Order Survives the Endpoint: A Locked Test of Experand-Relevant Trajectory Structure in a Transformer

Current evaluations of language models often privilege outputs or endpoints, while transformers can be highly sensitive to prompt wording, demonstration order, premise order, and positional structure. The Experand framework poses a different question: two histories can share an endpoint and a report while remaining different trajectories. This paper does not test consciousness. It asks whether an order-sensitive trajectory distinction survives when endpoint identity, report identity, token length, template form, and event inventory are controlled. A first locked study of Qwen2.5-0.5B-Instruct produced broader positive history-conditioned effects for path separation and recovery geometry, but a token-length diagnostic showed that the principal comparisons were not surface-matched. A second locked study therefore imposed exact tokenizer-length matching, a fixed four-line template, fixed endpoint and report text, symbolic event codes, two counterbalanced codebooks, and a matched order-reversal control. Under the stricter design, generic path separation and recovery geometry did not survive. Differential order sensitivity did: reversing the same Experand-coded event-token multiset produced greater final hidden-state separation than a matched control reversal. The effect was positive in both codebooks. The result provides initial empirical support for one Experand-relevant operational distinction while leaving the remainder of the framework open.