Previously, the system used a single set of candidate token IDs for scoring all prompt tokens uniformly, limiting the ability to tailor scoring per token position in causal language model prompts.

The lack of per-row candidate IDs meant that on-policy distillation teachers could not score each response position against the student's own top-K tokens, reducing scoring accuracy and flexibility.

Introduced a per-row integer array input for candidate token IDs, allowing each prompt token position to have its own candidate list. This array is validated, normalized to a C-contiguous int32 ndarray, and integrated into both Python and Rust frontends with consistent validation and zero-copy data paths. Padding with -1 is handled by masking to -inf in logprob outputs.

Using per-row candidate IDs significantly improves scoring precision in prompt token evaluation while maintaining performance through efficient array handling and consistent cross-language validation.