# Teen and child classifier recheck

Actually changed dataset PNGs: **241**, across **35 profiles**, including earlier stricter-filter replacements and the final seven-profile round. The [replacement audit](metadata/minor-safety/replacement-audit.json) distinguishes regenerated images from byte-identical retained expressions. A changed neutral requires a new coherent ten-expression series.

The [supplemental age19 report](metadata/minor-safety/age19-validation.json) additionally covers **41 existing19-year-old sprites**:40 original-base expressions and one previously public job-outfit pilot. Status: `passed`; flags: **0**. These young-adult-labelled profiles are outside the2760-category count but inside the literal under-20 check. Future human expression and professional-outfit generations with a numerical age below20 are screened inline with both models and both views before becoming uploadable. Raised candidates automatically retry with recorded actual seeds; failed searches cannot be uploaded. The [control-flow validation](metadata/minor-safety/future-teen-screening-validation.json) tests retry exhaustion, checksum binding, thresholds and concurrent plan updates; it is separate from actual model-output measurements.

Requested dataset scope: **276 identities × 10 expressions = 2,760 images**, covering babies, preschoolers, both school-age groups and both teen groups, including 18-year-olds in the older-teen category. **Six previously published child/teen model-gallery examples** are checked separately; their scores and any replacement commit are in the gallery reports. The published snapshot has checked **2,760 current dataset images**; **2,760 pass the operational criteria** and **0 require further processing**. This is an ongoing snapshot unless the linked completion report explicitly says complete.

The exact [policy](metadata/minor-safety/policy.json), [detector provenance](metadata/minor-safety/detector-provenance.json), [per-image scores](metadata/minor-safety/scores.jsonl) and [current validation](metadata/minor-safety/current-validation.json) are checksum-specific. They identify the actual PNG bytes, not only filenames. Already-public TAR contents are matched to their remote SHA256 and byte size before scoring the identical local images.

Two official LAION image models are used: the [CLIP ViT-L/14 binary classifier](https://github.com/LAION-AI/CLIP-based-NSFW-Detector) and the [LAION-SAFETY EfficientNetV2 B2 classifier](https://github.com/LAION-AI/LAION-SAFETY). The latter's drawing/neutral classes are SFW; its hentai/porn/sexy softmax scores are summed for the NSFW score. Each model sees the original portrait and a foreground crop fitted into a neutral square. CLIP uses L2-normalized projected OpenAI ViT-L/14 image embeddings; EfficientNet uses RGB0–1 and its documented260-pixel image input.

The active additional operational thresholds are **CLIP ≥0.10 or EfficientNet NSFW sum ≥0.30 in either view**. Ordinary positive classifier outputs are included by these conservative cutoffs; outputs at least0.50 are also recorded separately as positive. These additional thresholds are project criteria, not documented guarantees or a calibration for this illustrated domain. A flagged score does not establish that an image depicts sexual content. Model scores can produce false positives and false negatives.

The [baseline class-output analysis](metadata/minor-safety/baseline-classification-summary.json) separates these additional project flags from the models' ordinary classifications. Across all2,760 original images and both views, EfficientNet's highest-scoring class is always SFW:5,406 drawing views and114 neutral views, with no hentai/porn/sexy top labels. The maximum CLIP NSFW output is0.0545109. Thus all original top-label/binary classifications are SFW; the1,662 flags come from the extra0.10 sum threshold. This illustrated-domain sensitivity is an observed limitation, not evidence of1,662 sexual images.

The original stricter0.10 policy is preserved in the [baseline policy](metadata/minor-safety/baseline-policy.json). The [declared transition](metadata/minor-safety/policy-transition.json) documents the implementation judgment behind separate cutoffs after the optional threshold-preference question; no user-selected answer was inferred. Under the0.10 CLIP/0.30 EfficientNet rule,37 original images from seven identities receive additional precautionary regeneration. Saved raw detector outputs are reevaluated against the new criteria without changing pixels or model weights; historical original scores remain preserved. Earlier0.10-screened replacements remain valid under the new rule. The six model-gallery examples already passed the stricter0.10 check.

New teen/child TARs are held until the exact image checksums pass. During repair the default/base Datasets configurations whitelist already-uploaded, checksum-approved original TARs; entire flagged TARs are temporarily omitted. The [stream-filter snapshot](metadata/minor-safety/public-stream-filter.json) records that reduced subset. Previously published TARs with unresolved flags have also been temporarily removed from current main, as recorded in the [public quarantine audit](metadata/minor-safety/public-quarantine.json); old Git revisions remain historical versions. Screened replacements return under the same TAR names. The complete500-person plan remains the target, not a claim about this temporarily filtered stream. Full wildcard streaming is restored after the recheck and all replacement uploads complete. Replacements use fully covered, age-appropriate opaque clothing and are checked again. If a neutral reference changes, all ten emotions are regenerated against the new reference. When the original neutral already passes, its exact pixels and other passing expressions are normally retained; only flagged expressions are regenerated against that same neutral. Difficult cases may instead receive a coherent fully covering outfit variant for the entire ten-image family, recorded in the [recipe overrides](metadata/minor-safety/profile-recipe-overrides.json). Every released group still has ten passing expressions. Published replacement audits distinguish changed and retained PNG hashes, parameters and the corresponding TAR changes. Individual image assets remain inside TARs.

New neutral-reference searches prefer scores below0.20 to leave room for later expression edits. Within a16-candidate search, the best neutral that meets both active model/view criteria can also be used. Already accepted candidate PNGs are checkpointed with exact prompt, seed and reference linkage. New families generated under the active policy additionally wait for a policy-specific visual review gate before upload. Full PNG/TAR integrity validation and anonymous reference-index checks are separate from classifier judgments.