Maxwell's Demon as a Test Case for Finite Observer Records

A theoretical and computational analysis of when a physical correlation is usable as a record for a finite feedback controller. The paper separates boundary access, registration, joint context-record reliability, standalone and contextual information, decision value, and physical cost. It preserves exact stuck-memory, redundancy, XOR synergy, key-loss and cost-dominance examples. Version 10 corrects registration-time novelty as a universal gate, distinguishes target-conditioned endpoint reliability from uninterrupted path survival, places policy-dependent costs inside net-value optimization, and clarifies causal baselines, data processing, reference divergence and thermodynamic accounting. The finite target-only zero-conditional-information implication is retained with explicit assumptions. A reproducibility companion includes original v9 files, editable v10 source and equation definitions, and new executable checks: 10,201 binary registration-retention cases, 1,000 conditional-independence constructions, 1,000 joint-channel data-processing controls, 1,000 finite decision comparisons, and 12 named controls. All calculations are analytic or synthetic. The powered-latch and backaction campaign is proposed, not performed; no new Second-Law resolution, universal work bound, laboratory anchor or derivation of persistent uncertainty is claimed. Version 10 supersedes the supplied v9 working manuscript.