Hidden State Geometry Predicts Hallucination Before Generation: A Universal Pre-Generation Detector Across 7 Architectures

This record contains the full replication package for the paper"Hidden State Geometry Predicts Hallucination Before Generation:A Universal Pre-Generation Detector Across 7 Architectures"(Blaschuk, 2026). We demonstrate that the geometry of hidden states encodes a reliablesignal of hallucination before a single output token is generated.Using linear probes on residual stream activations from 7 languagemodel architectures (GPT-2-XL, Llama-3.1, Gemma-2, Mistral-7B, OPT,Phi-3, Falcon), we show that a pre-generation linear classifier predictshallucination with ROC-AUC 0.652-0.757 (mean 0.695) across all 7 models,significantly above chance (0.500). Live validation on Llama 3.1-8B-Instructconfirms AUC = 0.740. Layer sweep across all 32 layers reveals monotonicaccumulation of hallucination signal from layer 0 (AUC = 0.628) tolayer 26 (AUC = 0.757). Contents: Blaschuk_PreGeneration_Hallucination_2026.pdf — full paper replication_script.py — end-to-end replication pipeline (Python) pre_gen_hiddens_llama31.pkl — pre-generation hidden states, Llama 3.1-8B-Instruct, N=283 hallucination_detector_FINAL.pkl — numerical results fig1_pca_ef_separation.png through fig7_confusion_matrix.png — all figures This work is part of the DSAOP research series and builds directly on: Blaschuk (2026), DOI: 10.5281/zenodo.19455867 Alieksieienko (2026), DOI: 10.5281/zenodo.19415238