NoMMD: A Point-of-View-Bound Architecture for Recursive Mental Modeling

Abstract

Recursive mental modeling becomes difficult to evaluate when an observer's global record is mistaken for a participant's state, later evidence enters an earlier prediction, or fluent summaries silently replace incompatible hypotheses. We present NoMMD, a point-of-view-bound architecture that represents attributed social state as separate append-only ledgers. Its only semantic record kinds are actions, beliefs, desires, and emotions (ABDE); identity, time, provenance, evidence, uncertainty, and retrieval history remain metadata. Each ledger admits evidence available to its owner at the applicable frontier, preserves contradictory records, and derives a bounded working state through activation-weighted retrieval. Recursive depth is represented by nested perspective chains and measured by transitions between modeled points of view. The architecture supports narrative reconstruction, human–AI interaction, and acting agents while making no inference from a stored attribution to intrinsic mentality, qualia, or consciousness. We derive the representation's invariants, temporal and recursive semantics, failure tests, and an empirical program based on forward prediction, counterfactual consistency, calibration, revision latency, and information-leakage ablations. A worked deceptive-interaction example shows why a union of individually valid ledgers can be invalid for every participant.