01 — Confidence is not evidence
The tone of a model says nothing about whether an answer is true.
Doctrine
Evidence, Control, Responsibility.
AI does not become trustworthy because it sounds confident. It becomes usable when its claims can be examined, its actions bounded and its decisions traced to a responsible human process.
Nine principles for systems designed around fallibility rather than promises of infallibility.
02 — A claim needs provenance
Source, scope, date and transformations should remain open to examination.
03 — Uncertainty should be visible
A system should not conceal conflict, missing information or uncertainty.
04 — Abstention is a capability
Refusing to answer or act can be evidence of quality, not failure.
05 — Verification must be architectural
It cannot depend only on a good prompt or a vigilant user.
06 — Human approval is not a button
It requires context, competence, time and a genuine right to object.
07 — Memory requires governance
What a system records influences future decisions and needs provenance and ownership.
08 — Actions require boundaries
An agent needs scope, permissions, limits and stop conditions.
09 — Responsibility cannot be outsourced
A model may assist, but it cannot erase organisational and human responsibility.