The Architectural Decision: Why Deterministic Math Must Anchor Generative AI
Relying on stochastic language models to track psychological state invites hallucination and score drift.
When building complex cognitive systems, the most seductive trap is asking large language models to calculate personality vectors or maintain state directly.
LLMs are probabilistic token predictors, not mathematical state machines. Even with constrained JSON schemas, non-deterministic drift corrupts longitudinal scoring.
At QMatus, we instituted ADR-001: The Deterministic Core with Generative Coating. All vector deltas, confidence calculations, contradiction detection, and compatibility formulas run in pure deterministic Python algebra.
Generative language models are used strictly as a coating — rephrasing questions with empathy and tailoring conversational copy. The brain remains 100% verifiable.