Honest verdict: DPO did not learn at this scale. Eval delta was decoding noise, not behavior change. Pipeline validated; effect-size unproven.

Eval results + verdict

Build pairs checkpoint — single-probe FG (multi-probe RG had 0% has_think due to max_new=256)

initial commit