The same model scores 0.00 or 0.96 on tool calling depending only on the serving adapter

arXiv·high signalXBlueskyLinkedInCopy link

This paper shows the reported tool-call rate is a property of the model-interface stack, not the model. Holding weights, cases, decoding and seeds fixed on BFCL v4 and changing only the serving adapter moves the score between 0.00 and 0.96; a 2x2 over chat template and parser puts both main effects at exactly zero with all of the effect in the interaction. On tau-bench's 115 retail tasks the same swap moves server-parsed calls from 0 to 636, and inside verl's AgentLoop at 7B, 45 of 115 generations carry a complete call while 0 are accepted or executed. The authors release a 98-line preflight check that catches every silent failure they found.