{
  "study_id": "FL-STUDY-001",
  "candidates": {
    "Gemma 3 27B": "study/c1_gemma3",
    "MedGemma 27B": "study/c2b_medgemma_clean"
  },
  "n_common_images": 35,
  "reliability": {
    "Gemma 3 27B": {
      "n": 35,
      "clean_first_pass": 35,
      "repair_used": 0,
      "hard_fail": 0,
      "validity_pct": 100.0
    },
    "MedGemma 27B": {
      "n": 35,
      "clean_first_pass": 35,
      "repair_used": 0,
      "hard_fail": 0,
      "validity_pct": 100.0
    }
  },
  "tier_distribution": {
    "Gemma 3 27B": {
      "routine_log": 0,
      "watch_closely": 0,
      "contact_clinician": 0,
      "urgent_care": 35
    },
    "MedGemma 27B": {
      "routine_log": 1,
      "watch_closely": 0,
      "contact_clinician": 0,
      "urgent_care": 34
    }
  },
  "latency_proxy_s_per_image": {
    "Gemma 3 27B": 21.0,
    "MedGemma 27B": 21.0
  },
  "field_agreement": {
    "foot": {
      "agree_pct": 100.0,
      "kappa": 1.0
    },
    "surgery_site_visible": {
      "agree_pct": 100.0,
      "kappa": 1.0
    },
    "redness": {
      "agree_pct": 57.1,
      "kappa": 0.067
    },
    "swelling": {
      "agree_pct": 22.9,
      "kappa": 0.015
    },
    "drainage_or_fluid": {
      "agree_pct": 57.1,
      "kappa": 0.34
    },
    "discoloration": {
      "agree_pct": 94.3,
      "kappa": 0.314
    },
    "open_wound_or_skin_break": {
      "agree_pct": 80.0,
      "kappa": 0.0
    },
    "scab_or_crust": {
      "agree_pct": 91.4,
      "kappa": -0.019
    },
    "bruising": {
      "agree_pct": 60.0,
      "kappa": 0.167
    },
    "maceration": {
      "agree_pct": 71.4,
      "kappa": 0.448
    },
    "pressure_or_friction_signs": {
      "agree_pct": 0.0,
      "kappa": 0.0
    }
  },
  "tier_agreement": {
    "agree_pct": 97.1,
    "kappa": 0.0
  },
  "s2_diagnosis_dosing_flags": {
    "Gemma 3 27B": [],
    "MedGemma 27B": []
  },
  "wound_positive_labels": "PENDING owner/expert labeling (see PROTOCOL S1)",
  "note": "Deterministic gates measure reliability/safety/agreement, NOT clinical accuracy. Which model observes better is deferred to the Tier-2 blinded clinician packet.",
  "chain": {
    "links": 70,
    "verify_ok": true,
    "tamper_demo_breaks_at": 35
  }
}