AI REGULATION × FOL / SOVEREIGN AGENT LAB
● APPENDIX · ADVISORY EXECUTIVE BRIEF · USEFUL, NOT SUFFICIENT
AI REGULATION × FOL // EXECUTIVE ADVISORYFIVE LAYERS L1 → L5 VERDICT: USEFUL, NOT SUFFICIENTZ3 × CVC5 × SOUFFLÉ × LEAN × ACL2 SIMULATOR ≠ LEGAL COMPLIANCEA LAYER-SPECIFIC BLOCK ≠ A COMPLIANCE CERTIFICATE AI REGULATION × FOL // EXECUTIVE ADVISORYFIVE LAYERS L1 → L5 VERDICT: USEFUL, NOT SUFFICIENTZ3 × CVC5 × SOUFFLÉ × LEAN × ACL2 SIMULATOR ≠ LEGAL COMPLIANCEA LAYER-SPECIFIC BLOCK ≠ A COMPLIANCE CERTIFICATE
APPENDIX // ADVISORY // WHERE FORMAL METHODS HELP — AND WHERE THEY DO NOT

USEFUL, NOT SUFFICIENT.

Formal methods are useful in AI regulation, but not sufficient. This page separates the compliance-stack layers that admit mechanical verification from those that require post-deployment monitoring, attestations, and human judgement — and explains where formal tools fit in AI-related financial regulation, using financial compliance as the worked example.

← BEFORE FOL-LAB M1 STACK L1→L5 TOOLS Z3 · CVC5 · SOUFFLÉ · LEAN · ACL2 NEXT → NVIDIA × FOL NEXT → revolut.hot
SCOPE · SPEC · PROPOSED

A NEUROSYMBOLIC COMPLIANCE STACK — FROM SILICON TO NEURAL.

Where deterministic verification helps, where it fails, and how to integrate symbolic solvers into production AI systems to help address regulatory and auditability requirements — with explicit boundary conditions, and with nothing in this stack claiming to be a legal-compliance certificate.

00

The 2026 inflection

WHY NOW · MODEL CLAIM

The market has shifted. Through 2025 the gap between proprietary and local models was still obvious; through 2026 it has narrowed to the point where frontier-class inference is plausibly deployable on local AI endpoints such as the Olares One or the DGX Spark. Concrete path: Qwen3-Coder 30B A3B MoE is already at parity on many SWE-bench style tasks. For a bank, this changes the calculus from “which vendor do we trust with our data” to “which controls can we prove locally, and what still requires external assurance?”

01

The compliance stack

SELECT A LAYER
02

The verdict

WHAT THIS STACK IS — AND IS NOT
ADVISORY VERDICT · CLAIM · ATTRIBUTED TO THIS PAGE'S ANALYSIS

FORMAL METHODS ARE USEFUL IN AI REGULATION — BUT NOT SUFFICIENT.

What the stack is: a layered engineering proposal that combines formal verification (L1–L3), symbolic runtime governance (L4), and continuous neural decision-making (L5), with the explicit purpose of improving auditability, deterministic replay, and mechanised policy checking.

What the stack is not: a legal-compliance certificate. L1–L3 proofs and properties do not, by themselves, establish L4–L5 policy or legal compliance; those layers require post-deployment monitoring, attestations and human judgement.

03

Proof vs policy

THE HARD LINE

▚ L1 / L2 — FORMAL PROOFS

  • ACL2 — floating-point correctness (e.g. AMD FLD70 result: ~0.1% corner-case residual).
  • seL4 — functional correctness proofs, but limited to the specified property and model of the system.

▚ L4 / L5 — POLICY COMPLIANCE

  • Not equivalent. Proofs at L1/L2 do not by themselves establish L4/L5 policy or legal compliance — those layers require monitoring, attestations, human judgement.
CASE STUDY · WHY THE DISTINCTION MATTERS
2025 (speculative analogy): an L4 policy gate is intended to block prohibited model actions at runtime.
An adversarial prompt jailbreaks the L5 neural layer and bypasses the L4 gate.
The developer detects the breach, halts the session, and re-traces — effectively recalling that batch decision.

Is “we recalled the batch” a verified property of the overall system? No.
It is a post-deployment remediation: valuable, necessary — but not a formal guarantee.

The correct regulatory reading: L4 monitoring plus incident response supports Art. 73 EU AI Act-style
incident reporting and Art. 72 post-market monitoring; it does not substitute for them.
If a regulator asks “what did you verify, and what did you monitor?”, the answer must distinguish
verified properties from monitored behaviours.
04

Logic spectrum

EXPRESSIVENESS × TRACTABILITY · ILLUSTRATIVE

ILLUSTRATIVE POSITIONS — NOT BENCHMARKS. THE SAME FORMALISM CAN LAND IN DIFFERENT REGIONS ONCE OPTIMIZED.

05

Symbolic policy simulator

INTERACTIVE · DETERMINISTIC · NOT LEGAL COMPLIANCE
O = p·x + q·y ≤ B where p, q are permitted usages, x, y are usage quantities, and B is the allowed budget. Sliders set x, y; buttons set B. Evaluate checks each constraint separately and logs every decision.
0.30
0.20
∀ action a: (P(a) ∨ D(a)) → BLOCK(a) Literals: P = political-advice content; D = data-fabrication content. If no literal holds, allow.
derive k from TTL: k = 1 − (hours_since_issue / TTL) then check: k·a ≤ b where a = requested access level, b = bound.
6
8
5
5

DETERMINISTIC SIMULATION ONLY. IT IS NOT LEGAL COMPLIANCE. A LAYER-SPECIFIC L4 BLOCK IS NOT A COMPLIANCE CERTIFICATE.

06

Tooling catalog

FILTER BY PARADIGM
ToolParadigmRegulatory MappingComputational Complexity
07

Phased roadmap

INDEPENDENT TRACKS
PHASE P0 · FEB 2026

Feasibility

Z3/CVC5 constraint checks on synthetic policy sets + Soufflé prototype for training-data lineage on a small graph.

PHASE P1 · MAY 2026

Pilot

L4 governance gate on a narrow model task (transaction monitoring triage); L3 Datalog provenance integrated with MLOps metadata store; baseline metrics + first attestations.

PHASE P2 · AUG 2026

Hardening

Expand L4 rule coverage and move selected L1 accelerator checks into CI; security review of enclave integration; add countersigning to lineage graph.

PHASE P3 · NOV 2026

Scale

Rollout of verified gate policy to production endpoints; quarterly formal re-verification; regulator-facing technical documentation of guarantees.

08

Decision protocol — mechanise the check, never the judgement

CI GATES

▚ BLOCKER

Hard regulatory requirements (e.g., Art. 9 EU AI Act data-governance failures) block deployment if unsatisfied — no human override.

▚ MANDATORY

CI runs Z3/CVC5 on policy constraints; failures must be resolved or explicitly waived by a named accountable human.

▚ ADVISORY

Recommendations from L4 (e.g., "add counterexample to audit log") inform review; no direct block.

09

Reading the stack honestly

A FAIR READING

The stack is honest about limits. It does not claim that formal verification solves AI regulation; it claims that formal tools, applied at the right layers (L1 hardware, L3 provenance, L4 runtime policy checking), can materially improve auditability and deterministic replay. The neural layers (L5) still require post-deployment monitoring and human judgement. A fair reading: this is a useful, layered, engineering-first approach to AI governance — not a silver bullet.

10

Boundaries & context

WHERE THIS PAGE ENDS

MODEL CLAIM  Future directions. Extending the Datalog engine (L3) to continuous constraints; replacing threshold circuits in L1 with formal arithmetic circuits; integrating Lean proofs of L4 policy properties into release sign-off. All are speculative extensions.

Source code & lab notes: experiments and infrastructure live at /root/fol-lab on the Olares One home lab.

RECEIPT · FACT  Live use cases. This advisory stays general on purpose. Executed, narrower demonstrations live in their own case files:

FOL-LAB // M1 ↗ LC01 · decision plane ↗ Nemotron × FOL ↗ Working note ↗ LC02 · provisional authority ↗
11

References

PRIMARY SOURCES ONLY
DOI + PRIMARY LINKS
https://doi.org/10.6084/m9.figshare.33427318 — DOI-preprint (this analysis, with corrections)
arxiv.org/abs/2107.08209 — L1 formal verification of commercial FPUs (ACL2)
sel4.systems — L2 functional correctness (seL4)
microsoft.github.io/z3guide — L4 SMT constraint solving (Z3)
arxiv.org/abs/2402.01635 — neural-symbolic verification (Marabou)
eur-lex.europa.eu/eli/reg/2024/1689/oj — EU AI Act (Reg. 2024/1689)
ffiec.gov/press/pr083114.htm — vendor guidance (FFIEC)

NOTE: RELIANCE ON FFIEC GUIDANCE IS CONTEXTUAL, NOT ENDORSEMENT.