Output Interface, Clause Order, and Inference Budget in State-Conditioned Action Selection across Tuteo and Rioplatense Voseo

Executable language-model agents are often evaluated through interfaces that require an action rather than an explanation. We study how this output interface affects state-conditioned action selection and how inference budgets interact with a minimal within-Spanish contrast between tuteo and Rioplatense voseo. An initial 576-call experiment reveals widespread state-invariant behaviour despite valid command syntax. In a controlled binary-policy design, the same two mutually exclusive clauses are presented in opposite orders. Under direct command generation, pooled accuracy is 52.7%, but correctness is 90.6% when the true clause appears first and 14.8% when it appears second. A format-only control remains near chance (50.8%). Explicit deliberation removes most position dependence: raw end-to-end accuracy is 93.0% at a 256-token ceiling, with 93.8% versus 92.2% correctness when the true clause appears first versus second. Fifteen deliberative completions are truncated before the required final line. Repeating the same cells with a non-binding 1024-token ceiling changes only those failures: correctness becomes 253/256, the positional split becomes 100.0% versus 97.7%, and no previously successful final command changes. The apparent tuteo–voseo difference then is 127/128 versus 126/128 correct. Output interface, clause order, and inference budget therefore affect measured executable performance, including apparent differences between language varieties.