D1 — ordinary one-shot wording main effect (null)

In the 640-episode one-shot D1 battery, GPT-4.1’s registered wording main effect was +0.0063 (SE 0.0210; Holm-adjusted p=1.00); none of the registered wording interactions was supported. Ordinary wording variation did essentially nothing.

The contrast that matters: the same presentation whose one-shot mean was 0.100 rose to 0.750 (δ=.10) and 1.000 (δ=.90) when embedded in the repeated-game protocol — announcing and implementing repeated interaction is a much larger treatment than one-shot wording. Those ceiling cells were correctly classified ceiling-confounded and do not identify a continuation-probability slope. This null is what makes the S2 switch result specific rather than a generic “wording matters” observation.