We report an experimental test of level-k theory, applied to three simple games with non-neutral frames: Coordination, Discoordination and Hide and Seek. Using the same frame for all three games, we derive hypotheses that apply across the games and are independent of prior assumptions about salience. Those hypotheses are not confirmed by our experimental results. Our findings contrast with previous research which has fitted parameterised level-k models to Hide and Seek data. We show that, as a theory-testing criterion, the existence of a plausible model that replicates the main patterns in these data has a high probability of false positives.