GOVERN THE DECISION WHILE IT HAPPENS.
Use case · AI / regulation / control plane. This note compresses three public artifacts — built during my September 2026 week sprint — into one claim. It is an outsider working note for the unofficial & personal book project revolut.hot. Speculation is labelled.
NOT “EXPLAINABLE AI.”
RUNTIME GOVERNANCE.
The interesting problem is not “explainable AI”. It is whether a licensed digital bank can govern an automated decision while it is happening: which rule applies, who owns the override, what evidence survives, and what changes when the model, the product, or the law moves.
The claim
DEFINITION, NOT SLOGANCall that capability algorithmic sovereignty: not autarky, not “run every model on-prem”, and not a proof that the institution is legally compliant. It is the practical ability to constrain execution and show the trace.FRONTIER MODELS CAN STILL DRAFT, RETRIEVE, AND REASON. THEY SHOULD NOT BE THE LAST AUTHORITY ON A CREDIT, FRAUD, OR FINANCIAL-CRIME ACTION.
What the public record already forces
TWO DIFFERENT MACHINES▚ EU PATH RECEIPT · FACT
Still a horizontal risk statute — Reg. 2024/1689, plus the later Omnibus timetable. Not every banking model is Annex III high-risk: fraud systems sit in a different pocket from creditworthiness tools. Nightly retraining can still become a “substantial modification” problem if the same weights leak into a high-risk pipeline.
▚ UK PATH RECEIPT · FACT
Still sectoral conduct: Consumer Duty, SM&CR, PRA model-risk language. The United Kingdom has not passed a general AI Act for financial services.
Those are different machines. A bank that treats “AI policy” as one memo will fail one of them.
These corrections are in the DOI-preprint ↗ — they matter more than the neon.
A control plane, not a religion
THREE LAYERSModels propose
Score, draft, retrieve, cluster, write a first brief.
Deterministic layer evaluates
Structured facts against a versioned rule. SAT / fail-closed / escalate.
One packet, not a slide
Inputs, model version, policy version, reason codes, human override, outcome.
Datalog, policy-as-code, SMT, and proof assistants belong in layer 02 and in offline audit of layer 03. They do not belong on a 50 ms fraud path unless the rule is a cheap threshold. They do not translate “reasonable steps” into a theorem. FOL-page verdict: useful, not sufficient.
Three public artifacts → one working note
SEPTEMBER 2026 SPRINTLet's add runtime authority
Bounded, provisional, revocable authority — where the action is reversible.
What would count as grown-up
A MEASURABLE SPLIT OF WORK▚ CHEAP-ENOUGH TASKS
Tasks where a cheaper model is already enough: boilerplate, first draft, routing.
▚ LARGE-SET QUALITY
Tasks where quality is judged on a large held-out set — tens or hundreds of thousands of cases — not a demo.
▚ HOURS TO “SAFE TO SHIP”
Tasks where the scarce resource is not tokens but review, reason codes, and the incident path.
If open-weight models keep losing on the large-set comparison, that is not an argument against a policy layer. It is an argument that the policy layer should sit on top of whatever model currently wins — including a frontier API and an in-house fine-tune.
Sources in this bundle
BOOK-PROJECT CHAINThe NVIDIA Innovator's Dilemma → Europe's First Trillion-Dollar Startup? → FuturoShock.AI → + with a dash of compliance-background :) NEW: Algorithmic Sovereignty in Digital Banking — obligations → runtime controls.
Honest boundary. Independent commentary. Not legal advice. Not a Revolut document. Figures and product names that appear in the linked pages are public reconstructions and may be wrong in detail — treat them as hypotheses until confirmed by primary sources. Private conversations are excluded on purpose. Hugs, Slava Solodkiy. · g.dev/slavasolodkiy ↗