Research.
Each study is registered before it runs and published whatever it shows.
Posts.
Given the task, three Claude models let wrong runs through
Given the same 420 real OpenFOAM runs and the task, Courant accepted 0 of 264 wrong runs. Claude Opus, Sonnet and Haiku 5.5 accepted 69 of 264, 76 of 263 and 52 of 249.
1,827 real runs, and the two bars we missed
Of six hypotheses pre-registered on 1,827 real OpenFOAM runs, four held and two were refuted.
Exit zero is not evidence
Two real pitzDaily runs both end with End and exit 0. One is a valid result. The other stopped above its own convergence bar.