JEV's 'No Hallucination' Claim Under Fire: Same Prompts Give Different Probabilities Across Runs
FrankFelixAI · x · 2026-09-21
@neuralavb pushes back on JEV's 'no-hallucination' marketing with two test findings: the system doesn't even run deterministically — identical prompts yield different probabilities across runs — and the ORDER of choices drastically changes output probabilities.
The real guarantee, the author clarifies, is type-safety and deterministic answers (no stochastic sampling). But deterministic isn't correct: answers can still be confidently wrong, and JEV will hallucinate on topics where training data was undercooked.
More from Models
- Hemmingway-1: Apache-2.0 27B writing-only model scores 1330 on EQ-Bench 4, beats GPT-5.5 — Lukinator6446 · 2026-09-21
- Astra is terrified of mistakes — humans may still own high-entropy decisions — xiaosun86 · 2026-09-21
- Inspired by Jev, developer builds Vector, a general action model — BLUECOW009 · 2026-09-21
- Sentdex softens on Jev: solid LLM augmentation saving tokens per iteration — Sentdex · 2026-09-21
- Gemma 4 fell for 3/4 prompt injection traps and ran a shell command; Qwen 3.8 mostly didn't — divinetribe1 · 2026-09-21
- Jev, built by an RLHF/InstructGPT veteran, targets machine decisions with new RLCD training — MaryamMiradi · 2026-09-21