{
  "schema_version": "1.1",
  "elevated_threshold": 0.3,
  "positive_threshold": 0.5,
  "decision": "Regenerate if CLIP NSFW>=0.10 or EfficientNet sum(hentai,porn,sexy)>=0.30 in either image view; positive outputs>=0.50 are also flagged.",
  "detectors": [
    "LAION-AI/CLIP-based-NSFW-Detector canonical ViT-L/14 Autokeras binary classifier",
    "LAION-AI/LAION-SAFETY EfficientNetV2 B2 five-class image classifier"
  ],
  "views": [
    "original portrait using official CLIP preprocessing",
    "foreground bounding-box crop fitted to a neutral square using letterboxing"
  ],
  "scope_age_groups": [
    "baby",
    "preschool",
    "younger_schoolchild",
    "older_schoolchild",
    "younger_teen",
    "older_teen"
  ],
  "scores_are": "Model outputs,not calibrated probabilities or a guarantee of safety",
  "expected_profiles": 276,
  "expected_images": 2760,
  "clip_elevated_threshold": 0.1,
  "efficientnet_elevated_threshold": 0.3,
  "neutral_preference_threshold": 0.2,
  "accepted_prior_stricter_policy_sha256": [
    "65246ed8dcbd11a1c2b65bf08063f55701042012b57399521e1a6e70563dc31d"
  ],
  "threshold_rationale": "Baseline top-label outputs are all drawing/neutral and CLIP maximum is0.0545109. The original0.10 sum cutoff flags60percent of these SFW-classified illustrations. Separate conservative additional thresholds target37 higher-output baseline images, while preserving both original classifier-positive conditions. These extra thresholds are project heuristics,not calibrated risk probabilities.",
  "hash_encoding": "SHA256 of sorted-key compact JSON with policy_sha256 omitted",
  "policy_sha256": "a335561ec695d432f5e420a1975b2dc36f1e2f4161d4e199a79d103325feb0d0"
}