Verified Synthesis: From Agent Fan-Out to One Accountable Result

Organizations that dispatch one bounded request to several AI capacity providers receive back a set of outputs that partially agree, materially conflict, omit different evidence, and sometimes fail outright. This paper examines why possession of that fan-out set is not the same thing as possession of an accountable result, and it proposes — as unproven, author-owned design — a minimum verified-synthesis contract intended to close that gap. The proposed contract requires correlation-complete inputs; triage of generated content into load-bearing and supporting claims carrying provenance, evidence class, and locator; contradiction-first, bounded-depth verification against a coverage rule with a frozen denominator; explicit treatment of material contradiction, disagreement, uncertainty, missingness, and minority evidence as five distinct states; forced abstention under named conditions; bounded escalation that decays to abstention; consumption of a pre-committed protected reserve before any separately authorized release; and exactly two terminal outputs — an accountable result or a principled abstention — each preserving lineage, unresolved states, and limitations. The contract inherits authority from ARBITER, budget from PRETIUM, and execution returns from NIMBUS, and it hands its terminal state to the suite capstone without defining clearing. A fair counter-thesis is stated and preserved: published aggregation, judging, calibration, and logging practice may already be adequate at lower cost. The thesis is presented with explicit falsification and narrowing conditions, and nothing in this paper asserts implementation, adoption, superiority, or measured performance.