THE MODEL EXPLAINS — IT CANNOT OVERRULE.
Passport Evidence Integrity. A synthetic, locally executed demonstrator in which the model may perceive and explain — but cannot overrule failed evidence. Same decision plane, different evidence, different outcome.
SAME DECISION PLANE.
DIFFERENT EVIDENCE.
DIFFERENT OUTCOME.
The system has a valid ALLOW path. Nemotron-derived evidence simply failed to earn it — and the machine said so, on the record.
- human-readable fields: 9/9 correct
- MRZ integrity: FAIL
- override: NO
- control input: VALID
- MRZ integrity: PASS
- override: NO
The result
TWO PATHS · ONE PLANEThe system has a valid ALLOW path. Nemotron-derived evidence simply failed to earn it.
▚ RED · MODEL-DERIVED EVIDENCE NOT PROVEN
Nemotron correctly read the human-readable passport fields, but its MRZ transcription failed deterministic integrity checks — even after a dedicated enlarged MRZ crop.
Formal logic: request_better_evidence
OPA policy conditions: false · AI authorised to override: NO
▚ GREEN · PREVALIDATED SYNTHETIC CONTROL DEMONSTRATED
A deliberately prevalidated synthetic control passed the same downstream verification, formal-logic and institutional-policy stack.
Formal logic: accept_extraction
OPA policy conditions: true · AI authorised to override: NO
| RED · MODEL-DERIVED | GREEN · PREVALIDATED CONTROL | |
|---|---|---|
| Evidence source | Nemotron perception | Prevalidated synthetic input |
| Human-readable fields | 9 / 9 correct | VALID |
| MRZ integrity | FAIL | PASS |
| Formal logic | request_better_evidence | accept_extraction |
| Institutional action | ESCALATE_TO_HUMAN | ALLOW |
| AI can override? | NO | NO |
| Evidence packet | Generated + hashed | Generated + hashed |
The model reads. Deterministic code verifies. Formal logic decides what follows. Policy determines what may happen.THE MODEL EXPLAINS — BUT CANNOT OVERRULE.
Decision plane — six stations
EVIDENCE → RECORDWhat was actually executed
FOUR LAYERS + BOUNDARYNemotron 3 Nano Omni
Local, via Ollama on Olares One. Candidate extraction only — not final authority.
Python evidence checks
MRZ length, character set, check digits, cross-field consistency, uncertainty sanity.
SWI-Prolog / FOL
Converts pass / fail / conflict facts into an explicit formal decision.
OPA / Rego
Maps the formal state to ALLOW, ESCALATE_TO_HUMAN or BLOCK.
Authority stays outside
ai_authorised_to_override = false — in both the green and the red path.
One research program, not six random experiments
SPRINT 01–07 SEP 2026The useful part of the week was not any single model or solver. It was discovering where each kind of authority belongs.
FreeToken
Large MoE inference beyond nominal VRAM limits on consumer hardware.
FOL-Lab
Local LLM translation + deterministic Z3 / CVC5 / formal-method experiments.
Nemotron × Evidence Gate
The model reads; deterministic evidence refuses malformed confidence.
AI Regulation Navigator
A jurisdictional map of the regulatory surface around AI systems.
AI Regulation × FOL
Where formal logic helps with auditability — and where human/legal judgment remains mandatory.
Algorithmic Sovereignty
A runtime control-plane framing for regulated digital banking. P.S. human-factor & reality-checks — real-life compliance experience edits and add-ons.
The integration test
ESSAYS → EXECUTABLE PATTERNLive Case 01 is the point where the sprint stops being a set of essays and experiments and becomes one executable control pattern.
What this proves — and what it does not
READ BEFORE SHARINGDEMONSTRATED
- Perception without authority. A local multimodal model can be used as a perception layer without granting it final authority.
- Plausible ≠ valid. Machine-checkable evidence can reject a plausible but invalid model extraction.
- Both paths exist. The same downstream stack can produce an ALLOW path and an ESCALATE path.
- Separable roles. Formal logic and policy-as-code can be separated into distinct layers.
- Preserved trace. The final decision and artifact hashes can be preserved as an evidence-oriented record.
NOT PROVEN / NOT CLAIMED
- Not production. This is not a production KYC / AML system.
- Not legal advice. Not legal advice or regulatory approval.
- GREEN ≠ Nemotron win. The GREEN control is not presented as successful Nemotron MRZ extraction.
- Packet ≠ regulator report. An evidence packet alone is not a regulator-ready report.
- Formal ≠ legally complete. Formal correctness of a policy model does not prove the policy is legally complete or correct.
ALL IDENTITY DATA IN LIVE CASE 01 IS SYNTHETIC. THE GREEN PATH IS DELIBERATELY LABELLED A PREVALIDATED_SYNTHETIC_CONTROL. THE POINT IS NOT TO MANUFACTURE A SUCCESSFUL OCR BENCHMARK — IT IS TO DEMONSTRATE THAT MODEL OUTPUT MUST EARN AUTHORITY BY PASSING INDEPENDENT CONTROLS.
Product contract
THE MODEL MAY PROPOSE. IT CANNOT EXECUTE PAST A FAILED CONTROL.Source / reproducibility
PUBLIC BUNDLERED + GREEN inputs · deterministic validator · Prolog/FOL layer · Rego policy OPA decisions · evidence packets · SHA-256 manifest · run-demo.sh github.com/slavasolodkiy/sovereign-decision-plane-live-case-01
RELATED POSTS: @NansenID ↗ · @SolodkiyUK ↗
Bottom line
HARDWARE QUESTION → AUTHORITY QUESTIONThis week started with a hardware question and ended with an authority question.THE USEFUL ARCHITECTURE IS NOT “REPLACE THE INSTITUTION WITH AI”. IT IS THE OPPOSITE: MODELS DO THE PARTS THEY ARE GOOD AT, DETERMINISTIC TRUTH MOVES INTO INDEPENDENT CONTROLS, POLICY IS MADE EXPLICIT, THE TRACE IS PRESERVED — AND AUTHORITY STAYS OUTSIDE THE MODEL.