Run D of the terminator debug: full table LoRA with the tables at lr 3e-5 and the linear layers at lr 1e-4, 2 epochs. Failed. Acts in 16.7% of acceptance samples against 80-95% for adapters without table LoRA. Kept so the four failed recipes are all on record.

Qwen2.5-32B base never trained the ChatML control tokens. <|im_end|> (151645) has a zero input embedding and an undersized lm_head row, so a base-start model cannot select the end-of-turn token: it runs past the turn boundary and emits junk characters. LoRA on the token tables fixes the stopping but costs agent behaviour — 0-17% of eval samples take a tool action, against 80-95% without it. These arms are the search for a recipe that keeps both.

See data/misalignment-eval/table-lora-debug/ in the project repo for the per-run numbers and the measurement caveats.