EASA / FAA Aviation
AC 25.1309 · DO-178C · Part 23 / 25 / 27 / 29 flight-critical loops
EASA AI Concept Paper Issue 02 (2023) + ML Type-Cert Roadmap. No neural network has been certified in a Part 23/25/27/29 flight-critical loop to date.
Verify · Standards & Compliance · Pillar 03
Standards-aligned
by construction.
AQEA's patent family is structurally positioned to address regulatory frameworks where classical ML, by construction, cannot meet the bar — flight-critical aviation, EU AI Act high-risk systems, FDA SaMD, IEC 61508 SIL 2/3, GDPR Art. 22, UN R157.
Compliance density
6
regulatory frameworks
mapped to patent claims
25
USPTO filings
across 6 patent tracks
4
patents → 6 articles
compliance density of Track-F + M33
NDA
counsel-reviewable
claim-by-claim, drawings + draft PDFs
01 · The structural compliance gap
Why classical ML cannot satisfy these standards.
Regulators do not ask for accuracy — they ask for predictability under audit. Non-deterministic GPU kernels, batch-norm running statistics, opaque post-hoc explanations and weight-update procedures that no predetermined-change-control plan can cover sit at the heart of every standard ML stack. None of these are bugs. They are architectural commitments that the regulatory frameworks below treat as disqualifying.
AQEA's patent family commits to the opposite architecture: frozen substrate, deterministic execution, byte-identical replay, structural verifier-monopoly, and per-decision counterfactual records. The same properties that classical ML accepts as cost are what the standards explicitly require.
02 · Patent × Regulation matrix
Which patent answers which clause.
Mapping is mechanism-level, not marketing-level. Each row points to a specific patent claim and a specific regulatory article. Counsel-reviewable claim-by-claim under NDA.
| Patent | Mechanism | Framework | Article / Clause |
|---|---|---|---|
| N7 | Verifier-Only Trust-State A separate verifier — not the proposing model — is the sole authorized writer of trust-state. Oversight is mediated by structure, not policy. | EU AI Act Annex III | Art. 14 — Human Oversight |
| N9a | Multi-Modal Response Gate Per-decision write-ahead log plus counterfactual after-action record. Every authorized output carries a regulator-defensible trail. | EU AI Act Annex III | Art. 12 — Record-Keeping · Art. 13 — Transparency |
| N9a | Fail-Closed Gate Graceful degradation by construction: failure of any condition routes to a lower-severity outcome, never to a silent block. | EU AI Act Annex III | Art. 15 — Accuracy / Robustness / Cybersecurity |
| N9a | Counterfactual AAR Per-decision counterfactual record answers the 'had this input been different' question that Art. 22 requires — and that post-hoc SHAP cannot. | GDPR | Art. 22 — Right to Explanation of Automated Decisions |
| N8 | Anomaly-Streak Gate Deterministic + byte-identical replay + static-inspection-verifiable. The properties that block classical ML from SIL certification. | IEC 61508 | SIL 2 / SIL 3 — Industrial Functional Safety |
| M33 | Substrate (Track A) Substrate is frozen. Updates happen in the dictionary — an audit-conform predetermined-change-control mechanism, not a re-training. | FDA SaMD · EU MDR Annex VIII | Predetermined Change Control Plan |
| N6 | Pre-Inference Pipeline Stage-gated, byte-identically-replayable contextualization. Line-by-line traceability is a construction property, not a documentation task. | DO-178C | DAL-A / DAL-B Traceability |
| N9a | Per-Event Logging Reason for activation / deactivation recorded per event, with deterministic counterfactual. The class of evidence underwriters can price. | UNECE WP.29 | R157 — Automated Lane-Keeping (ALKS Level 3+) |
Patent application numbers, draft PDFs and prosecution status available under NDA-Full with strategic buyers, patent counsel and compliance officers.
03 · Framework coverage
Six regulatory frameworks. One architecture.
EU AI Act
Annex III high-risk systems — Banking, HR-tech, Education, Critical Infrastructure, Law Enforcement, Justice
Articles 6 + 8–15 phase in 2026–2027. Every regulated AI system must demonstrate compliance with risk management, logging, transparency, oversight, accuracy / robustness.
FDA SaMD · EU MDR
Software-as-Medical-Device Class II / III · Annex VIII intended use
FDA Predetermined Change Control Plan Guidance (2024) requires predictable update behaviour — incompatible with neural-network weight updates by construction.
IEC 61508
Industrial functional safety, SIL 2 / SIL 3
Non-deterministic ML cannot be certified under SIL 2/3. Deterministic substrate + byte-identical replay + static-inspection are the load-bearing requirements.
GDPR
Art. 22 — algorithmic decision explanation across EU jurisdictions
Post-hoc explanation methods (SHAP, LIME) are widely regarded as rationalisation, not the 'logic involved'. Counterfactual records answer the actual question.
UNECE WP.29
R157 — Type-Approval for Automated Lane-Keeping Systems Level 3+
Per-event 'reason for activation / deactivation' recording is mandated. Opaque ML decisions are practically uninsurable at acceptable premiums.
04 · Honest framing
Where we are not the right tool.
AQEA's moat is safety-critical, deterministic, auditable. Where that is not the painpoint, we are not the right tool.
- ·General-purpose LLM assistants are not our category. We are not a foundation model and will not become one.
- ·Image generation and creative ML pattern do not match our architecture. We classify and decide; we do not generate.
- ·Consumer-recommendation and marketing-optimisation accept opacity. Our determinism / audit guarantees are overhead there, not value.
This honesty is part of the substance guarantee. The painpoints we list above are the ones where the architecture is load-bearing.
Compliance briefing under NDA.
Patent × clause mapping with application numbers, drawings and draft PDFs. For compliance officers, counsel and pilot engineering teams.