Native unknown Outputs as a Scheduling Signal for Model Arrays: Small-Model Gatekeeping, Queryable Knowledge Maps, and the Failure Boundary of Same-Family Weighted Voting (English v1.1)
Chao Qin · Zenodo (CERN European Organization for Nuclear Research) · 2026
Companion short paper to "A 774M-Parameter Model Trained from Scratch Says 'I Don't Know'" (10.5281/zenodo.22191510). Moving the uncertainty signal from post-hoc estimation to generation time: unknown as a first-class output category becomes a metadata bit that a model array amplifies into three system capabilities — gatekeeping (37× leakage suppression; out-of-domain perception holds on all 55 concepts of a three-gradient probe set, with root shortcuts inert outside the boundary band), a queryable knowledge map (cross-seed consistency 0.946, disagreement self-localizing to the known/unknown critical band), and union-intersection two-expert routing (3,000-step unit expansion, zero forgetting). A clean negative result bounds the signal's value: within a same-family array, equal/quality/oracle weighting and two further fusion operations all degenerate to identical decisions (0.976) — while a single rule unit carrying out-of-array information lifts the same probe set to 41/41: overriding power requires out-of-array information sources (measured, M4v3). A six-route industry survey finds no counterpart for the native-metadata-bit + small-model-gatekeeper + state-judgment-array combination. v1.1 (2026-09-01): back-fills M4v3 heterogeneous overriding (41/41) and M2c sharp-boundary probes (11/12; first lexical-classification failure instance). Preregistered; negative details reported as-is. This record is the v1.1 of 10.5281/zenodo.22220040, published as a new deposition due to a Zenodo new-version pipeline fault.