Credence — Epistemic Guard

Epistemic guard that hooks into transformer hidden states to classify whether the model is recalling training data, using context, or confabulating. Exposes a probe layer trained on Fisher geometry (AUROC 0.9944 validated on Qwen2.5) that fires before token commitment. Blocks file writes and code commits when confidence drops below threshold. The symbolic layer tracks uncertain values across inference steps, so if a variable assignment gets flagged as confabulation risk, downstream operations using it stay blocked until verification. Reach for this when you're letting an agent write code or modify files and need circuit-breaker semantics tied to the model's actual epistemic state rather than parsing output text for hedging language.