AI oversight infrastructure
Make AI spend and oversight proportionate to risk.
ozDNA turns the risk-based logic of AI regulation into runtime architecture. complyDNA states the obligation; originDNA proves the decision. Low-risk work runs on the cheapest sufficient model; a high-risk decision goes to a multi-model Council — and every step is recorded as evidence.
The problem
AI spend is flat, oversight is manual, evidence is scattered
Three gaps a regulated operator feels at once — and no single tool closes them.
01
Spend ignores risk
The same model cost is paid whether a decision is routine or carries regulatory weight. Nothing routes by risk.
02
Oversight is manual
A human signs off, but the oversight mechanism is not itself software — so it can't scale and it can't be audited.
03
Evidence is scattered
The decision, its rationale, and the record a regulator will ask for live in different places, if they exist at all.
How it works
Classify → Route → Oversee → Prove
One runtime. Low-risk work takes the cheapest sufficient model; a high-risk decision is escalated to the Council, where a higher token cost is a deliberate insurance premium, not waste.
Classify
Each workload is placed on a risk tier, continuously across its lifecycle.
Route
Low risk → the cheapest model that clears the bar. High risk → the Council.
Oversee
Multi-model vote with a veto and a fail-closed judge. Over-reliance on automation is designed out.
Prove
Decision, votes, rationale and outcome are written to an append-only, hash-chained log and rendered as an attestation.
The economic thesis: proportionality for the CFO, evidence for the compliance officer — the same dashboard.
The distinctive part
Each component implements a specific AI Act article
Routing is a commodity; classification is static consulting. Combining them in one runtime — and mapping each part to an obligation — is the position no one else occupies.
| Layer | Role | AI Act article |
|---|---|---|
| ozDNA | Margin-aware, risk-proportionate AI infrastructure | Risk-based approach |
| complyDNA · DT 5.0 | Classification: workload → risk tier; routing decision | Art. 9 |
| complyDNA · Council | Multi-model vote + veto + fail-closed judge on high-risk decisions | Art. 14 |
| originDNA · Ledger | Decision, votes, rationale, outcome — hash-chained, immutable | Art. 12 |
| originDNA · Attestation | Evidence output an auditor or regulator can read | Art. 13 |
ozDNA implements, operationalises and evidences these obligations; it makes no absolute compliance claim. Timeline anchors: Art. 50 transparency duties apply 2 Aug 2026; Annex III high-risk areas 2 Dec 2027; systems embedded in regulated products 2 Aug 2028. Dates must be re-verified before any public material ships.
Live evidence
The classification engine is field-proven
The DT engine is not a slide. It runs in a live, paying product that detects AI-authored text in Turkish academic writing — real users, real revenue, in production.
A live product built on the same classification engine detects AI-authored text in Turkish academic writing — paying users, in the field.
Presented as technical validation of the DT engine, not as a product endorsement. See the Immortal MLRO use case for the same runtime applied to compliance.
Who it's for
Regulated operators under real audit pressure
Built for the compliance-officer-and-CTO pair who decide together — the material speaks both languages: the obligation and the integration.
Next step
Request a compliance audit
Tell us which AI systems you run. The audit output is itself a classification report — the product demonstrates itself.
Submissions reach us by email — no CRM, no tracking beyond the form itself.