{
  "schema": "ogi.weak-mask.reference.protocol.v1",
  "status": "draft_pending_root_review",
  "scope": "exploratory_autogenerated_binary_mask_reference",
  "target": "native published mask pixel != 0; algorithmic reference agreement only",
  "probability_semantics": "uncalibrated binary mask scores; no chemical confidence",
  "confirmation_allowed": false,
  "dataset": "baselab/controlled-diet-methane-dataset",
  "revision": "9475f96c32dd4764471e4358ca6659b0dd348ecd",
  "selection_sha256": "467c1c8595dc1a23365f80038771c86d35542640e77c6b9924640cb8655a2adb",
  "qualification_pins": {
    "access-boundary.json": "d5b503817451f6b86e207103bab9e8aeee2503d641e15f362b5ff69aea051183",
    "cd-filename-class-metadata.csv": "8371e6ce7c80a5e4a6d6060ec74951601953bb4210126c4446bddc524a53b6d5",
    "recording-asset-metadata.json": "5465ec7c10bc003dab8fe316767ef586fa399c031914f2aff4c61241572c5cb6",
    "additional-primary-metadata.json": "2aee168f4f65a7a8148b4535aa4caa10d1cfc887d7fe2471073b2d590a1bbb8f"
  },
  "roles": {"development_train": 18, "development_tune": 12, "development_compare": 12},
  "per_recording_counts": {"development_train": 6, "development_tune": 4, "development_compare": 4},
  "minimum_frame_gap": 30,
  "native_size_wh": [640, 480],
  "model_size_wh": [320, 240],
  "rgb_resampling": "Pillow RGB BILINEAR; uint8 then float64 / 255",
  "training_target": "native nonzero binary mask 2x2 mean foreground fraction; no nearest label loss",
  "metric_grid": "native 640x480 mask; model probabilities/predictions nearest repeated 2x in each axis",
  "coverage_diagnostic": "nearest resized binary model mask only for descriptive coverage; never used as training target or primary score",
  "features": ["R", "G", "B", "channel_mean_gray", "gray_3x3_mean", "gray_3x3_std", "abs_gray_dx", "abs_gray_dy"],
  "feature_edges": "edge padding; central differences / 2; population local standard deviation",
  "radiometric_claim": false,
  "arms": ["always_background", "always_foreground", "training_spatial_prior", "dark_gray", "bright_gray", "feature_logistic", "tiny_mlp"],
  "learned_l2_grid": [0.0, 0.001, 0.01],
  "threshold_grid": [0.01, 0.05, 0.1, 0.25, 0.5, 0.75, 0.9, 0.95, 0.99],
  "threshold_operator": ">=",
  "steps": 160,
  "max_sampled_pixels_per_frame": 8192,
  "sample_seed": "ogi-weak-mask-pixel-sample-v1",
  "sample_rule": "PCG64 seeded by first 16 SHA256 hex digits of sample_seed|filename; sorted choice without replacement; identical indices across learned candidates; no mask stratification",
  "standardization": "training sampled features only; population std floored at 1e-6; fixed for tune and compare",
  "positive_weight": "training sampled soft-negative mass / soft-positive mass; negative weight 1; refuse missing class mass",
  "loss": "mean(pos_weight*y*softplus(-logit)+(1-y)*softplus(logit)) + l2/2*sum(nonbias weights squared)",
  "logistic": {"learning_rate": 0.1, "initialization": "all zeros"},
  "mlp": {"hidden": 8, "activation": "tanh", "learning_rate": 0.05, "seed": 1107, "initialization": "PCG64 normal weights with std sqrt(1/fan_in), zero biases"},
  "candidate_budget": {"per_learned_arm": 3, "maximum_gradient_updates": 960, "candidate_wall_seconds": 180, "prepare_wall_seconds": 1500},
  "selection": "maximum development_tune mean of recording means of native frame IoU; ties keep candidate order then threshold order; no compare selection/refit",
  "metric_names": ["iou", "dice", "precision", "recall", "background_pixel_overprediction_ratio"],
  "empty_reference_rule": "empty prediction and reference: IoU=Dice=null; undefined precision/recall/background ratio=null with defined-count reporting; empty prediction against positive reference gives IoU=Dice=recall=0 and precision=null",
  "raw_probability_reload_tolerance": {"atol": 1e-10, "rtol": 1e-9},
  "input_byte_ceiling": 30000000,
  "threads": 1,
  "dependency_pins": {"numpy": "2.4.2", "Pillow": "12.1.0"},
  "code_license": "MIT",
  "data_attribution_license": "CC-BY-4.0; conflicting source metadata CC0-1.0 retained as unresolved",
  "pretrained_checkpoint": null,
  "no_network": true
}