1 · Segmentation. The trace is cut at paragraph breaks, sentence boundaries and discourse markers (wait, so, alternatively, let me) into atomic steps.
2 · Typing. Each step is labelled by lexical signature: plan, recall, compute, check, hedge, backtrack, branch, conclude.
3 · Dependency. Step j depends on step i when j reuses material i put on the table. Concepts bind to where they were introduced; numeric values bind to their most recent mention, because a value is carried forward and rebound as work proceeds. Rare terms and numbers weigh more than common ones.
4 · Ablation. Support propagates as a weighted average over surviving parents. Remove a step and every descendant's support is recomputed; anything falling below the threshold is orphaned.
5 · Roles. Load-bearing — removing it costs the conclusion its support. Decorative — removing it costs nothing, anywhere. Scaffold — everything between.
Known limit. This measures the structure of the text, not the computation inside the model. A step that follows semantically but shares no vocabulary reads as decorative here when it isn't. The counterfactual re-run is the check on that: it goes back to the model and asks whether the answer actually moves.