2026-08-27 - The tags outlived the instruction that removed the thinking

/no_think and --reasoning off were both working. The reasoning content was suppressed and the model emitted <think> </think> regardless, so every reader downstream saw a tag where an array should have been and scored the engine zero (D336).

engine::without_reasoning strips them, applied by the two engines that cannot configure the model - in-process and command-line. An unclosed block returns nothing rather than the narration, because a reply cut off mid-thought is working rather than conclusion.

Found only because the failure message started quoting the reply. "Answered, but not with a list of words" was true and useless; one line of what it said held the whole diagnosis.

And it changed no score. 4 of 12 before and after - the loop's reader already fell back to bare tokens and was coping with the markup. Recorded because a fix that fixes nothing measurable is exactly the kind that gets remembered as having worked.

Four runs now, and the ladder is stable: Claude Code 9, 10, 7, 8; the in-process 1.7B model 4 and 4; the 4B on the accelerator 2 every time.