I built a template auditor after noticing that comparable models received different reasoning defaults from their chat templates. Then I tested whether that explained the MiniCPM5 and Qwen AIME comparison. It didn’t. Turning Qwen’s thinking on didn’t remove the gap, and trying its recommended thinking sampler didn’t explain it either.The September run contains 259 attempts across four configurations on one RTX 4060. The long-form account, I thought the benchmark was unfair. Then I ran it., includes the method, result table, and the mistakes I made while running it.The code and saved results are public. The original report is FINDINGS.md; LAB_NOTES.md records setup failures, interruptions, and changes made during the experiment.The main limitation is scope. This was one machine and a small set of maths problems, with mid-run changes in the first stage and only answer tails retained. It reproduced a relative advantage in this harness, not the model card’s absolute scores.
