99XLAB / WORKING PAPER 01 / FORTHCOMING
Recommended
Is Not Qualified.
A recommendation can be authentic, well sourced, and still unfit for the action in front of it. Reliance begins where retrieval and verification stop.
- AUTHOR
- 99xlab
- STATUS
- Forthcoming working paper
- CANONICAL
- 99xlab.com/writing/recommended-is-not-qualified
AI systems are increasingly asked to move from answering questions to influencing or executing consequential actions. Most assurance methods remain centered on the claim: is it authentic, traceable, or supported? Those properties matter, but they do not establish whether a particular actor may rely on that claim for a particular purpose at a particular moment. This paper separates recommendation from qualification and proposes five reliance conditions — authentic, current, applicable, authorized, and evidenced — as the minimum structure for governed use.
Verification is a property of a claim. Reliance is a property of an action.
A signed price can be authentic and expired. A retrieved policy can be current and inapplicable to the customer in front of it. An authenticated agent can still be unauthorized to spend. Provenance can show where an answer came from without showing whether this actor may use it, for this purpose, now.
The consequential unit is therefore not the answer or even the session. It is the delegation: the bounded transfer of authority from a person or organization to a system that may affect the world.
A qualified answer must survive five different questions.
AUTHENTIC Is this the claim and source it purports to be?
CURRENT Is it valid at the time of intended use?
APPLICABLE Does it govern this entity, context, and purpose?
AUTHORIZED May this actor use it to take this action?
EVIDENCED Will the record show what was known, allowed, done, and reconsidered?
The model is a component. Qualification is an operating system.
No model can infer these conditions reliably from language alone. They live across identity, policy, rights, time, dependencies, records, and human accountability. A reliable system must preserve those distinctions rather than compress them into a confidence score.
This is the territory of NINE, the System of Reliance: know what can be used, see what change affects, and prove what happened next.