TraceForge Synthesizes Family-Native Reasoning Traces, Beats Fable 5's Own Traces by 3.7–7.1 Points on Code

TraceForge trains a generator and critic on Qwen3.8-Max traces alone, then writes 'family-native' reasoning for a closed teacher's answers without ever reading the teacher's hidden thinking. On Qwen3.5-9B-Base distilled from 5,000 Claude Code trajectories, swapping Fable 5's real traces for family-native ones with the same answers lifts HumanEval+ by 3.7 points and MBPP+ by 7.1 at the final checkpoint, and a judge rates the synthetic traces 0.68 real vs 0.66 for the originals — a legal workaround to frontier labs stripping reasoning from API responses.