Proof.
From a question to a justified answer.

A question to a collection is first a restriction: of all objects of the collection — letters, pages, sentences, persons — which can bear on it? famulus-proof builds an evidence space for each collection, a set of evidence objects with their attributes in DuckDB, and treats search as the restriction of those attributes. Every object has an address that is used verbatim in every claim (4,957 lines in the evidence space).

The filter is a funnel with named stages, and the probe lab shows it on single cases: which sentences survive, which are dropped, and why. See filtering of evidence spaces, evidence objects and attributes.

An answer is the conclusion of an argument. Addressed facts of the collection are its premises; critical questions search for objections; a conclusion is validated only when it follows from its premises. An objection against a reason becomes a checked lesson for every argument of its kind (4,514 lines in the argument layer).

Goal enfolding reads the proven evidence space and judges relative to a stated state of the collection; the result is a validity tree. The stages run as a score, the AI Partitur (12 pages), so that every step is named, versioned and repeatable.

The draft Famulus Reasoning (16 pages) describes evidence, reasoning chains and validation for questions to scholarly collections. The whole of famulus-proof is tested by 263 test functions; the map of the workspaces shows where each task is solved.