Audit-grade AI is decision-grade AI whose inputs, outputs, and executions can be verified by a third party without trusting the operator.
Why the term exists
AI in finance usually fails one of two tests: the model is opaque to the operator's compliance team, or the deployment is opaque to anyone outside the firm. Audit-grade inverts that. The model card, the benchmark, the deployment, and the trade log are public, addressable, and queryable without an NDA.
What audit-grade looks like in practice
- Reproducible benchmarks. Every model release competes head-to-head on a published cohort at duonlabs.com/arena.
- Calibrated outputs. Probability distributions whose accuracy is testable from external observation via extropy scoring.
- On-chain execution. Vault decisions land as signed transactions on public chains. Anyone can audit them.
- Public research loop. Training, ranking, and deployment visible at live.duonlabs.com.
Compliance becomes a dashboard
The on-chain transparency that scared institutional capital away from crypto is what makes a Duon Labs deployment underwritable. A compliance officer sees every position, entry, exit, and fee on a public ledger. The transparency is the trust mechanism. Multi-quarter diligence collapses into a dashboard review.
What it does not mean
Audit-grade is not open source. Weights stay proprietary. What is open is the evaluation surface: public market data in, public benchmarks and on-chain transactions out, scoring methodology documented. Third parties cannot copy the model. They can verify it works.
Questions
+ Is the Voyons model itself open source?
Weights are proprietary. The benchmark surface, evaluation methodology, on-chain executions, and live research loop are public.
+ Who can verify a Duon Labs deployment?
Anyone with a block explorer. Vault transactions live on public chains. Model rankings live at duonlabs.com/arena. System state lives at live.duonlabs.com. No NDA, no API key.
+ What stops Duon Labs from misreporting benchmarks?
The cohort is fixed before models compete. Scores compute deterministically from public data and public outputs. A third party can re-run the evaluation on the same cohort and verify the result.