13.7 Pretrained GPT-2 probe

Experiment: registered GPT-2 Medium pilot

Across three seeds, normalized commutator falls \(37.7\% \); pairwise logit and hidden order MSE fall \(58.5\% \) and \(58.6\% \); triple-permutation MSE falls \(57.9\% \). Mean task loss increases \(0.0026\) nats, and every geometric endpoint improves in every seed.

The pilot validates transfer of the bracket-control signature to pretrained attention adapters. Its short training schedule and low absolute SacreBLEU make it a geometric result, not evidence of high-quality semantic generation.