HELIX Chat runs Llama 3.1 8B Instruct with a certified upper bound on hallucination probability — not a classifier score, a proved ceiling derived from model geometry. Use this to prototype agent prompts before deploying against the production /v1/inference endpoint via x402. See benchmark numbers.