Action-Bound AI Safety: A Pre-Commitment Runtime Framework for Physical, Cyber-Physical, and Transactional Systems

This Version 3 Full Edition supersedes the earlier v4.8.1 Action-Bound AI Safety manuscript while preserving the original Action-Bound base manuscript as Part I. The release integrates Signal-Time-Authority clarifications and companion synthetic benchmark evidence. It clarifies detection versus redirectability, proxy-faithfulness decomposition, Safety Slack as a diagnostic margin, commitment-boundary calibration, human-latency constraints, falsification discipline, and companion synthetic benchmark evidence. This is a theory-first engineering framework and diagnostic research scaffold. It is not a deployed-system validation, not a safety guarantee, not a medical, industrial, financial, aerospace, transactional, or tool-use certification, and not a proof that all AI systems expose usable pre-commitment signals.