An answer needs a record.

01

Evidence and lineage

A defensible result needs its inputs, assumption set, model version and transformation history. Reports should distinguish calculated results, external evidence and interpretation.

Reviewers need a route back to the originating run rather than a number copied without context.

02

Interpretation is not authority

Language models interpret intent and propose approaches. Permissions, budgets and validation belong in enforceable controls outside the model’s discretion.

A plausible explanation does not grant access to a dataset or authorise a write.

03

Approval before consequence

Consequential actions should present scope, expected changes and relevant costs before execution. Approval applies to that scope, not to unrelated actions.

A changed plan may require a new decision. A user-visible state should distinguish a proposal, an approved action and a completed result.

04

Roles and ownership

Define who can view, prepare, approve and execute work. Assistant access and product access may differ. Review grants when responsibilities change or a person leaves.

Your organisation remains responsible for deciding who is authorised and for checking that the deployed controls match its requirements.

05

Limits and review

AI output can be incomplete or wrong. Check assumptions, units, provenance and suitability before using results in engineering, investment or policy decisions.

Ask for deployment-specific evidence of access testing, isolation, backups and incident procedures. This website does not certify those controls.

06

Related information

Security describes the assurance questions to ask. Privacy covers this public website. Application use, processing terms and service commitments require separate agreements.