An Experimental-Verification Program for the Corpus Mechanisms of the Register of Anticipated AI Failure Modes

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Companion document to the Register of Anticipated AI Failure Modes. Purpose: to turn each of the register's 12 domains from a general statement that "experimental verification is needed" into a concrete, honestly bounded plan — stating what is testable now, by what means, at what stage, and what is categorically not testable by a small group at present. Program design: (1) each proposed experiment carries an explicit evidence-class label, separating results that can move a domain’s Mapping status from results that only confirm a mechanism’s internal logic in isolation; (2) Phase 0 is limited to the one item in this program with genuine negative-control evidentiary weight, with lower-weight items moved to a parallel, explicitly non-gating track; (3) Domain 8’s custody/isolation component is treated as infrastructure supporting a combination test, not as a standalone deliverable; (4) a normative rule caps how much a single experiment may move a Mapping status and ties the cap to the evidence class actually produced; (5) Domain 3 carries an explicit cannot-close note. See §5 and §6 for the full rationale.