{ "schema_version": 1, "measured_on": "2026-10-07", "workload": { "name": "ContractNLI", "documents": 64, "decisions": 1088, "unique_rows_per_method": 1088, "source_sha256": "2582ce9de1925b85bc206419d2c51cf97519835ae04eb8e931f00f29549c7366", "text_truncated": false }, "runtime": { "gpu": "NVIDIA GeForce RTX 4090", "model": "JEV-9B BF16", "vllm": "0.31.0", "concurrency": 4, "prefix_cache": true, "cache_block_tokens": 528, "completion_tokens_per_request": 1 }, "batching": { "ready_batch_size": 64, "batches": 17, "coverage": "Every decision appears exactly once per layout; planning is performed independently inside each already-ready batch." }, "quality": { "original": { "correct": 720, "accuracy": 0.6617647058823529 }, "solo": { "correct": 804, "accuracy": 0.7389705882352942 }, "accuracy_delta": 0.07720588235294124, "paired": { "agreement": 0.7472426470588235, "same_correct": 640, "same_wrong": 173, "solo_fixes": 164, "solo_breaks": 80, "both_wrong_different": 31, "mcnemar_exact_two_sided_p": 8.088248458453298e-08, "accuracy_delta_95_percent_normal_approximation": [ 0.049442496545172, 0.10496926816071048 ] }, "labels": [ "Entailment", "Contradiction", "NotMentioned" ], "confusion_matrices_gold_rows_prediction_columns": { "original": [ [450, 30, 10], [16, 82, 6], [214, 92, 188] ], "solo": [ [398, 41, 51], [8, 88, 8], [81, 95, 318] ] } }, "fixed_work_performance": { "original_wall_seconds_sum": 293.64379544742405, "solo_wall_seconds_sum": 115.7259518885985, "original_rows_per_second": 3.705169381638792, "solo_rows_per_second": 9.401521285798916, "original_over_solo_speedup": 2.5374066115273344 }, "cache_accounting": { "solo_prompt_tokens": 2222443, "exact_reusable_prefix_tokens": 1893385, "exact_reusable_prefix_fraction": 0.8519386098991065, "block_quantized_ceiling_tokens": 1642080, "block_quantized_ceiling_fraction": 0.738862594001286, "measured_cached_tokens": 1610928, "measured_cached_fraction": 0.7248455865909722, "measured_over_block_ceiling": 0.9810289389067524, "original_measured_cached_fraction": 0.0 }, "planning_scope_boundary": { "description": "A separate diagnostic treated all 1,088 decisions as one global planning batch. The planner then preferred the 17-value hypothesis field before the 64 contracts, so contract prefixes were not exposed.", "original_accuracy": 0.6636029411764706, "solo_accuracy": 0.6571691176470589, "solo_cached_fraction": 0.02020330248368273, "original_wall_seconds": 292.8006931077689, "solo_wall_seconds": 291.07267580367625 }, "raw_artifacts": { "report": { "name": "contract-nli-quality-all-batch64-cache-enabled.json", "bytes": 2696566, "sha256": "c0e45b24aa056683fcfb5a36240e992650c6eee275caccd675b50abbdb29a5fa" }, "requests": { "name": "contract-nli-quality-all-batch64-cache-enabled-requests.jsonl", "rows": 2176, "bytes": 2144100, "sha256": "5f1e1c7351c21459119dab05c39d6bc9cca6bb808f46a3fdd4710041849c4d17" } } }