Executable Epistemic Contracts in Deterministic Agent Simulation: The ENIGMA Architecture

In most modern large language model (LLM) based agent simulations, agent beliefs are implemented as prompts or verbalization structures, and belief divergence arises as a side effect of stochastic text generation rather than as a designed, verifiable property. We present ENIGMA, a deterministic multi-agent architecture in which belief divergence is a first-class, causally efficacious, and executably verifiable property of the simulation. The core of the architecture consists of: ExperienceTrace, a unit of interpreted experience with first-class provenance (perception, testimony, introspection); a perception membrane that strictly separates world truth from agent observations; DeltaGate, the single legal interpretation-to-state entry point that isolates the LLM from the causal core; and BeliefCrystallizationEngine with asymmetric trauma (×6) and entropy decay (τ = 100 ticks). The methodological contribution is executable epistemic contracts (SUPERBOX): scenarios that turn epistemic claims about the system into binary, reproducible verdicts. We report reproducible runs under a fresh-extraction protocol: 45/45 core invariant probes, deterministic belief revision, the measured update asymmetry, the propagation of a single rumor with listener-dependent confidence, and the belief-to-decision causality confirmed by a positive control — four agents witnessing a theft form beliefs with confidence 0.9, disposition-specific decision modifiers, and a changed intent. We also present an honest audit of the scenarios that fail in the studied snapshot. The red verdicts are not concealed; they demonstrate that the contracts function as a diagnostic instrument rather than decoration. The work is positioned strictly as a systems contribution and a runnable reference implementation, not as a new psychological theory.