Trust, graded

Checking every claim a model produces would cost more of my time than the model saves. The system therefore treats verification as spent effort. It goes where a claim must hold, and every unchecked claim is carried, visibly, as unchecked.

The three grades

Every factual claim in the knowledge base holds one of three grades.

Work can proceed at any grade. A claim in a deliverable is stated at the grade it holds, never above it.

The two checks

Verification concentrates at two points. On the way in, a claim worth keeping is tested for whether it is true and whether it adds anything over what the model already produces on demand. On the way out, each factual claim in a deliverable is traced back to its source before delivery. Between the two points, claims can rest unverified without doing damage, because nothing leaves without the exit check.

A claim followed through

A note on an EU regulation stated a set of obligations under Title V of the text. The claim read correctly and had been copied into several other notes. The original document, held locally, places those obligations under Chapter V: the regulation's titles and chapters divide differently, and the wrong label read plausibly through every re-reading. The comparison against the original caught it, the correction was propagated to every copy, and the note's grade rose only after the fix.

The record that makes a claim checkable has four fields, usable in any tool. The claim itself, one sentence. A source pointer precise enough to open: document and article, section or page. The grade, one of the three above. And the check event: who verified it, how, and when. The record travels with the claim, so no claim moves without its state.

What stays human

Two acts stay human. Checking a fact against its source with my own eyes, and endorsing the reasoning built on it as my own. The machine verifies copies and enforces the checks. It does not certify anything.