Every result, including the nulls

Each experiment's decision rule is written before its code, and each is run against a matched control. We publish what came out, whichever way it came out.

25 October 2026 · Experiential learning · Positive result Only correct signal teaches Reward from a code-running checker taught a 1.5B model depth-4 composition it lacked. Coin-flip and failure-only reward, run identically, broke the model instead.