Does the compiler decide if the robot falls?
A pre-registered test of whether the build of the physics engine changes a policy's outcomes.
1 in 1,000episodes where a different build of the engine changed whether the robot fell
MuJoCo built with GCC, with fused multiply-add, and as a portable build diverges bitwise in 199 to 200 of 200 episodes. If floating point decided outcomes, a policy's success rate would depend on the compiler.
It does not. Across two trained policies, whether the robot fell changed in 1 of 1,000 episodes. Under the broader pre-registered definition, which also counts a fall time or final position that moved, a different build flips 0 to 2 percent of episodes, inside the range of nudging the initial state by one unit in the last place. Every McNemar test returns p = 1, and a fall-rate bias above about 1 percent is excluded.
The analysis script was hashed and time-stamped before any result was read. A clean null, written down before the data came in.