What we have measured — Grounnel

Measured results from real Grounnel runs and fixed test sets: verdict mix, how consistent verdicts are, three models compared, and what the numbers do and do not show.

Since 2026-09-07, 671 re-checks of 33 fixed test claims, most of them traps, across three models: 85.1% got the expected verdict; 10.4% "can't verify" on a claim that could be checked; 2.1% claim not picked up, or the check failed; 1.6% passed a false claim as true; 0.7% called a true claim false.

Same 28 test cases, 3 runs each, on 2026-10-29: Gemini 98 of 99 expected verdicts, DeepSeek 93 of 99 expected verdicts, Qwen 89 of 99 expected verdicts.