pvts-bench

Cumulative 2x2 benchmark (probe vs decode, HF vs vLLM), one row per (model, dataset, gpu, vLLM version). All cells of a run share the same full-schema harvest prompt; prefix caching off; one engine/model per process; max_new_tokens=256 uniformly (gen_hit_cap verifies the cap doesn't bind).

Accuracy: only Verified rows support a probe-vs-decode accuracy head-to-head (Self-Distill labels are the model's own outputs; decode there measures self-consistency). For Self-Distill and ToolACE rows the operative metrics are agree_probe_decode and the agree_cov* columns: fidelity of the probe to the decision the model itself would make, unconditionally and under confidence gating.

model_id	dataset_id	acc_probe	acc_decode	agree_probe_decode	agree_cov90	gen_hit_cap	ms_probe_vllm	ms_decode_vllm	probe_over_decode_vllm	probe_over_prefill_vllm	depth_ratio
llama-3.2-3b	toolace-self-distill	0.7392	0.9495	0.7337	0.7684	0.0078	8.2716	25.2229	0.3279	0.5758	0.5714
llama-3.2-3b	xlam-human-gold	0.9626	0.9919	0.9623	0.9890	0.0000	4.4108	11.2298	0.3928	0.5838	0.5714
llama-3.2-3b	xlam-self-distill	0.8639	0.9892	0.8569	0.8989	0.0000	4.3521	11.4892	0.3788	0.5815	0.5714
mistral-7b-v0.3	toolace-self-distill	0.7305	0.9320	0.7380	0.7703	0.1965	12.0094	52.1968	0.2301	0.5023	0.5000
mistral-7b-v0.3	xlam-human-gold	0.9542	0.9663	0.9298	0.9620	0.1123	6.5980	26.0303	0.2535	0.5040	0.5000
mistral-7b-v0.3	xlam-self-distill	0.7830	0.9785	0.7773	0.8069	0.0889	5.8097	23.1129	0.2514	0.5016	0.5000
qwen-2.5-7b	toolace-self-distill	0.5187	0.9923	0.5206	0.5383	0.0381	15.3468	37.5257	0.4090	0.5684	0.5714
qwen-2.5-7b	xlam-human-gold	0.9037	0.9951	0.9025	0.9423	0.0000	8.5390	19.7460	0.4324	0.5711	0.5714
qwen-2.5-7b	xlam-self-distill	0.7452	0.9997	0.7454	0.7780	0.0000	7.3090	17.3397	0.4215	0.5736	0.5714

raw/ holds per-run arm outputs (timings, decode predictions, inline probe predictions where 05a had no entry — see probe_pred_source per row).