Same model (grok-4-fast-reasoning), identical adversarial prompts.
Direct baseline: 6/30 hallucinations
nTGT system: 0/30
Difference is system architecture.
Paper: ntgt_epistemic_integrity.pdf
Full response logs are retained but not publicly released.