Which function
fits the request?

The same 1,253 public BFCL v4 multiple and live_multiple cases and official gold function names were used in all three views. Within each view, Kev and Jev received identical state, instruction, candidate order, and candidate text. All three views were specified before this Kev comparison began.

The model chooses one function name from the official candidates. This does not score generated arguments, execution, parallel calls, multi-step behavior, memory, or BFCL's overall leaderboard composite. The official dataset and answers are pinned to Gorilla commit f7cf7359b7ac615a0b294831c5ba2bc95ee4a000.

The prose adapter omits some schema fields. The three views therefore vary both representation and, in the description/prose conditions, information content. Their score differences cannot isolate punctuation or formatting alone. The prose view was post-hoc in the old CLM evaluation, but was specified in advance here.

These descriptive metrics use the returned probability distribution. ECE compares top confidence with observed accuracy in 10 fixed, equal-width bins. Multiclass Brier scores the full distribution; lower is better for both.

Coverage is the share with top probability at least 0.9. Error among accepted divides wrong accepted answers by all accepted answers; confident wrong is the count. The 0.9 threshold is a descriptive display choice, not a production gate fitted or validated on independent data. No model was recalibrated using these evaluation labels. One Jev prose response rounded its probabilities to two decimal places, summing to 0.99; its returned values were preserved without renormalization.

The request hashes and source-file hashes match the historical CLM evaluation in every view. The rows below compare the same adapted task. CLM was not rerun. These values also must not be equated with the CLM team's published 95.2% BFCL figure: its exact case set and scoring adapter were not available in the earlier report.