: a cryptographically provenanced record of every mutation, delegation, and governance decision | Logs are generated structurally and are tamper-evident, not reconstructed after the fact |
| Article 15, accuracy and robustness | Capability-aware executability: the agent computes whether execution can structurally occur given substrate-advertised conditions, distinguishing permission to act from ability to act | The system does not commit to actions that are structurally impossible under current conditions |
| Provenance and data-governance expectations | Training-level governance: the depth and selectivity of knowledge aggregation are controlled by semantic metadata, with cryptographic provenance linking updates | Knowledge provenance is controlled rather than coincidental |
The point is not that the architecture is purpose-built for any single regulation. It is that the same structural primitives that make an agent's behavior human-relatable, namely the ability to abstain, to record deviation honestly, to forecast before acting, and to treat execution as revocable, are the same primitives a regulator needs in order to oversee, interpret, log, and intervene. Structure built for one yields the other.
Embodiments and deployment options
The architecture is enabling across a range of deployments rather than a single instance. The specification describes embedded, co-resident, and hardware-assisted configurations, each connecting cognition to an admissibility gate, which a deployer can select according to latency, isolation, and certification needs:
- Embedded configuration. The cognitive state fields and governance gate run in-process with the model serving stack. This suits a high-risk system already operated within a controlled boundary, where the audit consumer is the operator's own compliance function and the lineage record feeds an internal logging pipeline.
- Co-resident configuration. The governance and lineage subsystem runs alongside the model as a separate component on the same host, isolating the audit trail and the admissibility decision from the inference engine. This supports an oversight model in which the party reading cognitive state and the party generating tokens are organizationally distinct, which is useful where a deployer must demonstrate independence between the system and its overseer.
- Hardware-assisted configuration. The admissibility gate and provenance record are anchored to a hardware trust boundary, strengthening the tamper-evidence of the lineage log. This addresses the cybersecurity and robustness expectations of Article 15, the Cyber Resilience Act, and common-criteria-style certification, where the integrity of the audit record itself must be defensible.
Across configurations, the deployer can expose the agent's readable cognitive state fields to a human overseer (satisfying the read side of Article 14), wire the lineage field into a regulated logging store (Article 12), and treat the governance gate's admissibility determinations as the documented, reviewable decision record (Articles 9 and 13). An implementer building any of these is working from disclosed mechanisms: the field schema, the composite governance gate, the deviation function for integrity, the confidence governor, the capability envelope, and the provenance-bearing lineage. No new mechanism beyond the disclosure is required to stand up a compliant deployment; the work is integration, not invention.
Why this matters for future-proofing
Regulatory requirements will tighten and proliferate. A deployer who satisfies them by documentation must re-document with each revision and must hope an auditor accepts assertions about an opaque core. A deployer whose system carries inspectable cognitive state, gates every action, and emits a tamper-evident lineage answers new oversight, logging, and risk questions by reading fields that already exist. The compliance posture is a property of the architecture, so it persists as the regime evolves. That is the durable advantage of building cognition from structural primitives rather than shaping the behavior of a structureless model.
Disclosure Scope
This article is a domain application of the Human-Relatable Intelligence architecture disclosed in United States Patent Application 19/647,395. The structural primitives described here, namely the seven-field cognitive schema, the composite governance gate, the integrity field and deviation function, the confidence governor, the capability envelope, inference-time and training-level governance, and the cryptographically provenanced lineage record, are disclosed in that application. The mapping of those primitives to the European Union AI Act, the General Data Protection Regulation, the Cyber Resilience Act, the NIS2 Directive, and related instruments is offered as a faithful application of the disclosed technology and is not legal advice. This article is published as a dated public disclosure tied to that application.
