source: the INT8 recipe decision is D818 (renumbered from D798)

source: convert_lib — a Detect head must hold the box branch (cv2.* / one2one_cv2.*)

source: D798 INT8 recipe (Detect head in float, MIXED, full bias correction); archive the 2026-10-08 convert_lib

INT8 IRs re-quantised with the D798 recipe: MIXED, full bias correction, Detect head in float (OpenVINO 2026.2.1 + NNCF 3.2.0)

archive: INT8 IRs before the D798 recipe (decode/head quantised, preset PERFORMANCE)

source: centred INT8 calibration; keep the pre-2026-10-08 converter for the IRs that kept it

INT8 IRs re-quantised with centred letterbox calibration (COCO val2017, OpenVINO 2026.2.1 + NNCF 3.2.0)

archive: INT8 IRs calibrated with top-left padding (before 2026-10-08 re-quantisation)

source: every CoreML recipe exports FP16 (YOLOv9 packages replaced in place)

YOLOv9 CoreML packages re-exported FP16 (run on the Apple Neural Engine), replaced in place

fix: reduced YOLO26 CoreML replaced in place; drop the -static copies

source: YOLO26 @320/@256 CoreML recipe is static (batch 1)

fix: YOLO26 @320/@256 CoreML re-exported with a static batch

edgetpuSets cards: the decision record is CamStack D724

Co-compiled Edge TPU sets root-plate-bird, face-plate-bird, plate-bird (CamStack D722)

Coral (Edge TPU) builds of yunet-2023mar and yolox-tiny-plate (CamStack D710 addendum)

docs: bird-classifier card lists the Edge TPU build

feat: bird-classifier Edge TPU build (tflite/bird-classifier_edgetpu.tflite, D710)

docs: notices, licences and source/ after the plate detector swap (D718)

chore: delete plateDetection/yolov8-plate (replaced by yolox-tiny-plate, D718)

feat: plateDetection/yolox-tiny-plate (open-lpr YOLOX-tiny, Apache-2.0)

source: one rebuild recipe per hosted AGPL file; yolov8n-plate provenance

docs: model notice, rewrap the Model Studio paragraph

docs: Model Studio's third-party import catalogs leave the model notice

Squash history: mirror holds only the models CamStack offers (2026-10-06)