gpu_results.json

{
  "paper": {
    "challenge_id": "64223",
    "openreview_id": "PoRigyDOcC",
    "arxiv_id": "2605.11744"
  },
  "environment": {
    "device": "cuda",
    "device_name": "Tesla T4",
    "dtype": "float32",
    "torch": "2.7.1+cu128",
    "numpy": "2.2.6",
    "trials": 256
  },
  "thresholds": {
    "gradient_relative_error_max": 2.9802322387695312e-05,
    "value_absolute_error_max": 2.9802322387695312e-05,
    "forward_absolute_error_max": 2.9802322387695312e-05,
    "misaligned_difference_min": 1e-05,
    "truncation_falsifier_min": 1e-07,
    "local_gradient_min": 1e-08,
    "retrieval_gradient_max": 0.0
  },
  "metrics": {
    "gradient_relative_error_max": 4.0176314452108995e-07,
    "gradient_relative_error_median": 7.584128335689134e-08,
    "value_absolute_error_max": 4.01415820050488e-06,
    "training_inference_forward_difference_max": 1.617075742199603e-07,
    "misaligned_reset_difference_min": 0.0009355164077803346,
    "truncated_vs_full_gradient_difference_min": 0.18395258298292394
  },
  "attention_audit": {
    "configuration": {
      "layers": 4,
      "heads": 8,
      "segment_tokens": 16,
      "long_layers": [
        1,
        3
      ],
      "long_heads": [
        0,
        1,
        2,
        3
      ],
      "carried_tokens": 4,
      "retrieved_tokens": 3
    },
    "executed_attention_repeats": 256,
    "access_rule_violations": 0,
    "local_carried_gradient_min": 3.452453851699829,
    "detached_retrieval_gradient_max": 0.0,
    "unique_access_patterns": [
      [
        "carried-local",
        4
      ],
      [
        "detached-retrieval",
        3
      ],
      [
        "within-only",
        0
      ]
    ]
  },
  "official_source_audit": {
    "locked_commit": "a875317b1fc1e2a00a8dee04a8dbc844219f4077",
    "checks": {
      "segment_loop": true,
      "head_partition": true,
      "retrieval_search": true,
      "retrieval_detached": true,
      "carried_kv_tail": true
    },
    "all_pass": true
  },
  "gates": {
    "claim_1_same_segment_forward_semantics": true,
    "claim_2_tbptt_exact_gradient": true,
    "claim_3_training_inference_alignment": true,
    "claim_4_sparse_retrieval_architecture": true
  },
  "all_pass": true,
  "duration_seconds": 47.437765085000024
}