Document new yolo11m-seg checkpoint and inference_config.json

README.md

@@ -6,9 +6,28 @@ Public assets for the [sea turtle behavior pipeline](https://github.com/bara495/
 
 | Path | Description |
 |------|-------------|
-| `models/yolo/best.pt` | YOLOv8m-seg model - sea turtle detection |
+| `models/yolo/best.pt` | YOLO11m-seg model — sea turtle detection |
+| `models/yolo/inference_config.json` | Calibrated `conf`/`iou`/`max_det` this checkpoint needs — see below |
 | `models/breathing/best_model.keras` | ResNet50 classifier - breathing detection |
 | `models/breathing/optimal_threshold.json` | F1-optimal threshold for breathing classifier |
 | `1min_sample/` | 1-minute sample clips (2 cameras, 2 days) for pipeline testing |
 
 Downloaded automatically by `setup.sh` -- no manual action needed.
+
+## YOLO inference settings
+
+Always run this model with `max_det=1` — every tank holds exactly one turtle,
+so this is a domain constraint, not a tuning choice. Use `inference_config.json`'s
+`conf` (0.40) rather than a generic default: it's calibrated against the full
+validation set under `max_det=1`, the model's actual deployment shape, and
+holds P=R=F1=1.0 across a wide safe range (0.05-0.75) rather than sitting at
+either edge of it.
+
+2026-09-05: replaced the previous checkpoint, which scored well on paper
+(fitness in the top 10 of ~40 sweep runs) but had 89 false positives and 80%
+recall when actually run at real inference settings. Root cause was two
+independent bugs — a checkpoint-selection bug in training.py that could save
+the wrong epoch's weights as best.pt, and 2 of 50 validation images carrying
+duplicate ground-truth labels — both since fixed. This checkpoint is a full
+retrain under the corrected code and corrected labels, verified with a clean
+P=R=F1=1.0 sweep on the full val set.