Every pipeline that thinks before it looks has this seam: the model writes a reasoning trace, fresher evidence arrives, and then something has to answer. Keep going inside the model’s own thinking block and it repeats the now-stale conclusion in 65.6% to 87.0% of conflicts. Ask again in a separate call with the new evidence last and it follows that evidence 29.9 to 60.6 points more often — for the same accuracy when the two agree. A trace being useful and a trace being obeyed are different properties, and your eval almost certainly measures only the first.