Sector
Defence
- AI assurance
- Simulation
- Edge intelligence
Systems that must work when the operating picture is degraded on purpose.
Defence AI is evaluated against an adversary who is selecting the inputs. That changes what a robustness claim has to mean: not average performance under perturbation, but demonstrated behaviour at the boundary of a declared operational domain, against a stated threat model, with the search procedure recorded well enough to argue about.
It also changes where the software can run. Accreditation environments do not permit outbound connections, so assurance tooling that assumes a licence server, a telemetry endpoint or a content delivery network is not deployable at all — a constraint that is usually discovered late.
What this means for evidence
Evidence that survives an accreditor
Every claim resolving to a signed, replayable record, verifiable offline by a reviewer with no account and no network.
Degraded-condition coverage
Coverage assessed per ODD cell with operational exposure weights traced to CONOPS, so that the sparse cells are visible rather than averaged away.
Separation of duties
The party that produced a piece of evidence cannot approve the gate that cites it. Enforced, and tested with a negative case.