TTP

Side-by-side chains of thought of the standard ε = 5 × 10⁻⁵ cell and the reasoning-budget cell on 180 test units, with blind judge verdicts: where the budget cell notices missing evidence and hedges, and where it loops.

Which strings count as the gold answer: exact match, token F1 and Zipf-weighted F1 measured on the 5,255 test questions, with the prior distribution under each rule and the wrongly placed mass estimated by an LM judge.

HotPotQA run of our framework at ε = 5 × 10⁻⁵, in the Campaign 4 QA layout: the 99 development units (33 questions × 0/1/2 gold paragraphs removed) across checkpoints, unit by unit, and the 417-step training stream with weight templates, trends and every group's rollouts.

ProtoQA parameter search of our framework: pick a cell (prior scale × ε) and follow it on the development set across checkpoints, question by question, and through its training stream beside the Campaign-4 Multi-RLCR rerun.

ProtoQA training dynamics of Campaign 3: the development set across checkpoints and the training stream beside Multi-RLCR. HotPotQA to follow.

Results tables, the ProtoQA frozen test set problem by problem with inspectable rollouts, and the ProtoQA training dynamics.

Prompt comparisons for ProtoQA, HotPotQA and TriviaQA with Qwen3-8B, including the historical generation prompts.