{ "schema": "cortiq-model-card-comparison/1", "model_file_sha256": "ed9b8ec2bbfe9e9fd30f14a5eaf82314f38bc7e7510a39772baa2de3801d79b1", "model_bytes": 304520292, "runtime": "cortiq 0.7.8", "oracle_in_hybrid_benchmark": "deepseek/deepseek-v4.1-flash", "benchmarks": [ { "id": "banking77", "name": "BANKING77", "n": 3080, "labels": 77, "local_correct": 2875, "local_accuracy_pct": 93.34, "jev_correct": 2636, "jev_accuracy_pct": 85.58, "local_wilson95_pct": [ 92.41, 94.17 ], "jev_wilson95_pct": [ 84.3, 86.78 ], "mcnemar_p": 9.72e-33, "accepted": 2791, "accepted_correct": 2714, "accepted_accuracy_pct": 97.24, "coverage_pct": 90.62, "local_p50_ms": 3.92, "local_p95_ms": 5.27, "cortiq_https_p50_ms": 78.4, "cortiq_https_p95_ms": 141.6, "jev_http_p50_ms": 348.33, "jev_http_p95_ms": 426.03, "jev_latency_n": 3080, "jev_latency_concurrency": 4, "hybrid_correct": 2893, "hybrid_accuracy_pct": 93.93, "hybrid_oracle_calls": 282, "hybrid_api_usd_per_million": 3.01, "jev_api_usd_per_million": 183.69 }, { "id": "clinc150", "name": "CLINC150", "n": 4500, "labels": 150, "local_correct": 4328, "local_accuracy_pct": 96.18, "jev_correct": 4354, "jev_accuracy_pct": 96.76, "local_wilson95_pct": [ 95.58, 96.7 ], "jev_wilson95_pct": [ 96.2, 97.23 ], "mcnemar_p": 0.099, "accepted": 4145, "accepted_correct": 4091, "accepted_accuracy_pct": 98.7, "coverage_pct": 92.11, "local_p50_ms": 3.79, "local_p95_ms": 4.57, "cortiq_https_p50_ms": 78.5, "cortiq_https_p95_ms": 134.6, "jev_http_p50_ms": 391.51, "jev_http_p95_ms": 528.49, "jev_latency_n": 200, "jev_latency_concurrency": 1, "hybrid_correct": 4386, "hybrid_accuracy_pct": 97.47, "hybrid_oracle_calls": 351, "hybrid_api_usd_per_million": 3.53, "jev_api_usd_per_million": 271.61 }, { "id": "massive", "name": "MASSIVE", "n": 2974, "labels": 60, "local_correct": 2562, "local_accuracy_pct": 86.15, "jev_correct": 2551, "jev_accuracy_pct": 85.78, "local_wilson95_pct": [ 84.86, 87.34 ], "jev_wilson95_pct": [ 84.48, 86.99 ], "mcnemar_p": 0.607, "accepted": 1615, "accepted_correct": 1581, "accepted_accuracy_pct": 97.89, "coverage_pct": 54.3, "local_p50_ms": 3.0, "local_p95_ms": 3.58, "cortiq_https_p50_ms": 70.4, "cortiq_https_p95_ms": 117.9, "jev_http_p50_ms": 346.37, "jev_http_p95_ms": 492.93, "jev_latency_n": 200, "jev_latency_concurrency": 1, "hybrid_correct": 2617, "hybrid_accuracy_pct": 88.0, "hybrid_oracle_calls": 1335, "hybrid_api_usd_per_million": 8.48, "jev_api_usd_per_million": 110.86 } ], "notes": [ "Local accuracy is top-1 over every test row before abstention; gate accuracy and coverage are separate.", "Local CPU latency: full text to decision, Apple M4, one CPU thread, 50 warmups, all test rows.", "Cortiq HTTPS latency: same client Mac to deployed server, sequential, no oracle, all test rows.", "Jev latency includes network; BANKING77 uses four workers, CLINC150 and MASSIVE serial samples of 200.", "Different hardware and concurrency: this is not a hardware-normalized throughput comparison.", "Hybrid costs: served cache/learning run with stored DeepSeek answers replayed and 24 live MASSIVE calls; not a timing run.", "Per-million fees extrapolate recorded costs, excluding hardware, electricity and operations.", "Reused public benchmarks, unequal supervision: CMF train+dev vs two examples per label for Jev.", "Current model uses BAAI bge-small-en-v1.5 compressed by Cortiq NVG, not the separate Embryo experiment." ], "sources": [ { "path_in_project": "artifacts/decision-v4-20260926/gates-release.json", "sha256": "47d3039559e0f87dee0c6dae9adc3bbd63acc52b05d4859f2de82cf7a0a9ba03" }, { "path_in_project": "reports/decision-v4-20260926/live/live.json", "sha256": "c8769af100e3138291675ab3eed20caa3ecaa780b6b81f4cecbde96dc29f87f5" }, { "path_in_project": "reports/decision-v4-20260926/article/stored.json", "sha256": "4d6918a2d98f16c8b4cdda7759fef98e6343f091fa7723e8ea938d9fb696db47" }, { "path_in_project": "reports/decision-v4-20260926/eval/stream-final.json", "sha256": "0bb19de37cee315a5a5ad4563abcf2944a80c49ef05984554a4fc31012702f50" } ] }