AI Control Under Partial Observability: A Minimal Impossibility Result and a Verifiable Governance Proposal (AI 통제 위험의 구조적 한계와 검증가능 거버넌스 제안)

This note argues that under partial observability, worst-case AI control can be impossible (“fiber conflict”), and proposes a practical governance response: high-stakes Action Boundaries with evidence-gated ALLOW/HOLD/DENY and HOLD = fail-closed (no execution). The decision kernel may remain private, but outputs are proof-carrying and third-party verifiable at v1 level: Format (public Receipt/Profile schema + example encodings), Integrity (tamper-evidence via SHA256SUMS), and Provenance linkage (run metadata via RUN_INFO.json). Claims: artifact-level verifiability + enforceable HOLD gating.Non-claims: correctness of decisions, internal alignment, or full prevention of evaluation gaming. Public verification links: Receipt/Profile spec (raw): https://raw.githubusercontent.com/edict-lab/edict-intake/main/docs/spec/edrp-receipt-v0.1.md Examples: https://github.com/edict-lab/edict-intake/tree/main/docs/spec/examples Live outputs: https://github.com/edict-lab/edict-sandbox/tree/main/docs/latest SHA256SUMS: https://github.com/edict-lab/edict-sandbox/blob/main/docs/latest/SHA256SUMS RUN_INFO: https://github.com/edict-lab/edict-sandbox/blob/main/docs/latest/RUN_INFO.json Concept DOI (all versions): https://doi.org/10.5281/zenodo.18604700Version DOI (v0.1): https://doi.org/10.5281/zenodo.18604701