Trust record / Responsible AI

Model capability does not replace accountable judgement.

AI is used inside a named workflow with explicit inputs, evaluation, review points, unsupported-answer behaviour and human release.
Responsible AI posturereviewed 19 August 2026
Current posture
Provider, model, data-use and human-review requirements are determined by the specific workflow and its consequence.
Applies to
The website or a specifically scoped engagement.
What this cannot conclude
Universal accuracy, autonomous authority or fitness outside the tested workflow boundary.

Operating detail

Responsible AI is recorded through concrete decisions.

The fields below describe the current method. Engagement-specific evidence and terms still have to be produced for the workflow in question.
Control / 01Responsible AI

Purpose and boundary

The model has a named role inside the workflow. It does not silently acquire authority because its output appears plausible.

  • Permitted inputs and tools
  • Actions the model cannot take
  • Named decision owner for release
Control / 02Responsible AI

Evaluation

System behaviour is checked against representative tasks, unsupported states and known failure modes before release.

  • Acceptance criteria before build
  • Refusal and escalation cases included
  • Material changes trigger re-evaluation
Control / 03Responsible AI

Human accountability

AI may draft, retrieve, classify or recommend. Consequential judgement remains assigned to people who can understand and override the output.

  • Visible review state
  • Correction and escalation path
  • No automation theatre disguised as oversight

Evidence request

Ask for the evidence behind the statement.

A procurement or technical question should name the control, workflow and evidence requested. The answer should state the current fact and the uncovered gap.
Responsible AI question recorddirect founder response
Question
The exact control, policy or evidence requested.
Applies to
The named workflow, system and engagement.
Response
Current fact, available evidence and any gap.
What this cannot conclude
That one scoped response covers another workflow, provider, environment or contract.

Send one precise trust question.

Discuss one workflow