Resolve endpoint asymmetry in entropy trajectories

Determine whether omitting the final answer-bearing unit from the entropy trajectory for completed chains, but not for truncated chains, materially changes the coherence and magnitude results.

Background

When a continuation from the final prefix returns no parseable answer, the study uses the preceding prefix as the final trajectory point. In the traced sample this occurs predominantly for completed chains, whose final answer unit is consequently omitted, whereas truncated chains are treated differently. Since coherence is computed from the first and final entropy values, this creates an endpoint asymmetry in the magnitude statistic.

The paper states that resolving the issue would require recomputing trajectory statistics from retained continuation text, but the main panel did not retain that text. Thus, the potential effect of the asymmetry on the magnitude findings remains unresolved.

References

Whether it moves the result of Section~\ref{sec:magnitude} is untested, and it is not the kind of asymmetry a rank correlation ignores, since the omitted unit is the one at which entropy would be lowest and it is omitted from one group and not the other. Settling it means recomputing the trajectory statistics from retained continuation text, which the panel did not keep.

— Chain-of-Thought Entropy as a Reliability Signal: A Preregistered Reproduction  (2609.19606 - Cochran, 17 Sep 2026) in Section 6.2, Step Segmentation and the Step Cap; Section 9, Conclusions