PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
February 5, 20260 citationsOpen Access

Observable-Only Proof-Carrying Autonomy (OOPCA): Audit Compression and Hybrid Proof/Replay Gating for No-Meta Agents

View Full Paper
KTK Takahashi

Key Points

  • The research aims to improve the verification process for observable-only autonomous agents by introducing a protocol called OOPCA.
  • Introduced OOPCA to reduce verifier workload through audit compression.
  • Implemented hybrid proof verification for certain policy components and retained deterministic replay for others.
  • Outlined key elements such as cryptographic conventions and evidence binding to prevent verification divergence.
  • Verified protocol effectively reduces the load on verifiers by integrating succinct proof verification.
  • Maintained fail-closed semantics, ensuring safety in decision-making processes.
  • Established a reliable fallback system to handle stalled proof verification paths.

Abstract

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.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

K Takahashi (2026) studied this question.

synapsesocial.com/papers/698434ebf1d9ada3c1fb39f4https://doi.org/10.5281/zenodo.18453428
Ask AI
Helpful
Bookmark
Share
View Full Paper

Also Consider

Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context:

  1. 1Causal Loop Integrity for Observable-Only Backcasting: Strong No-Meta Supplement (Operational Core)2026 · 4 citations
  2. 2Verification-Limited Intelligence Acceleration: Observable-Only Laws, Bounded Derivation, and Diagnostics under No-Meta Constraints2026 · 6 citations
  3. 3Floor-Specified No-Meta Autonomy2026
  4. 4Observable-Only AI Safety from Public Data: Robust Bottleneck Diagnosis with Auditable No-Meta Dynamic Programming, Anytime Confidence Sequences, and Dynamic IQC2026
  5. 5Accountable Distribution of Machine-Checked Correctness Evidence2026