nx_coco_eval.nx

nx_coco_eval.nx -- REAL industry COCO keypoint scorer (GT-bbox variant). Reads the sovereign-fetched COCO GT JSONL (whyen-wang/coco_keypoints -> keypoints_validation.jsonl, format per line: {"image":"..jpg","bboxes":[[x,y,w,h],..],"keypoints":[[[x,y,v]x17],..]}) and a predictions JSONL in the SAME shape, and computes per-person OKS (official COCO sigmas, s^2 = bbox area) + COCO AP@[.5:.95] via the gated nx_coco_oks / nx_coco_ap. This is the EXTERNAL grader that ends the self- validation: GT is COCO's real annotations, not our own model's output. GT-bbox eval: predictions are produced by running the pose net on each GT bbox, so pred person i <-> GT person i (no detector, no FPs) => AP@t = fraction of persons with OKS>=t (what nx_coco_ap computes). This is the standard, DISCLOSED "pose accuracy with GT bbox" number (not detector AP). Area = bbox w*h (this dataset carries bbox, not mask area -- documented approximation; official COCO uses segmentation area). Modes (build/run launcher forwards no argv -> config files, cwd=nxc2): data/mp_eval_gt.txt -- GT JSONL path (required) data/mp_eval_pred.txt -- predictions JSONL path (optional). Present+nonempty => REAL eval (pair by image name). Absent/empty => SELF-TEST on real GT line 1 (identity=1000 + shift). nx_safety_envelope: intended_use: offline scoring of pose predictions vs real COCO GT sil_target: SIL1 verdict: NOT_YET_EVALUATED genealogy_id: cocodataset/cocoeval(OKS) lineage_id: nx_coco_eval_v1