by SLAVA SOLODKIY.cv / LIVE CASE 01
Google Cloud × NVIDIA Recorded run verified Synthetic data only #NVIDIAGTC

The decision stays local.
The cloud witnesses the receipt.

NVIDIA Nemotron proposes candidate evidence on a local Olares system. Deterministic verification, formal logic and policy-as-code decide. An optional Google Cloud Storage witness preserves a minimal post-decision receipt and verifies the exact stored generation by SHA-256.

Open GitHub repo ↗ Original architecture Nemotron × FOL experiment Potential use-case: Revolut
Local AI decision path
01
NVIDIA Nemotron perception Local candidate extraction from recorded synthetic evidence.
↓
02
Deterministic verification MRZ structure, check digits, syntax and cross-field consistency.
↓
03
SWI-Prolog / FOL Verified facts become formal decisions — not model prose.
↓
04
OPA / Rego institutional policy ALLOW or ESCALATE_TO_HUMAN. AI override remains false.
Local authority The decision core remains on-device / local.
→
Minimal receipt only No passport image. No extracted identity fields.
Local vs Cloud in Compliance :: for the Google route

One cloud layer.
No change in authority.

The Google extension was deliberately narrow: preserve a minimal receipt externally without moving sensitive perception (GDPR, etc) or decision authority into the cloud.

NV

Local NVIDIA model

Nemotron remains the perception and explanation layer in the recorded local experiment. It proposes candidate evidence; it does not decide.

FOL

Deterministic decision core

Python verification, SWI-Prolog and OPA/Rego remain the layers that determine whether evidence earns ALLOW or requires human escalation.

GCS

Additional trust-layer (to local AI): cloud witness

A post-decision receipt is written to Google Cloud Storage, then the exact object generation is read back and hashed.

Authority boundary

Different jobs.
Different trust.

This is the point of the experiment: use local models and cloud services where they are useful without silently granting either of them final authority.

Local AI: decision authority

NVIDIA + deterministic stack

✓Nemotron proposes candidate observations.
✓Python verifies machine-checkable evidence.
✓SWI-Prolog derives the formal decision.
✓OPA/Rego maps the decision to institutional action.
✓AI override remains false.
External: storage witness

Google Cloud Storage

✓Receives only a minimal synthetic decision receipt.
✓Create-only semantics prevent silent overwrite of the content-addressed object.
✓Exact object generation is read back after upload.
✓Downloaded bytes are matched against the local SHA-256.
✓Does not certify whether the decision itself is correct.
Recorded witness

Enough to inspect.

(Useful for inter- and correspondent- banking compliance communication.) The public repository records the receipt and a sanitized result from the real Google Cloud Shell run.

StatusGCS_READBACK_VERIFIED
Stored receipt545 bytes
GCS generation1788942747393502
Created2026-09-09T08:32:27.399Z
Readback verified2026-09-09T08:32:28.058972+00:00
Receipt SHA-256771cbab54cd029e8c65ad520e1bdc72ce219cc14eaeb6a491edef8bed1999381
Source evidence commit6d60a418725d313c66fb1eff3d40aa8ef2f4f724
Public cloud accessDisabled / private bucket
The model proposes. Evidence decides.
The cloud preserves the receipt.

Research demonstrator only. Synthetic identity data. Not a production KYC/AML system, not regulatory approval, not a compliance certification, and not a claim that Google Cloud independently validates decision correctness.

Open artifacts

Inspect the evidence trail.

The public replay does not re-run Nemotron; it starts from the captured candidate evidence and reproduces the deterministic decision layers.

GitHub — Sovereign Decision Plane / Live Case 01 ↗ Code, CI, receipts, sanitized GCS witness result and reproducibility notes. Cloud Witness documentation ↗ Exact claim boundary, execution notes, tests and limitations. Original Sovereign Decision Plane page The broader local authority-separation architecture. Nemotron × FOL experiment The local multimodal experiment that produced the model-derived RED path. Potential use-case: Revolut