MLX_BUN_FILL_TRACE=<file.jsonl> — additionally append one record per
proposal: the proposed ids/text next to what the model's OWN logits said
at every span position (position 0 = the in-flight sample, j>0 = argmax
after ids[j-1]), for BOTH policies. Under assert the readback is
trace-only (one lm_head over the span) — the served path never pays it.
This is the list that answers "would the model have produced this span?"
without a verify pass in production.
MLX_BUN_FILL_TRACE=<file.jsonl> — additionally append one record per proposal: the proposed ids/text next to what the model's OWN logits said at every span position (position 0 = the in-flight sample, j>0 = argmax after ids[j-1]), for BOTH policies. Under
assertthe readback is trace-only (one lm_head over the span) — the served path never pays it. This is the list that answers "would the model have produced this span?" without a verify pass in production.