Meta-reasoning v6 (9B): side-by-side rollouts

Same 10 problems (5 TCSBench-2026, 5 ArxivMath-0626), sample 0, v6 scaffold at depth 8 × width 6 with explicit budgets and 12-slot rolling compaction. Columns: the v6 SFT (midtrain) model, the outcome-only RL run at step 150, and the same run at step 250 (final). The executor (E) is the frozen v6 SFT in all three; the RL policy also writes the compactions. Scores: TCS rubric (GPT-5.6-Luna high), ArxivMath correctness (Luna medium).

Each layer shows: compaction run before the planner turn (if any) → frontier the planner saw → planner thinking → planner output (directions) → each direction with E's public result. E's private reasoning is omitted; long compactor/final-answer reasoning is truncated in the middle.