PRISM: An Observation Layer for Agentic and Long-Running Computational Systems

PRISM introduces an Observation Layer positioned before the expensive controller in agentic and long-running computational systems. Semantic projection and temporal event selection transform process and log history into decision-facing events. The manuscript separates authoritative system success, controller decision fidelity, and optional controller self-report, and reports conditional efficiency effects within the tested scope. It distinguishes the primary M4 Max experiment and clean-v6 long-inference replay from secondary panels and cross-platform replay/live replication. This record contains the authoritative manuscript PDF and citation/documentation materials. It does not distribute source code or experimental provenance artifacts.