RegTech · enterprise
Every deliberation is an immutable, attributed record by construction. Human oversight, convergence, dissent and kill-switch traces are byproducts of how Noolog runs — not paperwork assembled for an auditor afterwards. Trust like that has to hold at two layers: the protocol that runs the deliberation, and the hardware it runs on. Two pillars, one guarantee.
Governance, human oversight and an immutable audit trail are properties of the deliberation protocol — mapped directly onto the EU AI Act, FCA, FINRA, MAS, GDPR and SOC2.
A tamper-responsive sealed enclosure with a certified integrity anchor. The record can’t be silently rewritten — because the hardware physically won’t let it.
The software layer. Because oversight, provenance and logging are how the deliberation runs, the regulatory artifacts fall out of it — no post-hoc paperwork. Here is the full mapping, by jurisdiction.
Every framework below reads across three beats: the requirement it imposes, what Noolog provides against it, and how — the concrete mechanism that makes the claim real. Deadlines are in the timeline above.
| Framework · requirement | What Noolog provides | How — the mechanism |
|---|---|---|
| EU AI Act Art. 14 Human oversight of high-risk AI | Attributed human-in-the-loop trace, by construction. | HITL trace · kill-switch · chained recordNamed operators inject / redirect / approve / reject / kill every round; each action attributed to an individual and written to the tamper-evident chain. |
| EU AI Act Art. 12 Transparency & record-keeping | Immutable, attributed transcript of every deliberation. | Immutable chained recordNATS JetStream audit trail anchored to a Poseidon-BN254 commitment chain; per-session exportable transcript. |
| EU AI Act Art. 26 Log retention for deployers | Configurable, timestamp-verified retention you own. | Immutable chained recordImmutable logs export to your archive; a shared chain_head makes any later edit detectable. |
| EU AI Act Art. 50 Content provenance (C2PA) | A signed C2PA provenance manifest per output. | C2PA manifest · signed audit envelopeEach output carries models, round, operator and inputs in a DualSigner (Ed25519 + ML-DSA-65) AuditEnvelope. |
| FCA SM&CR Senior-manager accountability | Named-operator audit trail at every decision point. | HITL trace · kill-switchEvery decision attributed to a person; halting a session emits a full kill-switch trace. |
| FCA Consumer Duty Good outcomes for consumers | Convergence and dissent captured per round. | Convergence + dissent recordPer-round convergence scores and preserved dissent show how a conclusion was reached, who disagreed, and why. |
| FINRA AI Guidance Model governance & supervision | Provider-agnostic, per-round model attribution. | Immutable chained recordEvery model call logged with version + params to the chain; models are swappable, attribution is not. |
| MAS AIRM AI risk mgmt & data sovereignty | On-prem sovereign node — nothing leaves. | Region pinning · sovereign nodeAir-gapped option, jurisdiction/region pinning, zero external API dependency. |
| GDPR Art. 22 · PDPO Automated-decision rights & data locality | Human-in-the-loop and on-prem locality. | HITL trace · region pinning · kill-switchExplainable multi-agent chain, immediate halt on demand, and data pinned to the sovereign node. |
| SOC2 Type II Security, availability, confidentiality | Immutable audit trail with role-based access. | Immutable chained recordRBAC, on-prem option, and source-available code for independent review. |
On-prem removes CTP risk. Named-operator HITL trace satisfies SM&CR accountability. Dissent logging demonstrates Consumer-Duty good outcomes.
Art. 14 oversight is the core architecture; Art. 12 logging is a byproduct; Art. 50 provenance per output; kill switch = emergency stop.
Per-round model attribution satisfies FINRA governance. SOC2-ready trail. Source-available enables independent review for federal procurement.
Air-gapped on-prem for full data sovereignty. Zero external API satisfies MAS outsourcing. Configurable retention for PDPA.
Kill switch + HITL override meet the AI Basic Act’s high-risk requirements. On-prem for PIPA data localization. Trail for FSC reporting.
HKMA expects explainability + human oversight in banking — the transcript and HITL trace provide both. On-prem for PDPO localization.
Complete agent conversation across all rounds — every prompt, response and reasoning step, timestamped.
Agreement heatmaps, per-round convergence scores, final consensus metrics — exactly how agents aligned or diverged.
Every human intervention timestamped and attributed — injections, redirections, approvals, rejections — to named operators.
Which agents disagreed, on what, and why — preserved even when consensus is reached. Critical for FCA Consumer Duty.
On halt: a full state snapshot — partial results, active agents, reason for termination, operator identity. Art. 14 & SM&CR ready.
Not a mockup. This is the C2PA provenance manifest Noolog writes for a finished deliberation — carrying verbatim field names and the real claim-ledger tally. The same report renderer also produces IEC 62304 and FDA algorithm-record documents.
// C2PA provenance manifest — generated per deliberation { "claim_generator": "noolog/1.0", "format": "application/vnd.noolog.deliberation+json", "title": "Deliberation — Raise prices 10%?", "generated_at": "2026-08-01T00:00:00Z", "assertions": [ { "label": "c2pa.actions", "data": /* C2PA-standard action set */ }, { "label": "noolog.deliberation", "data": { "winner": "EngineerBot", "rounds": 3, "convergence_pct": 92, "evaluator_agreement_pct": 78 } }, { "label": "noolog.claim_ledger", "data": { "verified": 5, "contested": 1, "wrong": 0, "unverified": 2 } } ], "provenance": { "chain_head": "a3dbbad…" } }
Emitted live to NATS nsed.{session}.audit.* as a signed AuditEnvelope — DualSigner: Ed25519 + ML-DSA-65 (post-quantum) — and anchored to a Poseidon-BN254 commitment chain. The same chain_head rides every event of the deliberation. Tamper-evident by construction.
Today's export is tamper-evident and machine-readable. Next it becomes regulator-ready and independently verifiable. This is on our roadmap — a designed premium-enterprise direction shown here as intent, not a shipped feature.
Requirement → proposal → PM approval → QA evaluation → QA-lead sign-off, each row cryptographically signed. Export as JSON, CSV, or a PDF regulators can ingest.
One deliberation, rendered into IEC 62304, DO-178C and 21 CFR Part 11 submission formats.
The DualSigner envelopes are emitted today; a recipient-side verifier turns "tamper-evident" into "independently verifiable — any post-export edit is detectable."
Verdict, dissents, per-string provenance and explicit liability mapping bundled with a replayable, chained audit trail.
Compliance isn't only the protocol and the box. The interface an operator actually touches is built in open-source Slint — a statically-typed UI toolkit that compiles to native Rust/C++ with no garbage-collected runtime and deterministic execution, shipped with a GUI test-automation framework. That engineering discipline is exactly what lets a safety-critical UI be certified — so the interface is built to map onto the major functional-safety and software-quality frameworks.
Not every standard applies to every deployment — several are embedded- or hardware-specific. The point isn't that the desktop app is certified to all of them; it's that the UI is engineered to that bar, on a toolkit built and tested for exactly these regimes.
A protocol is only as trustworthy as the machine underneath it. So the second pillar moves the guarantee into silicon: the device is a sealed, tamper-responsive enclosure whose on-board integrity anchor makes the audit trail append-only and tamper-evident. The log can’t be quietly rewritten — the hardware physically resists it.
Everything that touches your data — SSD, NPU, TPM, and an optional HSM — lives inside one sealed boundary. Integrity is anchored in one of two ways: an HSM that owns an append-only tamper log, or internal buses hardened so nothing can sit between the components undetected.
An on-device Hardware Security Module owns an append-only, tamper-evident log. We target certifying it to a recognised standard so the tamper-log guarantee is independently verifiable, not just asserted.
Drop the HSM if every internal path — SSD ↔ NPU ↔ TPM — is secured against man-in-the-middle by physical design, using the IDE (Integrity & Data Encryption) and SPDM protocols end to end.
Chasing full standard certification is, in part, a marketing move — a badge buyers and regulators recognise. We say so plainly. But the reason it’s worth doing is that the standard forces the integrity to be real and independently checkable, not a claim in a datasheet.