Responsible AI

A platform that assesses AI should be careful about its own use of it

DUN-AI reaches its conclusions deterministically. A language model may help phrase a narrative section; it is never permitted to decide anything, and the boundary is enforced by the architecture rather than by a policy someone has to remember.

The separation is structural

Policies drift. Architecture does not. The decision engines have no route to the AI gateway: they do not import it, cannot reach it, and a test in the suite reads the module imports of every engine to prove it. If someone adds that import, the build fails.

Never model-determined
  • The final verdict
  • Any score or its band
  • Any financial calculation
  • Regulatory classification
  • Whether a hard blocker applies
  • Whether a risk is acceptable
  • The ranking of alternatives
  • The confidence calculation
May be model-assisted
  • Phrasing of narrative sections in a report
  • Summarising evidence a user supplied
  • Suggesting a clearer problem statement
  • Rewording a question for a specific audience

Any assisted text is marked as such in the report, and the underlying determination remains the deterministic one.

Stated in every report

Each report carries an explicit field recording whether a language model influenced the verdict. It reads no, because the code permits no other value. That field exists so a reader does not have to take this page on trust.

Assistance is optional and off by default

The platform runs completely without any model. AI assistance is disabled unless an operator supplies a provider credential, and every deployment can state which mode it is in. There is no hidden dependency on an external service.

When assistance is enabled, prompts and responses are handled under the same data classification rules as the rest of the assessment. Confidential answers are not sent to an external provider.

How we apply our own framework to ourselves

The principles the assessment framework applies, and how the platform meets them in its own construction.
PrincipleHow it applies here
Meaningful human controlA verdict is a recommendation until a named person seals it. Overriding it towards a more permissive outcome requires a written reason, a specific permission and a second approver.
ExplainabilityEvery score carries a rationale, every gate carries its reasoning, and the decision path from input to verdict is recorded stage by stage.
ContestabilityThe inputs, the content versions and the logic are all recoverable, so a conclusion can be examined rather than merely disputed.
ReproducibilityContent versions are pinned to the assessment. Re-running produces the same result, and reports carry an integrity hash.
Honesty about uncertaintyConfidence is calculated, ranges are used instead of point estimates, and a thin evidence base produces a request for evidence rather than a confident answer.
Independence from commercial interestAssessment engines cannot read commercial or vendor data. A test verifies this by inspecting imports, for the same reason the AI boundary is tested.

Security and data handling → · Known limitations →