This record provides a machine-auditable specification for observable-only, no-meta autonomy: decision-making and coordination protocols that do not rely on privileged evaluators, hidden scoring, or unverifiable authority. The core design goal is fail-closed safety with replayable evidence. All eligibility, caps, closure, and action authorization are derived from deterministic, bounded, and publicly checkable rules over a canonicalized event log. Key contributions Deterministic canonicalization (encoding + normalization + domain-separated hashing) so independent parties can reproduce identical bytes and hashes. Authority-bearing commitments with non-circular roots (rejection noise and test channels are explicitly excluded from authority roots). Bounded selection and metering: per-item and per-decision work is capped and replayable, reducing gaming incentives and denial-of-service via unbounded evaluation. Deposits, bonds, and escrowed obligations: observable enforcement primitives that align incentives without hidden adjudication. Caps and pessimistic release: explicit loss bounds and conservative release rules to prevent runaway exposure under uncertainty. Finality-gated external effects: real-world actions are authorized only after deterministic replay checks and rule-fixed receipts. Included artifacts PDF specification, schemas (machine-readable identifiers and bounded strings), and pseudocode-level procedures for audit, replay, and enforcement. This work targets implementers of multi-agent systems, distributed governance, transparency logs, and safety-critical autonomous services where “trust me” is not a valid primitive.
K Takahashi (Tue,) studied this question.