Trajectories That Segment Themselves: Agent-Declared Boundaries as a Training Unit

The paper proposes collection-time semantic self-segmentation for long-horizon coding-agent trajectories. During execution, the agent declares falsifiable causal hypotheses whose adoption defines variable-length phases, avoiding fixed windows, milestone vocabularies, replay, teacher logits, or retrospective segmentation. After removing the declarations, models attribute action blocks to their governing hypotheses at more than twice chance, with a paired sign-test p-value of 0.0002; a code-blind annotator matches 24 of 40 boundaries versus 11.5 expected from random placement. DPO on 2,551 phase-boundary pairs changes no decisions on 91 adversarial items, but changes four of 60 matched-construction items from wrong to right.