Observable-Only No-Meta Causal Autonomy Protocol (ONCAP)
K Takahashi · Zenodo (CERN European Organization for Nuclear Research) · 2026
Observable-Only No-Meta Causal Autonomy Protocol (ONCAP) is an implementation-ready protocol stack for autonomous agents that must operate without any privileged evaluator (“no-meta”) and can rely only on observable evidence (“observable-only”). It treats the future as a verification horizon to operationally tame the arrow of time: agents estimate expected futures and causal pathways, then deterministically exclude trajectories that are predictably not conducive to wellbeing—without introducing a reward channel over wellbeing. ONCAP integrates four mutually reinforcing components into a single turn-based control loop: (1) a Self-Recognition Kernel (SRK) enforcing four safety floors—Visibility, Contraction, Contact, and Dissipation—using SAT/VIOL/UNCERT statuses with fail-closed privilege handling; (2) deterministic, audit-ready replay and observable-only backcasting under a pinned Deterministic Replay Profile (DRP); (3) closure-gated admissibility via commit–open time anchors, deterministic action hashing, and explicit closure obligations with deadlines; and (4) friction-minimal intersubjective contracts that degrade safely under missing evidence and non-adopting counterparties. Anti-gaming is enforced structurally rather than motivationally: decision outcomes become a deterministic function of committed artifacts and pinned policies; non-authoritative proposal computation is separated from authoritative gates; schemas/registries are pinned; chaff is bounded and deduplicated; and uncertainty triggers monotone cap-tightening rather than privilege escalation. Wellbeing is implemented as an exclude-only mask driven by policy-pinned predicate descriptors and conservative abstention semantics—never as a scalar reward signal. The result is a self-contained, verifiable, and audit-friendly autonomy protocol intended for adversarial multi-agent environments where unverifiable internal states, narrative justifications, or subjective self-reports cannot be granted authority.