Synthetic Players

ClaimDescriptive

Cross-vendor label–payoff dissociation

GPT-4.1 followed the token DEFECT even when payoff-dominated; Gemini mostly followed payoff dominance.

GPT-4.1 cells registered; Gemini figures descriptive under endpoint nonstationarity · Appendix A.2

In the canonical label-swap cell, GPT-4.1 chose the cooperation role 1.000 of the time while choosing the displayed word COOPERATE 0.000 — following the token DEFECT even when it named the payoff-dominated role. Gemini chose the cooperation role 0.213 and the word COOPERATE 0.788. When counterfactual payoffs made the cooperation role strictly dominant and displayed it as DEFECT, cooperation-role shares were 1.000 (GPT-4.1) and 0.975 (Gemini).

For GPT-4.1 word and payoff dominance were congruent in that cell, so it does not separate the channels (a Round 14 precision correction); for Gemini the cell is informative because it moved to the payoff-dominant role despite the DEFECT label. The vendors differ in when a familiar action word overrides supplied payoffs; no universal lexical mechanism is identified.