No-meta, observable-only autonomous agents cannot rely on privileged external evaluators and must ground safety-critical decisions in evidence that is publicly inspectable and deterministically checkable. As such systems scale, shared verification becomes a dominant bottleneck: verifiers must ingest, replay, and audit growing logs while facing adversarial participation, divergence risks, and denial-of-service pressure. This work introduces Observable-Only Proof-Carrying Autonomy (OOPCA), a protocol-level audit mode that reduces verifier workload (“audit compression”) by replacing selected parts of deterministic replay with succinct proof verification, while preserving fail-closed semantics and non-inflationary progress accounting. OOPCA is hybrid by default: proofs are used only for policy components whose semantics are pinned to a Quantized Deterministic Semantics (QDS) profile and compiled into a pinned relation artifact; deterministic replay remains the canonical fallback for components that are not safely compiled (e.g., floating-point dynamics, large unstructured search, debugging/diagnostics). Key elements include: Pinned cryptographic and encoding conventions (hash algorithm, domain separation strings, canonical JSON via RFC 8785 JCS, canonical base64) to prevent verifier divergence. Action-intent and evidence-manifest binding, with explicit evidence-set closure rules so proofs cannot be reused for different actions or hidden inputs. Positive cutoff certificates based on deterministic work-counter state updates, explicitly avoiding succinct “absence claims” as a safety foundation. Bounded liveness through positive fallback-unlock receipts and an observable fallback ledger, keeping the default posture fail-closed while ensuring a deterministic “escape hatch” under stalled proof paths. Accountability without meta via proof-carrying minimal diagnostics and Merkle-based controlled trace disclosure (openable trace fragments). The paper includes a self-contained, machine-consumable JSON Schema bundle aligned with the ONCAP vocabulary, enabling interoperable implementations and deterministic verification across independent verifiers.
K Takahashi (Mon,) studied this question.