Research · 2026
A verdict form lifts an untrained model, and the library lifts it further.
No training. Given the library as a tool and a verdict it must fill in, Qwen2.5 7B catches 22 impossible claims instead of 9. Mistral 7B, which almost never called the library, still rose from 12 to 16 on the form alone.
Qwen2.5 7B, words only9 of 47
library + verdict form22 of 47
One square = one impossible claim, 47 in all. Filled: caught and the broken limit named.
9 → 22Qwen2.5 7B; it called the library on 83 of 100 claims
12 → 16Mistral 7B; it called the library on 1 of 100
1 → 4Qwen false alarms, of 36: the cost
How we measured
ClaimsThe same closed 100 real-world claims.
ChangeOnly the environment: the library declared as a tool, and the answer given as a form with three choices (possible, impossible, need data). Same weights.
SetupBoth models 4-bit on one laptop GPU, greedy. In full precision on a cloud GPU the untrained Qwen scored 5 of 47 with words only, so read the laptop numbers within about ±5.
ScoringThe same two blind judges, a third on disagreements.
What it gives
Structure is a free lever, but it trades catches for false alarms. Training has to keep the gain and drop the cost.