New Model for Autonomous System Execution Under Partial Observability

A recent study published on arXiv presents the Reconstructive Authority Model (RAM), which tackles the challenge faced by autonomous systems functioning under conditions of partial observability, where the relevant execution state is never completely visible. The researchers contend that current governance methods, including trusted execution environments, oracle-signed state proofs, and cryptographic attestation, ensure computational integrity and state projections but are fundamentally inadequate, as an authenticated state projection alone does not guarantee execution validity. RAM distinguishes integrity from coverage by introducing a reconstruction gate that assesses an explicit coverage envelope, which includes proven state, declared assumptions, and recognized unobservable residuals. Execution is allowed only when coverage meets the requirements for the action class, while insufficient coverage results in dynamic privilege reduction or a fail-safe mechanism. The paper differentiates attestation, which validates trust in measurements, from RAM, which verifies the adequacy of observable elements.