{
  "status": "passed",
  "date": "2026-10-04",
  "execution": "server GPU; pre-existing cached backbone; network disabled",
  "method": "QIME-TF-MMR",
  "num_questions": 8855,
  "topk": 256,
  "mmr_lambda": 0.7,
  "backbone_revision": "963121bfb9c625475f65b08fb54990ce9c4e7a1a",
  "questions_sha256": "79b00c0675cdf38a0a5de0bbae7c055a0c3072bf8a868ae362d739fe85ea5e74",
  "native_source_sha256": "b5916b6b6e724ca7663917229f8d3979700b23ad3690a18bf2bd2607aac30e1c",
  "smoke_input_count": 6,
  "output_shape": [
    6,
    8855
  ],
  "output_dtype": "float32",
  "active_dimensions_per_text": [
    256,
    256,
    256,
    256,
    256,
    256
  ],
  "native_comparison": {
    "same_backbone_and_question_tensors": true,
    "binary_disagreements": 0,
    "mmr_selection_order_equal": true
  },
  "native_full_batch_selection_order_equal": true,
  "chunked_batch_comparison": {
    "batch_size": 2,
    "binary_disagreements": 0
  },
  "synthetic_mmr_configurations_passed": 9,
  "additional_checks": [
    "single-string input",
    "empty input",
    "invalid input validation",
    "question-bank tamper detection"
  ],
  "versions": {
    "torch": "2.6.0+cu124",
    "sentence_transformers": "3.4.1",
    "huggingface_hub": "0.36.0",
    "numpy": "1.26.4"
  },
  "paper_benchmark_reproduction": "reproduced",
  "paper_benchmark_reproduction_details": "benchmark_reproduction.json",
  "paper_benchmark_completed_tasks": 12,
  "paper_benchmark_total_tasks": 12,
  "full_twelve_task_reproduction_complete": true,
  "paper_benchmark_author_reported_completed_tasks": [
    "MedrxivClusteringP2P",
    "MedrxivClusteringS2S",
    "ClusTREC-Covid"
  ]
}