Information Authority: A Knowledge-Provenance Architecture for Non-Cheating Tactical AI, with Unity-Free Reproducible Verification

Tactical AI in games and training simulations frequently relies on privileged access to complete world state — informally, "cheating" — which undermines believability and the training validity of simulated forces. This paper presents two contributions. First, an information-authority architecture that assigns every AI-facing datum a knowledge-provenance level and structurally blocks privileged-truth data from driving runtime decisions, so that non-cheating is enforced by fail-closed defaults and a central runtime validator rather than left to per-agent discipline. Second, a Unity-free, reproducible verification that the no-cheat invariant holds for the covered scenarios: a central output validator rejects any decision derived from privileged truth at both the aggregate and per-decision level; an auditable "bait" scenario presents privileged enemy data that the honest decision path declines (zero privileged-truth decisions while recording the blocked accesses), verified across five decision-intent types; and a fail-closed negative control deliberately consumes privileged truth and is caught and classified. The verification uses a deterministic, trace-only headless evidence path with manifest-driven scenario families, repeated runs with confidence intervals, runner-level negative controls, artifact-hash checking, same-machine clean-clone reproduction, and containerized CI clean-environment reproduction that reproduces the reported numbers bit-for-bit. All guarantees are conditional on faithful provenance tagging at the perception layer: the validator enforces that no decision labeled privileged passes, not the correctness of the labels themselves. These results support only the constrained claims that (a) the architecture rejects privileged-truth decisions and (b) this rejection is reproducibly auditable without Unity. They do not support human-realistic behavior, tactical superiority, real-world operational validity, medical/targeting/surveillance/weapon-control validity, or approval of learned live action. Reproduction artifact (code subset, manifests, Dockerfile, CI workflow, results): https://github.com/anomalia0287-ai/information-authority-repro — archived at https://doi.org/10.5281/zenodo.20405019