How a system answers only what it can show

Deterministic computation, exposed step by step. The model does not compute.

• The computing core is deterministic: the same input returns the same output, always.
• Every step of the computation is exposed, so a person can repeat it by hand or in another implementation.
• The language model never performs the computation itself.

Facts in an addressable store with origin and licence. Empty is an answer.

• Facts live in an addressable store, separate from the model's weights.
• Every fact carries its origin: source, author or issuer, date, licence.
• An empty result is a valid answer and must be returned as such, never filled in by generation.

Every finding comes with the chance of it happening at random, over the real search space.

• Every claimed finding is reported together with how often it would occur by chance.
• The size of the search space is counted and disclosed, and corrected for multiple comparisons.
• A negative result is published on the same terms as a positive one.

Refusal rules in code, machine-readable, about people, the future and missing sources.

• Refusal rules live in the code, not in a prompt, and are machine-readable to the caller.
• The rules name what the system will not say about a person, about the future, and about a source it does not have.
• A caller can read the rules before it presents anything to a human.

A hash and a reproducible call under every answer.

• Every answer carries a record: a hash of its content and the exact call that reproduces it.
• The record changes only when the input, the data or the engine version changes.
• The engine version is part of the record.