LFM tool-calling — models × frameworks × tests

scoring is driven by a reference LFM2 parser: the model's raw output is parsed by the reference and compared to the framework's own parse (✓/✗ in the detail pane). fail = the two disagree (the framework's parser is wrong). When they agree the parse is trusted, so the case check decides: pass, or model failed = the parse was faithful but the call is wrong (the model's own mistake) — counted as neither pass nor fail, and excluded from %.

the raw is the model's actual generation in every case: llama.cpp via __verbose, SGLang via the chat response's own logprobs token stream, vLLM via its server /tokenize prompt — so the framework-vs-reference comparison is exact.