add rfdetr_v13 (384px fp16): 92% real-secret recall @0.8 vs v12 0%, 0% false-fire

fp16 re-export: 54MB, 2x smaller+faster, -0.6pp vs v11 fp32

restore rfdetr_v9.onnx — pinned by the Tinfoil privacy-filter Dockerfile (do not delete until that pin moves to v11)

chore: prune superseded v8/v9 weights + stale v8 inference script (keep v11)

chore: rewrite NOTICE (RF-DETR attribution, not OPF; drop dataset disclosure); prune superseded v8/v9 weights, keep v11

docs: rewrite card — screenpipe's own detector, v11 numbers, drop training-data internals

add rfdetr_v11 (realworld-augmented, 512x512) — fixes v9 under-detection on real screens

add rfdetr_v9.safetensors (MLX runtime, sibling of rfdetr_v9.onnx)

add rfdetr_v9.onnx — retrained on 905-img augmented bench, 0% FP across 42+ surfaces at thr 0.50

examples: end-to-end inference + redaction demo

notice: attribution to RF-DETR + WebPII

license: CC BY-NC 4.0

docs: model card with bench numbers + privacy contract

model: rfdetr_v8 ONNX (108 MB)

initial commit