Sentinel — A Methods & Positioning Note

Christopher I. V. Farmer, LL.M. (University of Worcester) · Independent Researcher · civfarmer@gmail.com

Positioning note. Version 1.0, July 2026. Companion to the Sentinel readiness assessor. Law stated as-of July 2026. Decision-support, not legal advice.

Scope note. This note explains what Sentinel is, how it reaches its conclusions, and — just as importantly — what it is not. Sentinel is an offline decision-support instrument, not a legal adviser and not an official export-control authority. Every position it takes is dated and labelled; none of it should be relied upon without confirming the current primary sources.

The problem: two legal questions, answered in one pass

Any serious AI or autonomous-system programme has to satisfy two distinct bodies of law that are almost never assessed together. The first asks whether it is lawful to build and use the system — its risk tier under the EU AI Act (Regulation (EU) 2024/1689) and the obligations that tier triggers. The second asks whether it is lawful to move the system, its source, its model weights, or its underlying compute across a border — the EU Dual-Use Regulation (Regulation (EU) 2021/821) and the Wassenaar Arrangement control lists it implements, with US EAR/ITAR exposure layered on top.

In practice these are handled by different teams, different advisers, and different timelines, and the seam between them is exactly where a compliant-looking product becomes an unlawful export — or where an export-cleared item quietly breaches a prohibited-practice rule. Sentinel's premise is that the two questions draw on almost identical inputs — what the system does, its data and autonomy, its sectors, and the jurisdictions of build and use — and can therefore be answered together, in a single structured pass, early enough to change the design.

The method: a deterministic rules engine

Sentinel is not a model and makes no probabilistic prediction. It is a deterministic rules engine — a set of pure functions that map a structured description of a deployment onto the applicable legal categories, byte-for-byte reproducibly and entirely offline, with no network calls. The AI Act classifier walks Article 5 (prohibited practices), Annex III and Annex I (high-risk), and Article 50 (limited-risk transparency), returning exactly one of four tiers — prohibited, high-risk, limited-risk, minimal-risk — together with the specific article or annex that fired. A parallel export screen tests the same facts against the Dual-Use Regulation's Article 4 general catch-all and Article 5 cyber-surveillance catch-all, the Annex I control-list categories, and the Wassenaar Category 3/4/5 controls, with an honest, unresolved flag wherever US EAR or ITAR could bite extraterritorially. The two results are combined into a single verdict — deployable, deployable with controls, blocked, or more information needed — plus an exportable, audit-ready control record.

Every line is tagged — and dated

The engine's defining discipline is that no assertion is left unlabelled. Every reason it emits carries one of four tags, so the reader can always see where the law ends and the reasoning begins:

Every position is dated to July 2026 and pinned to the 2025–26 Digital Omnibus — the amending package adopted by the European Parliament on 16 June and the Council on 29 June 2026 and signed on 8 July 2026, which moves the Annex III high-risk obligations to 2 December 2027 and adds a new Article 5 prohibition (AI generating non-consensual intimate imagery or CSAM) from 2 December 2026. Because that package was, as of writing, adopted but not yet in force — awaiting Official-Journal publication — the lines that depend on it are tagged verify, not legal fact. That is the whole point of the tagging: the tool refuses to launder a pending date into a settled rule.

The thesis: regulated by design

Sentinel's normative claim is that governance is not a disclaimer appended to a finished product but a design layer built into it — regulated by design. The same engine that classifies the system also names the controls that make the answer defensible: meaningful human authority (a real override, pause, and stop authority proportionate to the system's autonomy — Art 14 / Art 26); immutable auditability (tamper-evident event logs, decision provenance, and, where triggered, a Fundamental Rights Impact Assessment under Art 27 or a GDPR DPIA under Art 35); and export screening wired into the control layer rather than bolted on at the loading dock. A system built this way is not merely compliant on the day it ships — it can prove it was governed, which is the question a regulator or a court actually asks after the fact.

Scope and limits

Sentinel is decision-support, not legal advice, and nothing it returns is an official control classification. It creates no lawyer–client relationship, binds no authority, and does not substitute for a formal export-control ruling (BIS / BAFA / ECJU) or a qualified legal opinion. Its outputs are only as good as the description entered; the law moves; and the verify-tagged lines mark exactly where the user must consult a primary source or a competent adviser before relying on the result. It is offered as a structured first pass — a way to see both legal questions at once, while the design can still be changed — not as the last word on either.

— Christopher I. V. Farmer, LL.M. (University of Worcester) · Independent Researcher · civfarmer@gmail.com