data/delhi_cd/week1_gate.json

{
  "date": "2026-07-16",
  "week1_gate_criteria": {
    "labeled_pairs_ge_25": {
      "required": 25,
      "actual": 32,
      "pass": true
    },
    "calibration_measurable": {
      "note": "best_params.json present; Feature-Based mean F1 ~0.04\u00e2\u20ac\u201c0.11 depending on sweep set",
      "artifact": "runs/calibration/best_params.json",
      "pass": true
    },
    "finetune_pipeline_clean": {
      "smoke": "runs/day4_smoke/metrics.json",
      "full_run": "runs/day5_full/",
      "pass": true
    }
  },
  "uday_day1_to_day5": {
    "day1": "compare_methods manifest mode + finetune scaffold \u00e2\u20ac\u201d done",
    "day2": "delhi_eval loader + baseline + validate_detection --delhi \u00e2\u20ac\u201d done",
    "day3": "grid search + baseline metrics + smoke train \u00e2\u20ac\u201d done",
    "day4": "data/delhi_cd 70/15/15 + smoke metrics \u00e2\u20ac\u201d done",
    "day5": "preset day5 tuning + 30-epoch full train complete"
  },
  "gates_verified_2026_07_16": {
    "parked_cars_feature_based_f1": 1.0,
    "delhi_feature_based_mean_f1": 0.0413,
    "delhi_n_pairs": 32
  },
  "decision": "GO for Week 2 (finish Day5/6 full train; Priyanka promote Feature-Based calibrated defaults if car-FP stays 1.0)",
  "caveats": [
    "Fine-tune val/test F1 still near zero on CPU smoke \u00e2\u20ac\u201d Day 5 full run in progress",
    "Do not promote AI/Hybrid defaults until fine-tune beats pretrained on held-out Delhi test",
    "inserted_buildings synthetic F1=0 under Feature-Based \u00e2\u20ac\u201d expected classical under-detect; keep as regression visibility"
  ],
  "day5_full_result": {
    "source_run": "runs/day5_full/20260716_142643",
    "epochs_ran": 30,
    "lr": 3e-05,
    "augment": true,
    "loss": "bce_dice",
    "best_val_f1": 0.0001,
    "test_mean_f1": 0.0025,
    "device": "cpu",
    "train_tiles": 22
  }
}