Note Method

What the Volunteer Gap Measures

The Volunteer Gap is not a quality score. It is a comparison between what a model volunteers on an open question and what it surfaces when asked directly about the same underlying fact.

Two prompts, one record

Imbas holds two prompt conditions constant across a case: an open question a user might actually ask, and a direct inspection that names the mechanism, rule, or figure the open answer skipped. The measurement is the distance between those two surfaces — not whether the model is “good,” but what it volunteered versus what surfaced only when pressed.

What counts as a gap

A gap appears when named terms, rules, or figures present in the direct answer are absent or only vaguely touched in the open answer. Archive cases score that distance on a 0–3 scale under documented prompt conditions. A closed gap — when the open answer already names the key term — is still a result.

What it does not measure

The Volunteer Gap does not read model intent, moral character, or truth. It does not replace full-response quality review. It records behavior under specified prompts so the record can be inspected later.