Conditionally Approved: Proof-Bound Branchable Universes for Long-Horizon AI Agents Under Changing Evidence, Authority, and Time

Preliminary-results preprint (17 pages) and a snapshot of the public repository at commit 12250a9. Conditionally Approved is a fully synthetic, proof-bound benchmark for long-horizon AI agents under changing evidence, authority, and time: five task editions (document truth under pushback, epistemic residue, competing legitimate authority, temporal portfolio control, honest handoff) bound to frozen task, model, runtime, evaluator, world-root, projection, and restore-lineage identities. The paper reports a completed cost-capped live execution across four identity-pinned frontier providers: 55 physical attempts, 37 gradable, 18 infrastructure exclusions, zero perfect-rubric completions, $705.88 recorded spend, with runtime-cohort drift detected and pooled comparisons refused by the verifier. A redacted 55-row evidence index is committed so every published aggregate is independently recomputable (gradia-universe cost-capped-verify); raw trajectories remain private pending rights and blinded human adjudication, and that reconstruction boundary is disclosed. Machine-scored criterion surfaces (1,375 assignments; 486 green / 889 red) are reported as unresolved observations, not model-attributed failures. Licensing: manuscript text CC BY 4.0; code and synthetic fixtures in the archive Apache-2.0. Repository: https://github.com/rudycelekli/gradia-universes-work-sample