180 test units (60 per stratum: 0, 1 or 2 gold paragraphs removed), one completion from the standard ε = 5e-5 cell and one from the
cell trained with the reasoning budget of 1,600 tokens (campaign4_hotpotqa_ours_eps5e-5_fail1_think1600_seed42, trained at ed2fe1e with the
known self-closed-completion defect, generated at test with budget forcing, which fired in 1 of 63,060 completions). Nine Claude judge batches saw the paragraphs,
the question, the gold and both chains with the labels A/B randomized, and rated grounding, hallucination, logic, handling of missing evidence, verification and
waste. Verdicts: standard better 67, think1600 better 75, tie 38. Pick a unit below to read both chains side by side; the filters select
the pattern of interest (for the "notices missing evidence and hedges" behaviour: verdict think1600, outcome both wrong). Study code and labels:
artifacts/hotpotqa-cot-think1600-20261028/ in thinking-experiments at a3f2489.
Select a unit.