{
  "schema": "semif-phase1-summary-v1",
  "date": "2026-09-16",
  "claim": "Direct option logits are the strongest tested open general-decision baseline; a native reranker remains a retrieval control and is not a drop-in universal decision operator.",
  "models": {
    "direct_logits": {
      "source": "Qwen/Qwen3.5-4B",
      "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a"
    },
    "reranker": {
      "source": "Qwen/Qwen3-Reranker-4B",
      "revision": "22e683669bc0f0bd69640a1354a6d0aebcfeede5"
    }
  },
  "semantic_quality": {
    "authored_144_mean_family_balanced_accuracy": {
      "direct_logits": 0.8132381608,
      "reranker": 0.6250733042
    },
    "wanli_256_balanced_accuracy": {
      "direct_logits": 0.6365253078,
      "reranker": 0.5219790242
    },
    "typesafe_public_102_equal_case_modal_agreement": {
      "direct_logits": 0.8453205128,
      "reranker": 0.5595421245,
      "published_jev": 0.8831410256
    },
    "typesafe_public_102_tv_distance": {
      "direct_logits": 0.1770346368,
      "reranker": 0.4440789631,
      "published_jev": 0.1268250717
    },
    "every_judge_grid_36_accuracy": {
      "direct_logits": 0.8055555556,
      "reranker": 0.6944444444
    },
    "every_action_firewall_10_accuracy": {
      "direct_logits": 0.7,
      "reranker": 0.7
    },
    "every_code_rag_recall_at_1": {
      "direct_logits": 1.0,
      "reranker": 1.0
    },
    "every_company_brain_recall_at_1": {
      "direct_logits": 0.9285714286,
      "reranker": 0.9285714286
    }
  },
  "paired_reranker_minus_direct": {
    "authored_balanced_accuracy": {
      "difference": -0.1881648566,
      "bootstrap_95": [
        -0.2557270628,
        -0.1200396825
      ]
    },
    "wanli_balanced_accuracy": {
      "difference": -0.1145462836,
      "bootstrap_95": [
        -0.1844227325,
        -0.0444315396
      ]
    }
  },
  "perturbations_36": {
    "direct_logits": {
      "base_balanced_accuracy": 0.7227513228,
      "option_reversal": {
        "balanced_accuracy": 0.8132275132,
        "argmax_flips": 10
      },
      "criterion_wrapper": {
        "balanced_accuracy": 0.7058201058,
        "argmax_flips": 9
      },
      "irrelevant_context": {
        "balanced_accuracy": 0.8206349206,
        "argmax_flips": 4
      },
      "missing_evidence_confident_non_insufficient_at_0_8": 1
    },
    "reranker": {
      "base_balanced_accuracy": 0.5301587302,
      "option_reversal": {
        "balanced_accuracy": 0.4984126984,
        "argmax_flips": 2
      },
      "criterion_wrapper": {
        "balanced_accuracy": 0.6465608466,
        "argmax_flips": 9
      },
      "irrelevant_context": {
        "balanced_accuracy": 0.5634920635,
        "argmax_flips": 13
      },
      "missing_evidence_confident_non_insufficient_at_0_8": 1
    }
  },
  "shape777": {
    "fixture_sha256": "8dcf414b12fc2684e3c4ca5f3ebfd3f525f5346fec4a9bc67eb65138101f55f1",
    "hardware": "NVIDIA GeForce RTX 3090",
    "timing_scope": "Warm-loaded BF16 model; prompt construction/tokenization, transfers, forward and CPU readout included; model load and output write excluded.",
    "direct": [
      {
        "mode": "fresh_batch1",
        "wall_seconds": 333.063398068,
        "judgments_per_second": 2.3328891872,
        "state_p50_seconds": 8.985686894,
        "peak_cuda_bytes": 8894241280,
        "argmax_flips_vs_fresh": 0
      },
      {
        "mode": "serial_prefix",
        "wall_seconds": 72.263639052,
        "judgments_per_second": 10.7522954863,
        "state_p50_seconds": 1.932188766,
        "peak_cuda_bytes": 8991294976,
        "argmax_flips_vs_fresh": 5
      },
      {
        "mode": "parallel_suffix",
        "wall_seconds": 38.792141172,
        "judgments_per_second": 20.0298301802,
        "state_p50_seconds": 1.047665117,
        "peak_cuda_bytes": 11611157504,
        "argmax_flips_vs_fresh": 6
      }
    ],
    "reranker": [
      {
        "pair_batch_size": 1,
        "wall_seconds": 417.30154415,
        "judgments_per_second": 1.861962916,
        "state_p50_seconds": 11.284789991,
        "peak_cuda_bytes": 8192491520,
        "argmax_flips_vs_batch1": 0
      },
      {
        "pair_batch_size": 4,
        "wall_seconds": 441.661915525,
        "judgments_per_second": 1.759264208,
        "state_p50_seconds": 11.94427757,
        "peak_cuda_bytes": 8619127296,
        "argmax_flips_vs_batch1": 51
      },
      {
        "pair_batch_size": 8,
        "wall_seconds": 435.194096286,
        "judgments_per_second": 1.7854102494,
        "state_p50_seconds": 11.756876173,
        "peak_cuda_bytes": 9185073152,
        "argmax_flips_vs_batch1": 54
      }
    ]
  },
  "boundaries": [
    "No live Jev endpoint was run.",
    "The TypeSafe comparison is 102 public rows across 20 cases, not the reported 711-row aggregate.",
    "The shape777 fixture matches the 37x21 count only; inputs, serving stack and hardware are not equivalent to Jev.",
    "Conditional option scores are not established as calibrated confidence."
  ],
  "decision_vs_compact_generation21": {
    "model": "Qwen/Qwen3.5-4B",
    "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
    "hardware": "NVIDIA GeForce RTX 3090",
    "workload": "Same frozen model, exact state, criteria, hardware and process. Direct path returns two-option distributions; generation emits only an ordered yes/no JSON array.",
    "direct_parallel": {
      "runs_seconds": [
        1.036716750008054,
        1.019257096981164,
        1.023019488027785
      ],
      "median_seconds": 1.023019488027785,
      "generated_output_tokens": 0,
      "result": "21 two-option probability distributions"
    },
    "compact_generation": {
      "runs_seconds": [
        5.286041187006049,
        5.331690029008314,
        5.413413154950831
      ],
      "time_to_first_token_seconds": [
        0.48786429001484066,
        0.4888657660339959,
        0.4883691109716892
      ],
      "output_tokens": [
        111,
        111,
        111
      ],
      "median_seconds": 5.331690029008314,
      "median_output_tokens": 111,
      "all_runs_valid_complete_arrays": true,
      "agreement_with_direct_argmax": 0.8571428571428571
    },
    "wall_time_ratio_generation_over_direct": 5.211718927551365
  },
  "verbose_generation_reference21": {
    "model": "Qwen/Qwen3.5-4B",
    "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
    "hardware": "NVIDIA GeForce RTX 3090",
    "workload": "Same frozen BF16 model, owned state, and 21 binary criteria; direct two-option distributions versus one autoregressively generated JSON object.",
    "direct_parallel": {
      "runs_seconds": [
        1.06307857,
        1.065514043,
        1.070278497
      ],
      "median_seconds": 1.065514043,
      "generated_output_tokens": 0,
      "result": "21 two-option probability distributions"
    },
    "naive_generation": {
      "input_tokens": 2927,
      "time_to_first_token_seconds": 0.642879314,
      "total_seconds": 18.229238557,
      "output_tokens": 405,
      "tokens_per_second_after_first": 22.9726836662,
      "valid_json": true,
      "complete_21_key_schema": true
    },
    "wall_time_ratio_generation_over_direct": 17.1083982207,
    "status": "verbose reference; superseded as headline by compact ordered-array baseline"
  }
}