Research · 2026
A library of physical limits barely moves a frontier model.
On 101 real claims never used for development, a frontier model found 25 of 29 impossible claims alone and 27 of 29 with our library of limits. The difference is not significant. An earlier result of 42 to 47 of 47 came from a setup that overstated it, and we withdrew it.
Frontier model25 of 29
+ ONTO library27 of 29
One square = one impossible claim, 29 in all. Filled: caught and the broken limit named.
25 to 27impossible claims found, of 29 (exact McNemar p = 0.625)
0 of 44false alarms on possible claims, both arms
14 to 20said not enough data, of 28; not significant
How we measured
Claims101 real claims never used for development: 29 impossible, 44 possible, 28 lack the figure that decides them.
Two armsThe same model in a fresh chat, cold, then with the library pasted in. Preregistered before the run.
ScoringAgainst the key, with no AI judge. One pass in a chat, not an API run with fixed versions.
What it gives
The library does not make a strong model better at physics. What the model lacks is saying not enough data when that is the true answer.