diff --git a/README.md b/README.md
index c826906..d52f236 100644
--- a/README.md
+++ b/README.md
@@ -1,101 +1,98 @@
 # Saqr — PPE Safety Detection on Unitree G1
 
-Real-time PPE compliance (helmet, vest, boots, gloves, goggles) using YOLO11n,
-designed to run on a Unitree G1 humanoid with an Intel RealSense D435I. On
-UNSAFE the robot speaks a warning and plays the `reject` arm action.
+Real-time PPE compliance (helmet, vest, boots, gloves, goggles) using YOLO11n.
+Runs on a Unitree G1 humanoid with an Intel RealSense D435I. On UNSAFE the
+robot speaks a warning and plays the `reject` arm action.
 
 ## Layout
 
 ```
-saqr/           # python package
-  core/         # detection + tracking + events (shared by CLI/GUI/bridge)
-  apps/         # CLI entry points (saqr, detect, train, manager, view_stream)
-  gui/          # PySide6 desktop GUI
-  robot/        # G1 bridge + DDS controller
-  utils/        # logger
-scripts/        # deploy.sh, start_saqr.sh, run_local.sh, run_robot.sh, systemd unit
-config/         # logging.json
-data/           # dataset/, models/   (gitignored)
-runtime/        # captures/, logs/, runs/   (gitignored)
-docs/           # DEPLOY.md, start.md, use_case_catalogue.pdf
+Saqr/
+├── core/           # detection + tracking + events (shared by CLI/GUI/bridge)
+├── apps/           # CLI modules (saqr_cli, detect_cli, train_cli, manager_cli, view_stream)
+├── gui/            # PySide6 desktop GUI (dev-machine only)
+├── robot/          # G1 bridge + DDS controller
+├── utils/          # logger, config loader
+├── scripts/
+│   ├── start_saqr.sh         # the single entry point
+│   ├── saqr-bridge.service   # systemd unit (wraps start_saqr.sh)
+│   └── deploy.sh             # push code dev machine → robot
+├── config/         # logging.json, core_config.json, robot_config.json
+├── data/           # dataset/, models/          (gitignored)
+├── runtime/        # captures/, runs/           (gitignored)
+├── logs/           # per-module .log files      (gitignored)
+├── docs/           # DEPLOY.md, start.md, use_case_catalogue.pdf
+├── pyproject.toml
+└── README.md
 ```
 
-## Quick start
+The project root is auto-detected from `core/paths.py::PROJECT_ROOT` — drop
+the `Saqr/` folder anywhere on disk and the code finds itself. Override
+with `SAQR_ROOT=/custom/path` if needed.
+
+## Run
+
+The project only runs through **[scripts/start_saqr.sh](scripts/start_saqr.sh)**
+(directly or under the `saqr-bridge` systemd unit):
 
 ```bash
-# Install the package (editable)
-pip install -e .
-
-# Local dev run (webcam)
-saqr --source 0
-
-# PySide6 GUI
-pip install -e ".[gui]"
-saqr-gui
-
-# On the Unitree G1 (bridge owns the R2+X / R2+Y flow)
-saqr-bridge --iface eth0 --source realsense --headless -- --stream 8080
+# On the robot:
+sudo systemctl start saqr-bridge       # production
+# or
+~/Saqr/scripts/start_saqr.sh           # foreground / debug
 ```
 
-Without installing, everything still works via `python -m`:
+Then on the wireless remote:
+- **R2 + X** → start detection
+- **R2 + Y** → stop detection
 
+See [docs/DEPLOY.md](docs/DEPLOY.md) for first-time deploy and
+[docs/start.md](docs/start.md) for the systemd workflow.
+
+## Deploy
+
+From the dev machine:
 ```bash
-python -m saqr.apps.saqr_cli --source 0
-python -m saqr.robot.bridge --iface eth0 --source realsense --headless
+scripts/deploy.sh            # rsync + pip install -e . in the robot's conda env
+scripts/deploy.sh --ip …     # custom robot IP
 ```
 
-## Docs
+## Configure
 
-- [docs/DEPLOY.md](docs/DEPLOY.md) — full deploy + robot setup.
-- [docs/start.md](docs/start.md) — systemd auto-start workflow.
-- [docs/use_case_catalogue.pdf](docs/use_case_catalogue.pdf) — PPE use-case spec.
+All tunable values live in JSON — no code edits needed:
+
+- [config/core_config.json](config/core_config.json) — detection, tracking,
+  camera, stream, training.
+- [config/robot_config.json](config/robot_config.json) — bridge, TTS,
+  phrases, arm actions, deploy, start_saqr defaults.
+- [config/logging.json](config/logging.json) — log levels per category.
+
+Precedence: **env var > config JSON > code fallback**. Most defaults can be
+overridden via env vars without editing the files (`CONDA_ENV`,
+`SAQR_SOURCE`, `STREAM_PORT`, `DDS_IFACE`, `ROBOT_IP`, …).
+
+After editing any JSON:
+```bash
+sudo systemctl restart saqr-bridge
+```
 
 ## Data & Models
 
-The `data/` and `runtime/` directories are excluded from git (too large).
-Download them separately before training or running inference.
+`data/` and `runtime/` are gitignored (too large). Download separately:
 
-### `data/` — dataset and pre-trained weights
+- **Dataset**: [testcasque/ppe-detection-qlq3d](https://universe.roboflow.com/testcasque/ppe-detection-qlq3d)
+  → YOLOv11 format → unzip to `data/dataset/`.
+- **Base weights**: [Ultralytics releases](https://github.com/ultralytics/assets/releases)
+  → `yolo11n.pt` into `data/models/`.
+- **Saqr fine-tuned weights** (`saqr_best.pt`): produced by training — see
+  [docs/DEPLOY.md](docs/DEPLOY.md).
 
-Expected contents:
+## Training (dev machine, off-path)
 
-```
-data/
-  dataset/
-    train/{images,labels}/
-    valid/{images,labels}/
-    test/{images,labels}/
-    data.yaml
-  models/
-    saqr_best.pt       # Saqr YOLO11n fine-tuned on PPE
-    saqr_last.pt
-    yolo11n.pt         # base YOLO11n
-    yolo26n.pt         # base YOLO26n
-```
-
-Download:
-
-- **Dataset** (PPE, Roboflow): [testcasque/ppe-detection-qlq3d](https://universe.roboflow.com/testcasque/ppe-detection-qlq3d)
-  Open the Roboflow link → *Download Dataset* → format **YOLOv11** → unzip into `data/dataset/`.
-- **Base YOLO weights**: [Ultralytics assets releases](https://github.com/ultralytics/assets/releases)
-  Grab `yolo11n.pt` (and optionally `yolo26n.pt`) into `data/models/`.
-- **Saqr fine-tuned weights** (`saqr_best.pt`, `saqr_last.pt`):
-  Produced by training — see "Training" below. Or request from the maintainer.
-
-Place everything under `data/` so the tree matches above.
-
-### `runtime/` — training output (optional)
-
-Auto-generated when you run training. Not required for inference.
-Contains confusion matrices, PR curves, batch previews, and the raw weights
-under `runtime/runs/train/saqr_det/weights/`.
-
-### Training
+Training is done on a workstation, not on the robot. It's a one-off, not
+part of the normal run flow:
 
 ```bash
-# after placing the dataset in data/dataset/ and base weights in data/models/
-python -m saqr.apps.train_cli --data data/dataset/data.yaml --weights data/models/yolo11n.pt
+python -m apps.train_cli --epochs 100 --batch 16
+# best weights land at data/models/saqr_best.pt; deploy with scripts/deploy.sh
 ```
-
-Outputs land in `runtime/runs/train/saqr_det/`. Copy the best checkpoint to
-`data/models/saqr_best.pt` to use it at inference time.
diff --git a/saqr/apps/__init__.py b/apps/__init__.py
similarity index 100%
rename from saqr/apps/__init__.py
rename to apps/__init__.py
diff --git a/saqr/apps/detect_cli.py b/apps/detect_cli.py
similarity index 83%
rename from saqr/apps/detect_cli.py
rename to apps/detect_cli.py
index ffe0340..0bbb824 100644
--- a/saqr/apps/detect_cli.py
+++ b/apps/detect_cli.py
@@ -8,11 +8,14 @@ from pathlib import Path
 import cv2
 from ultralytics import YOLO
 
-from saqr.core.detection import get_inference_config, set_inference_config
-from saqr.core.model import resolve_model_path
-from saqr.utils.logger import get_logger
+from core.detection import get_inference_config, set_inference_config
+from core.model import resolve_model_path
+from utils.config import load_config
+from utils.logger import get_logger
 
 log = get_logger("Inference", "detect")
+_DET = load_config("core")["detection"]
+_CAM = load_config("core")["camera"]
 
 VIOLATION = {"no-helmet", "no-vest", "no-boots", "no-gloves", "no-goggles"}
 COMPLIANT = {"helmet", "vest", "boots", "gloves", "goggles"}
@@ -97,12 +100,12 @@ def run_image(model, path, conf):
 
 def main():
     parser = argparse.ArgumentParser(description="Saqr simple PPE detection")
-    parser.add_argument("--source", default="0")
-    parser.add_argument("--model",  default="saqr_best.pt")
-    parser.add_argument("--conf",   type=float, default=0.35)
-    parser.add_argument("--device", default="0", help="'cpu', '0', 'cuda:0'")
-    parser.add_argument("--half",   action="store_true")
-    parser.add_argument("--imgsz",  type=int, default=320)
+    parser.add_argument("--source", default=_CAM["default_source"])
+    parser.add_argument("--model",  default=_DET["default_model"])
+    parser.add_argument("--conf",   type=float, default=_DET["conf"])
+    parser.add_argument("--device", default=_DET["device"], help="'cpu', '0', 'cuda:0'")
+    parser.add_argument("--half",   action="store_true", default=_DET["half"])
+    parser.add_argument("--imgsz",  type=int, default=_DET["imgsz"])
     args = parser.parse_args()
 
     set_inference_config(device=args.device, half=args.half, imgsz=args.imgsz)
diff --git a/saqr/apps/manager_cli.py b/apps/manager_cli.py
similarity index 99%
rename from saqr/apps/manager_cli.py
rename to apps/manager_cli.py
index 170168c..770625e 100644
--- a/saqr/apps/manager_cli.py
+++ b/apps/manager_cli.py
@@ -10,8 +10,8 @@ from pathlib import Path
 
 import cv2
 
-from saqr.core.paths import CAPTURES_DIR, PROJECT_ROOT
-from saqr.utils.logger import get_logger
+from core.paths import CAPTURES_DIR, PROJECT_ROOT
+from utils.logger import get_logger
 
 log = get_logger("Manager", "manager")
 
diff --git a/saqr/apps/saqr_cli.py b/apps/saqr_cli.py
similarity index 54%
rename from saqr/apps/saqr_cli.py
rename to apps/saqr_cli.py
index 2b3307d..14321fe 100644
--- a/saqr/apps/saqr_cli.py
+++ b/apps/saqr_cli.py
@@ -9,23 +9,28 @@ from typing import Dict
 import cv2
 from ultralytics import YOLO
 
-from saqr.core.camera import open_capture
-from saqr.core.capture import setup_capture_dirs
-from saqr.core.detection import set_inference_config
-from saqr.core.drawing import draw_counters
-from saqr.core.events import EventLogger, write_result_csv
-from saqr.core.model import resolve_model_path
-from saqr.core.paths import EVENTS_CSV, RESULT_CSV
-from saqr.core.pipeline import process_frame
-from saqr.core.streaming import start_stream_server, update_stream_frame
-from saqr.core.tracking import PersonTracker
-from saqr.utils.logger import get_logger
+from core.camera import RealSenseCapture, open_capture
+from core.capture import setup_capture_dirs, setup_snapshot_dirs
+from core.detection import set_inference_config
+from core.drawing import draw_counters
+from core.events import EventLogger, write_result_csv
+from core.model import resolve_model_path
+from core.paths import EVENTS_CSV, RESULT_CSV
+from core.pipeline import process_frame
+from core.streaming import start_stream_server, update_stream_frame
+from core.tracking import PersonTracker
+from utils.config import load_config
+from utils.logger import get_logger
 
 log = get_logger("Inference", "saqr")
+_CORE = load_config("core")
 
 
-def run_video(model, source, conf, capture_dirs: Dict[str, Path], show_gui, csv_every_frame,
-              max_missing, match_distance, status_confirm_frames, stream_port=0):
+def run_video(model, source, conf, capture_dirs: Dict[str, Path], show_gui,
+              csv_interval, max_missing, match_distance, status_confirm_frames,
+              *,
+              snapshot_dirs=None, max_distance_m: float = 0.0,
+              stream_port: int = 0):
     cap = open_capture(source)
     if not cap.isOpened():
         log.error(f"Cannot open source: {source}")
@@ -37,6 +42,9 @@ def run_video(model, source, conf, capture_dirs: Dict[str, Path], show_gui, csv_
         cap.release()
         return
 
+    is_realsense = isinstance(cap, RealSenseCapture) and cap.has_depth
+    depth_scale = cap.depth_scale if is_realsense else 0.001
+
     event_logger = EventLogger(EVENTS_CSV)
     tracker = PersonTracker(
         event_logger=event_logger,
@@ -48,7 +56,11 @@ def run_video(model, source, conf, capture_dirs: Dict[str, Path], show_gui, csv_
     if stream_port > 0:
         start_stream_server(stream_port)
 
-    log.info(f"Session started | source={source}")
+    log.info(
+        f"Session started | source={source} depth={is_realsense} "
+        f"max_distance_m={max_distance_m if max_distance_m > 0 else 'off'} "
+        f"csv_interval={csv_interval}"
+    )
     if show_gui:
         print("Running - press q to quit, s to save frame.")
 
@@ -58,10 +70,16 @@ def run_video(model, source, conf, capture_dirs: Dict[str, Path], show_gui, csv_
 
     while True:
         frame_idx += 1
+        depth_frame = cap.latest_depth if is_realsense else None
+        write_csv_this_frame = csv_interval > 0 and (frame_idx % csv_interval == 0)
+
         try:
             annotated, visible = process_frame(
                 frame, model, tracker, frame_idx, conf,
-                capture_dirs, write_csv=csv_every_frame,
+                capture_dirs, write_csv=write_csv_this_frame,
+                snapshot_dirs=snapshot_dirs,
+                depth_frame=depth_frame, depth_scale=depth_scale,
+                max_distance_m=max_distance_m,
             )
         except Exception as e:
             log.exception(f"Frame error #{frame_idx}: {e}")
@@ -94,11 +112,12 @@ def run_video(model, source, conf, capture_dirs: Dict[str, Path], show_gui, csv_
     if show_gui:
         cv2.destroyAllWindows()
 
+    # Always write final state on exit so the last tracked people are recorded.
     write_result_csv(list(tracker.tracks.values()), RESULT_CSV)
     log.info(f"Session ended | frames={frame_idx} tracks_created={tracker.next_id - 1}")
 
 
-def run_image(model, path, conf, capture_dirs: Dict[str, Path], show_gui):
+def run_image(model, path, conf, capture_dirs: Dict[str, Path], show_gui, snapshot_dirs=None):
     frame = cv2.imread(path)
     if frame is None:
         log.error(f"Cannot read image: {path}")
@@ -107,7 +126,10 @@ def run_image(model, path, conf, capture_dirs: Dict[str, Path], show_gui):
     event_logger = EventLogger(EVENTS_CSV)
     tracker = PersonTracker(event_logger=event_logger)
 
-    annotated, visible = process_frame(frame, model, tracker, 1, conf, capture_dirs)
+    annotated, visible = process_frame(
+        frame, model, tracker, 1, conf, capture_dirs,
+        snapshot_dirs=snapshot_dirs,
+    )
     draw_counters(annotated, visible, 0.0)
 
     out = Path(path).stem + "_saqr.jpg"
@@ -122,20 +144,32 @@ def run_image(model, path, conf, capture_dirs: Dict[str, Path], show_gui):
 
 def main():
     parser = argparse.ArgumentParser(description="Saqr PPE detection with tracking")
-    parser.add_argument("--source",  default="0",
+    det  = _CORE["detection"]
+    trk  = _CORE["tracking"]
+    cam  = _CORE["camera"]
+    cap_cfg = _CORE["capture"]
+
+    parser.add_argument("--source", default=cam["default_source"],
                         help="0/1=webcam, realsense, realsense:SERIAL, /dev/videoX, or video path")
-    parser.add_argument("--model",   default="saqr_best.pt",
+    parser.add_argument("--model",  default=det["default_model"],
                         help="Trained YOLO weights (resolved under data/models/ by default)")
-    parser.add_argument("--conf",    type=float, default=0.35)
-    parser.add_argument("--max-missing", type=int, default=90)
-    parser.add_argument("--match-distance", type=float, default=250.0)
-    parser.add_argument("--status-confirm-frames", type=int, default=5)
-    parser.add_argument("--headless", action="store_true")
-    parser.add_argument("--stream",   type=int, default=0, metavar="PORT")
-    parser.add_argument("--csv-on-exit", action="store_true")
-    parser.add_argument("--device", default="0")
-    parser.add_argument("--half",   action="store_true")
-    parser.add_argument("--imgsz",  type=int, default=320)
+    parser.add_argument("--conf",                  type=float, default=det["conf"])
+    parser.add_argument("--max-missing",           type=int,   default=trk["max_missing"])
+    parser.add_argument("--match-distance",        type=float, default=trk["match_distance"])
+    parser.add_argument("--status-confirm-frames", type=int,   default=trk["status_confirm_frames"])
+    parser.add_argument("--max-distance-m",        type=float, default=det.get("max_distance_m", 0.0),
+                        help="RealSense-only: drop candidates beyond this depth (0 = off)")
+    parser.add_argument("--headless",              action="store_true")
+    parser.add_argument("--stream",                type=int,   default=0, metavar="PORT")
+    parser.add_argument("--csv-interval",          type=int,   default=trk.get("csv_write_every_n_frames", 30),
+                        help="Write result.csv every N frames (0 = only on exit)")
+    parser.add_argument("--csv-on-exit",           action="store_true",
+                        help="Alias for --csv-interval 0")
+    parser.add_argument("--no-snapshots",          action="store_true",
+                        help="Disable full-frame snapshot on transitions")
+    parser.add_argument("--device",                default=det["device"])
+    parser.add_argument("--half",                  action="store_true", default=det["half"])
+    parser.add_argument("--imgsz",                 type=int,   default=det["imgsz"])
     args = parser.parse_args()
 
     set_inference_config(device=args.device, half=args.half, imgsz=args.imgsz)
@@ -154,12 +188,17 @@ def main():
         log.warning("PyTorch not found")
 
     capture_dirs = setup_capture_dirs()
+    snapshot_dirs = None
+    if cap_cfg.get("save_event_snapshot", True) and not args.no_snapshots:
+        snapshot_dirs = setup_snapshot_dirs()
+
+    csv_interval = 0 if args.csv_on_exit else max(0, args.csv_interval)
 
     try:
         model_path = resolve_model_path(args.model)
     except FileNotFoundError as e:
         log.error(str(e))
-        log.error("Train first: saqr-train --dataset data/dataset")
+        log.error("Train first: python -m apps.train_cli --dataset data/dataset")
         raise SystemExit(1)
 
     log.info(f"Loading model: {model_path}")
@@ -176,14 +215,17 @@ def main():
         run_video(
             model, source, args.conf, capture_dirs,
             show_gui=not args.headless,
-            csv_every_frame=not args.csv_on_exit,
+            csv_interval=csv_interval,
             max_missing=args.max_missing,
             match_distance=args.match_distance,
             status_confirm_frames=args.status_confirm_frames,
+            snapshot_dirs=snapshot_dirs,
+            max_distance_m=args.max_distance_m,
             stream_port=args.stream,
         )
     elif Path(source).exists():
-        run_image(model, source, args.conf, capture_dirs, show_gui=not args.headless)
+        run_image(model, source, args.conf, capture_dirs,
+                  show_gui=not args.headless, snapshot_dirs=snapshot_dirs)
     else:
         log.error(f"Source not found: {source}")
         raise SystemExit(1)
diff --git a/saqr/apps/train_cli.py b/apps/train_cli.py
similarity index 84%
rename from saqr/apps/train_cli.py
rename to apps/train_cli.py
index 55fb294..5369610 100644
--- a/saqr/apps/train_cli.py
+++ b/apps/train_cli.py
@@ -7,10 +7,12 @@ from pathlib import Path
 
 import yaml
 
-from saqr.core.paths import DATASET_DIR, MODELS_DIR, PROJECT_ROOT, RUNS_DIR
-from saqr.utils.logger import get_logger
+from core.paths import DATASET_DIR, MODELS_DIR, PROJECT_ROOT, RUNS_DIR
+from utils.config import load_config
+from utils.logger import get_logger
 
 log = get_logger("Training", "train")
+_TRAIN = load_config("core")["training"]
 
 EXPECTED_CLASSES = [
     "boots", "gloves", "goggles", "helmet", "no-boots",
@@ -52,13 +54,13 @@ def main():
     parser = argparse.ArgumentParser(description="Train Saqr PPE detector (YOLO11n)")
     parser.add_argument("--dataset", default=str(DATASET_DIR),
                         help="Root folder containing data.yaml + train/valid/test")
-    parser.add_argument("--epochs",  type=int, default=100)
-    parser.add_argument("--imgsz",   type=int, default=640)
-    parser.add_argument("--batch",   type=int, default=16)
-    parser.add_argument("--model",   default="yolo11n.pt",
+    parser.add_argument("--epochs",  type=int, default=_TRAIN["epochs"])
+    parser.add_argument("--imgsz",   type=int, default=_TRAIN["imgsz"])
+    parser.add_argument("--batch",   type=int, default=_TRAIN["batch"])
+    parser.add_argument("--model",   default=_TRAIN["base_model"],
                         help="Base YOLO model (auto-downloaded if not present)")
-    parser.add_argument("--name",    default="saqr_det")
-    parser.add_argument("--device",  default="0")
+    parser.add_argument("--name",    default=_TRAIN["run_name"])
+    parser.add_argument("--device",  default=_TRAIN["device"])
     args = parser.parse_args()
 
     dataset_root = Path(args.dataset)
diff --git a/saqr/apps/view_stream.py b/apps/view_stream.py
similarity index 69%
rename from saqr/apps/view_stream.py
rename to apps/view_stream.py
index 2a31d35..4d1c89c 100644
--- a/saqr/apps/view_stream.py
+++ b/apps/view_stream.py
@@ -1,15 +1,20 @@
-"""View the robot's MJPEG stream on a laptop."""
+"""View the robot's MJPEG stream from a laptop."""
 from __future__ import annotations
 
 import argparse
 
 import cv2
 
+from utils.config import load_config
+
 
 def main():
+    default_ip   = load_config("robot")["deploy"]["robot_ip"]
+    default_port = load_config("core")["stream"]["port"]
+
     parser = argparse.ArgumentParser(description="View Saqr PPE stream from robot")
-    parser.add_argument("--ip",   default="192.168.123.164", help="Robot IP address")
-    parser.add_argument("--port", default="8080",            help="Stream port")
+    parser.add_argument("--ip",   default=default_ip,         help="Robot IP address")
+    parser.add_argument("--port", default=str(default_port),  help="Stream port")
     args = parser.parse_args()
 
     url = f"http://{args.ip}:{args.port}/stream"
diff --git a/assets/audio/README.md b/assets/audio/README.md
new file mode 100644
index 0000000..4d41144
--- /dev/null
+++ b/assets/audio/README.md
@@ -0,0 +1,88 @@
+# Pre-recorded audio library
+
+WAV clips played via `AudioClient.PlayStream` on the G1 speaker. Bypassing
+`TtsMaker` saves ~200–700 ms of firmware synthesis buffer per announcement
+and eliminates 3104 "device busy" errors.
+
+## Required format
+
+Every file **must** be:
+
+- **16 kHz** sample rate
+- **mono** (1 channel)
+- **16-bit signed PCM** (`int16`)
+- `.wav` container
+
+Any file not matching this is logged as a warning and skipped (the bridge
+falls back to `TtsMaker` for that phrase per `config/robot_config.json.tts.mode`).
+
+## Expected files
+
+Drop WAVs under the right `<category>/<key>.wav` path so `AudioPlayer` finds them:
+
+```
+fixed/
+  ready.wav             "Saqr is running. Press R2 plus X to start."
+  deactivated.wav       "Saqr deactivated."
+  no_camera.wav         "Camera not connected. Please plug in the camera and try again."
+  safe.wav              "Safe to enter. Have a good day."
+  unsafe_generic.wav    "Please stop. Wear your proper safety equipment."
+
+unsafe_missing/
+  helmet.wav            "Please stop. Wear your proper safety equipment. You are missing helmet."
+  vest.wav              "Please stop. Wear your proper safety equipment. You are missing vest."
+  helmet_vest.wav       "Please stop. Wear your proper safety equipment. You are missing helmet and vest."
+```
+
+**Naming rule for `unsafe_missing/`**: the `<key>` is the missing PPE names
+sorted alphabetically and joined with `_`. So if someone misses both helmet
+and vest, the bridge looks up `unsafe_missing/helmet_vest.wav` (not
+`vest_helmet.wav`). If you extend `compliance.required_ppe` later, add clips
+for every subset — for 3 required items that's 7 combinations (2³−1).
+
+## Converting existing recordings
+
+If your source file is at a different sample rate or stereo, convert with
+`ffmpeg`:
+
+```bash
+ffmpeg -y -i input.m4a -ac 1 -ar 16000 -sample_fmt s16 fixed/safe.wav
+```
+
+Validate with:
+
+```bash
+python3 - <<'EOF'
+import wave, sys
+with wave.open("fixed/safe.wav", "rb") as wf:
+    print(wf.getframerate(), "Hz,", wf.getnchannels(), "ch,", wf.getsampwidth()*8, "bit")
+EOF
+# must print: 16000 Hz, 1 ch, 16 bit
+```
+
+## Recording tips
+
+- Quiet room; no echo.
+- Don't clip — keep peaks below 0 dBFS.
+- Leave ~100 ms of silence at the start and end so the clip doesn't pop.
+- Target speaking rate: ~3 syllables/sec. The shortest clip (`deactivated`)
+  should be ~2 s; longest (`no_camera`) around 5–6 s.
+
+## Runtime behaviour
+
+In `config/robot_config.json`:
+
+```json
+"tts": {
+    "mode": "recorded_or_tts"
+}
+```
+
+Modes:
+- `recorded_or_tts` — play WAV if the clip exists, otherwise fall back to `TtsMaker`.
+- `recorded_only` — play WAV or stay silent. Useful for demos where you want
+  deterministic audio. Will skip any phrase whose clip is missing.
+- `tts_only` — ignore the WAV library entirely (current legacy behaviour).
+
+After adding or replacing WAVs, restart the bridge to reload the library
+(`sudo systemctl restart saqr-bridge`).
diff --git a/saqr/core/__init__.py b/assets/audio/fixed/.gitkeep
similarity index 100%
rename from saqr/core/__init__.py
rename to assets/audio/fixed/.gitkeep
diff --git a/assets/audio/fixed/deactivated.wav b/assets/audio/fixed/deactivated.wav
new file mode 100644
index 0000000..78ed258
Binary files /dev/null and b/assets/audio/fixed/deactivated.wav differ
diff --git a/assets/audio/fixed/no_camera.wav b/assets/audio/fixed/no_camera.wav
new file mode 100644
index 0000000..9382ad8
Binary files /dev/null and b/assets/audio/fixed/no_camera.wav differ
diff --git a/assets/audio/fixed/ready.wav b/assets/audio/fixed/ready.wav
new file mode 100644
index 0000000..52be2c0
Binary files /dev/null and b/assets/audio/fixed/ready.wav differ
diff --git a/assets/audio/fixed/safe.wav b/assets/audio/fixed/safe.wav
new file mode 100644
index 0000000..16b02ee
Binary files /dev/null and b/assets/audio/fixed/safe.wav differ
diff --git a/assets/audio/fixed/unsafe_generic.wav b/assets/audio/fixed/unsafe_generic.wav
new file mode 100644
index 0000000..f4d5548
Binary files /dev/null and b/assets/audio/fixed/unsafe_generic.wav differ
diff --git a/saqr/gui/__init__.py b/assets/audio/unsafe_missing/.gitkeep
similarity index 100%
rename from saqr/gui/__init__.py
rename to assets/audio/unsafe_missing/.gitkeep
diff --git a/assets/audio/unsafe_missing/helmet.wav b/assets/audio/unsafe_missing/helmet.wav
new file mode 100644
index 0000000..ebc439c
Binary files /dev/null and b/assets/audio/unsafe_missing/helmet.wav differ
diff --git a/assets/audio/unsafe_missing/helmet_vest.wav b/assets/audio/unsafe_missing/helmet_vest.wav
new file mode 100644
index 0000000..63e082b
Binary files /dev/null and b/assets/audio/unsafe_missing/helmet_vest.wav differ
diff --git a/assets/audio/unsafe_missing/vest.wav b/assets/audio/unsafe_missing/vest.wav
new file mode 100644
index 0000000..abf4cd9
Binary files /dev/null and b/assets/audio/unsafe_missing/vest.wav differ
diff --git a/config/core_config.json b/config/core_config.json
new file mode 100644
index 0000000..c049a36
--- /dev/null
+++ b/config/core_config.json
@@ -0,0 +1,44 @@
+{
+    "detection": {
+        "conf":           0.35,
+        "device":         "0",
+        "half":           false,
+        "imgsz":          320,
+        "default_model":  "saqr_best.pt",
+        "max_distance_m": 3.0
+    },
+    "tracking": {
+        "max_missing":              90,
+        "match_distance":           250.0,
+        "status_confirm_frames":    5,
+        "csv_write_every_n_frames": 30,
+        "stationary_check_frames":  15,
+        "stationary_tolerance_px":  40.0
+    },
+    "compliance": {
+        "required_ppe": ["helmet", "vest"]
+    },
+    "capture": {
+        "save_event_snapshot": true
+    },
+    "camera": {
+        "width":          640,
+        "height":         480,
+        "fps":            30,
+        "default_source": "0",
+        "enable_depth":   true
+    },
+    "stream": {
+        "host":         "0.0.0.0",
+        "port":         8080,
+        "jpeg_quality": 70
+    },
+    "training": {
+        "epochs":     100,
+        "imgsz":      640,
+        "batch":      16,
+        "device":     "0",
+        "base_model": "yolo11n.pt",
+        "run_name":   "saqr_det"
+    }
+}
diff --git a/config/robot_config.json b/config/robot_config.json
new file mode 100644
index 0000000..3ad700e
--- /dev/null
+++ b/config/robot_config.json
@@ -0,0 +1,51 @@
+{
+    "bridge": {
+        "iface":             "eth0",
+        "timeout":           10.0,
+        "cooldown":          8.0,
+        "release_after":     0.5,
+        "audio_lead_s":      0.3,
+        "quick_fail_window": 8.0,
+        "trigger_poll_hz":   50.0
+    },
+    "tts": {
+        "mode":             "recorded_only",
+        "speaker_id":       2,
+        "volume":           100,
+        "seconds_per_char": 0.12,
+        "min_seconds":      2.5,
+        "queue_max":        4,
+        "busy_factor": {
+            "min":  1.0,
+            "max":  2.5,
+            "up":   1.20,
+            "down": 0.97
+        }
+    },
+    "phrases": {
+        "safe":           "Safe to enter. Have a good day.",
+        "unsafe_generic": "Please stop. Wear your proper safety equipment.",
+        "unsafe_missing": "Please stop. Wear your proper safety equipment. You are missing {items}.",
+        "deactivated":    "Saqr deactivated.",
+        "ready":          "Saqr is running. Press R2 plus X to start.",
+        "no_camera":      "Camera not connected. Please plug in the camera and try again."
+    },
+    "arm": {
+        "reject_action":  "reject",
+        "release_action": "release arm"
+    },
+    "deploy": {
+        "robot_ip":   "192.168.123.164",
+        "robot_user": "unitree",
+        "robot_env":  "marcus",
+        "remote_dir": "/home/unitree/Saqr",
+        "ssh_opts":   "-o StrictHostKeyChecking=no -o ConnectTimeout=10"
+    },
+    "start_saqr": {
+        "conda_root":  "$HOME/miniconda3",
+        "conda_env":   "marcus",
+        "dds_iface":   "eth0",
+        "saqr_source": "realsense",
+        "stream_port": 8080
+    }
+}
diff --git a/saqr/robot/__init__.py b/core/__init__.py
similarity index 100%
rename from saqr/robot/__init__.py
rename to core/__init__.py
diff --git a/core/camera.py b/core/camera.py
new file mode 100644
index 0000000..ebe3b23
--- /dev/null
+++ b/core/camera.py
@@ -0,0 +1,157 @@
+"""Camera backends: RealSense SDK (with optional aligned depth) and OpenCV V4L2.
+
+RealSenseCapture exposes the latest depth frame via ``latest_depth`` (a
+numpy uint16 array in millimetres) and ``depth_scale`` (meters per raw unit).
+Non-RealSense captures leave ``latest_depth = None`` — callers must guard.
+"""
+from __future__ import annotations
+
+from typing import Optional
+
+import cv2
+import numpy as np
+
+from utils.config import load_config
+from utils.logger import get_logger
+
+log = get_logger("Inference", "camera")
+_CFG = load_config("core")["camera"]
+
+try:
+    import pyrealsense2 as rs
+    HAS_REALSENSE = True
+except ImportError:
+    HAS_REALSENSE = False
+
+
+class RealSenseCapture:
+    """pyrealsense2 pipeline with an OpenCV-like read() interface.
+
+    When ``enable_depth`` is true (default from core_config.camera.enable_depth),
+    the pipeline also streams depth and aligns each frame to the color view.
+    The aligned depth frame is stored on ``self.latest_depth`` as a numpy array.
+    """
+
+    def __init__(self, width: int = _CFG["width"], height: int = _CFG["height"],
+                 fps: int = _CFG["fps"], serial: Optional[str] = None,
+                 enable_depth: bool = _CFG.get("enable_depth", True)):
+        if not HAS_REALSENSE:
+            raise RuntimeError("pyrealsense2 not installed")
+        self.pipeline = rs.pipeline()
+        cfg = rs.config()
+        if serial:
+            cfg.enable_device(serial)
+        cfg.enable_stream(rs.stream.color, width, height, rs.format.bgr8, fps)
+
+        self.has_depth = bool(enable_depth)
+        self._align = None
+        self.depth_scale = 0.001  # mm per raw unit default; overwritten below
+
+        if self.has_depth:
+            cfg.enable_stream(rs.stream.depth, width, height, rs.format.z16, fps)
+
+        self.profile = self.pipeline.start(cfg)
+        self._open = True
+
+        if self.has_depth:
+            try:
+                depth_sensor = self.profile.get_device().first_depth_sensor()
+                self.depth_scale = float(depth_sensor.get_depth_scale())
+                self._align = rs.align(rs.stream.color)
+            except Exception as e:
+                log.warning(f"Depth init failed ({e}); disabling depth")
+                self.has_depth = False
+                self._align = None
+
+        self.latest_depth: Optional[np.ndarray] = None
+
+        dev = self.profile.get_device()
+        log.info(
+            f"RealSense opened | {dev.get_info(rs.camera_info.name)} "
+            f"serial={dev.get_info(rs.camera_info.serial_number)} "
+            f"{width}x{height}@{fps} depth={self.has_depth}"
+        )
+
+    def isOpened(self) -> bool:
+        return self._open
+
+    def read(self):
+        if not self._open:
+            return False, None
+        try:
+            frames = self.pipeline.wait_for_frames(timeout_ms=3000)
+            if self._align is not None:
+                frames = self._align.process(frames)
+            color = frames.get_color_frame()
+            if not color:
+                return False, None
+            if self.has_depth:
+                depth = frames.get_depth_frame()
+                self.latest_depth = (
+                    np.asanyarray(depth.get_data()) if depth else None
+                )
+            return True, np.asanyarray(color.get_data())
+        except Exception:
+            return False, None
+
+    def release(self):
+        if self._open:
+            self.pipeline.stop()
+            self._open = False
+
+
+def open_capture(source: str):
+    if source.lower().startswith("realsense"):
+        serial = None
+        if ":" in source:
+            serial = source.split(":", 1)[1]
+        return RealSenseCapture(
+            width=_CFG["width"], height=_CFG["height"], fps=_CFG["fps"],
+            serial=serial,
+        )
+
+    if str(source).isdigit():
+        idx = int(source)
+        cap = cv2.VideoCapture(idx)
+        if cap.isOpened():
+            return cap
+        cap = cv2.VideoCapture(idx, cv2.CAP_ANY)
+        if cap.isOpened():
+            return cap
+        cap = cv2.VideoCapture(idx, cv2.CAP_V4L2)
+        return cap
+
+    if source.startswith("/dev/video"):
+        cap = cv2.VideoCapture(source, cv2.CAP_V4L2)
+        if cap.isOpened():
+            cap.set(cv2.CAP_PROP_BUFFERSIZE, 1)
+            cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc(*"MJPG"))
+            cap.set(cv2.CAP_PROP_FRAME_WIDTH, _CFG["width"])
+            cap.set(cv2.CAP_PROP_FRAME_HEIGHT, _CFG["height"])
+            cap.set(cv2.CAP_PROP_FPS, _CFG["fps"])
+        return cap
+
+    return cv2.VideoCapture(source)
+
+
+def estimate_person_distance_m(depth_frame, bbox, depth_scale: float) -> Optional[float]:
+    """Median distance in metres inside ``bbox`` on the depth frame.
+
+    Returns ``None`` if depth is unavailable or the bbox falls outside the
+    valid area. Ignores zero-depth pixels (RealSense's "no return" marker).
+    """
+    if depth_frame is None:
+        return None
+    h, w = depth_frame.shape[:2]
+    x1, y1, x2, y2 = bbox
+    x1 = max(0, min(int(x1), w - 1))
+    x2 = max(0, min(int(x2), w))
+    y1 = max(0, min(int(y1), h - 1))
+    y2 = max(0, min(int(y2), h))
+    if x2 <= x1 or y2 <= y1:
+        return None
+    roi = depth_frame[y1:y2, x1:x2]
+    valid = roi[roi > 0]
+    if valid.size == 0:
+        return None
+    return float(np.median(valid) * depth_scale)
diff --git a/core/capture.py b/core/capture.py
new file mode 100644
index 0000000..35559e3
--- /dev/null
+++ b/core/capture.py
@@ -0,0 +1,67 @@
+"""Per-track image cropping, capture dirs, and full-frame event snapshots."""
+from __future__ import annotations
+
+from datetime import datetime
+from pathlib import Path
+from typing import Dict, Optional
+
+import cv2
+
+from core.detection import STATUSES
+from core.geometry import clamp_bbox
+from core.paths import CAPTURES_DIR, SNAPSHOTS_DIR
+
+
+def setup_capture_dirs() -> Dict[str, Path]:
+    """runtime/captures/{SAFE,PARTIAL,UNSAFE}/ — latest crop per track."""
+    dirs: Dict[str, Path] = {}
+    for s in STATUSES:
+        d = CAPTURES_DIR / s
+        d.mkdir(parents=True, exist_ok=True)
+        dirs[s] = d
+    return dirs
+
+
+def setup_snapshot_dirs() -> Dict[str, Path]:
+    """runtime/snapshots/{SAFE,PARTIAL,UNSAFE}/ — annotated full frame per transition."""
+    dirs: Dict[str, Path] = {}
+    for s in STATUSES:
+        d = SNAPSHOTS_DIR / s
+        d.mkdir(parents=True, exist_ok=True)
+        dirs[s] = d
+    return dirs
+
+
+def save_track_image(frame, track, capture_dirs: Dict[str, Path]) -> Optional[Path]:
+    """Save the latest crop for a track. Overwritten each frame."""
+    h, w = frame.shape[:2]
+    x1, y1, x2, y2 = clamp_bbox(track.bbox, w, h)
+    if x2 <= x1 or y2 <= y1:
+        return None
+    crop = frame[y1:y2, x1:x2]
+    if crop.size == 0:
+        return None
+
+    target = capture_dirs[track.status] / f"track_{track.track_id:04d}.jpg"
+    if track.photo_path and track.photo_path != target and track.photo_path.exists():
+        try:
+            track.photo_path.unlink()
+        except OSError:
+            pass
+
+    cv2.imwrite(str(target), crop)
+    track.photo_path = target
+    return target
+
+
+def save_event_snapshot(annotated_frame, track, snapshot_dirs: Dict[str, Path]) -> Optional[Path]:
+    """Save the full annotated frame at the moment of a NEW / STATUS_CHANGE event.
+
+    Timestamped filename so a history is preserved across events.
+    """
+    if annotated_frame is None or annotated_frame.size == 0:
+        return None
+    ts = datetime.now().strftime("%Y%m%d_%H%M%S_%f")[:-3]
+    target = snapshot_dirs[track.status] / f"track_{track.track_id:04d}_{ts}.jpg"
+    cv2.imwrite(str(target), annotated_frame)
+    return target
diff --git a/core/compliance.py b/core/compliance.py
new file mode 100644
index 0000000..1784ad3
--- /dev/null
+++ b/core/compliance.py
@@ -0,0 +1,55 @@
+"""Binary SAFE / UNSAFE classification — every item in REQUIRED_PPE must be worn.
+
+The required list comes from ``config/core_config.json::compliance.required_ppe``
+so it can be retuned without a code change.
+"""
+from __future__ import annotations
+
+from typing import Dict, List, Tuple
+
+from core.detection import POSITIVE_TO_NEGATIVE, PPE_DISPLAY_ORDER
+from utils.config import load_config
+
+# PPE items required for SAFE status. Other items (gloves/goggles/boots) are
+# still detected and logged but don't affect compliance.
+REQUIRED_PPE: Tuple[str, ...] = tuple(
+    load_config("core").get("compliance", {}).get("required_ppe", ["helmet", "vest"])
+)
+
+
+def _is_wearing(items: Dict[str, float], ppe: str) -> bool:
+    pos = items.get(ppe, 0.0)
+    neg = items.get(POSITIVE_TO_NEGATIVE[ppe], 0.0)
+    return pos > neg and pos > 0
+
+
+def status_from_items(items: Dict[str, float]) -> str:
+    """SAFE only when every required item is confidently worn."""
+    if all(_is_wearing(items, p) for p in REQUIRED_PPE):
+        return "SAFE"
+    return "UNSAFE"
+
+
+def split_wearing_missing(items: Dict[str, float]) -> Tuple[List[str], List[str], List[str]]:
+    """Bucket each PPE item into wearing / missing / unknown.
+
+    ``missing`` only contains REQUIRED items that aren't confidently worn —
+    this is what the TTS announcement keys off of, so it maps cleanly to a
+    recorded clip. Non-required items (gloves/goggles/boots) are never in
+    ``missing`` regardless of whether their ``no-X`` class was detected; they
+    go to ``unknown`` so they're still visible in the event log.
+    """
+    wearing, missing, unknown = [], [], []
+    for pos in PPE_DISPLAY_ORDER:
+        neg = POSITIVE_TO_NEGATIVE[pos]
+        pos_conf = items.get(pos, 0.0)
+        neg_conf = items.get(neg, 0.0)
+
+        worn = pos_conf > neg_conf and pos_conf > 0
+        if worn:
+            wearing.append(pos)
+        elif pos in REQUIRED_PPE:
+            missing.append(pos)
+        else:
+            unknown.append(pos)
+    return wearing, missing, unknown
diff --git a/saqr/core/detection.py b/core/detection.py
similarity index 100%
rename from saqr/core/detection.py
rename to core/detection.py
diff --git a/saqr/core/drawing.py b/core/drawing.py
similarity index 95%
rename from saqr/core/drawing.py
rename to core/drawing.py
index e09763a..7e124a9 100644
--- a/saqr/core/drawing.py
+++ b/core/drawing.py
@@ -5,8 +5,8 @@ from typing import List, Tuple
 
 import cv2
 
-from saqr.core.compliance import split_wearing_missing
-from saqr.core.detection import STATUSES
+from core.compliance import split_wearing_missing
+from core.detection import STATUSES
 
 GREEN  = (0, 200, 0)
 YELLOW = (0, 200, 255)
diff --git a/saqr/core/events.py b/core/events.py
similarity index 85%
rename from saqr/core/events.py
rename to core/events.py
index 04d433e..0c9e7ab 100644
--- a/saqr/core/events.py
+++ b/core/events.py
@@ -1,6 +1,6 @@
 """Event-line emission and CSV writers.
 
-The ``emit_event`` output line is a contract with ``saqr.robot.bridge`` — its
+The ``emit_event`` output line is a contract with ``robot.bridge`` — its
 regex parses this exact format. Don't change the field order without updating
 the bridge.
 """
@@ -11,8 +11,8 @@ from datetime import datetime
 from pathlib import Path
 from typing import Dict, List
 
-from saqr.core.compliance import split_wearing_missing
-from saqr.core.detection import CLASS_ORDER
+from core.compliance import split_wearing_missing
+from core.detection import CLASS_ORDER
 
 
 def now_iso() -> str:
@@ -21,7 +21,7 @@ def now_iso() -> str:
 
 class EventLogger:
     FIELDS = ["timestamp", "track_id", "event_type", "status",
-              "wearing", "missing", "unknown", "photo", "path"]
+              "wearing", "missing", "unknown", "photo", "path", "snapshot"]
 
     def __init__(self, path: Path):
         self.path = path
@@ -63,15 +63,16 @@ def write_result_csv(tracks: List, output: Path) -> None:
 
 
 def emit_event(track, event_logger: EventLogger, event_type: str = "STATUS_CHANGE",
-               force: bool = False) -> None:
+               force: bool = False, snapshot_path: Path = None) -> None:
     if track.photo_path is None:
         return
     if not force and track.announced_status == track.status:
         return
 
     wearing, missing, unknown = split_wearing_missing(track.items)
+    ts = datetime.now().strftime("%H:%M:%S.%f")[:-3]
     msg = (
-        f"ID {track.track_id:04d} | {event_type} | {track.status} | "
+        f"[{ts}] ID {track.track_id:04d} | {event_type} | {track.status} | "
         f"wearing: {', '.join(wearing) or 'none'} | "
         f"missing: {', '.join(missing) or 'none'} | "
         f"unknown: {', '.join(unknown) or 'none'}"
@@ -88,6 +89,7 @@ def emit_event(track, event_logger: EventLogger, event_type: str = "STATUS_CHANG
         "unknown":    ", ".join(unknown),
         "photo":      track.photo_path.name if track.photo_path else "",
         "path":       str(track.photo_path) if track.photo_path else "",
+        "snapshot":   str(snapshot_path) if snapshot_path else "",
     })
     track.announced_status = track.status
     track.event_count += 1
diff --git a/saqr/core/geometry.py b/core/geometry.py
similarity index 100%
rename from saqr/core/geometry.py
rename to core/geometry.py
diff --git a/saqr/core/grouping.py b/core/grouping.py
similarity index 95%
rename from saqr/core/grouping.py
rename to core/grouping.py
index 9360779..f1842a8 100644
--- a/saqr/core/grouping.py
+++ b/core/grouping.py
@@ -4,8 +4,8 @@ from __future__ import annotations
 from dataclasses import dataclass, field
 from typing import Dict, List, Tuple
 
-from saqr.core.detection import PPEItem
-from saqr.core.geometry import box_distance, expand_bbox, merge_boxes
+from core.detection import PPEItem
+from core.geometry import box_distance, expand_bbox, merge_boxes
 
 
 @dataclass
diff --git a/saqr/core/model.py b/core/model.py
similarity index 90%
rename from saqr/core/model.py
rename to core/model.py
index b24d3b3..340c58b 100644
--- a/saqr/core/model.py
+++ b/core/model.py
@@ -3,7 +3,7 @@ from __future__ import annotations
 
 from pathlib import Path
 
-from saqr.core.paths import MODELS_DIR, PROJECT_ROOT
+from core.paths import MODELS_DIR, PROJECT_ROOT
 
 
 def resolve_model_path(model_arg: str) -> Path:
diff --git a/core/paths.py b/core/paths.py
new file mode 100644
index 0000000..d81a7ab
--- /dev/null
+++ b/core/paths.py
@@ -0,0 +1,28 @@
+"""Canonical project paths, resolved dynamically at import time.
+
+By default the project root is the parent of the ``core/`` package —
+whatever filesystem location ``Saqr/`` currently lives at. Override by
+setting the ``SAQR_ROOT`` environment variable.
+"""
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+_DEFAULT_ROOT = Path(__file__).resolve().parents[1]
+PROJECT_ROOT = Path(os.environ.get("SAQR_ROOT", _DEFAULT_ROOT)).resolve()
+
+CONFIG_DIR   = PROJECT_ROOT / "config"
+DATA_DIR     = PROJECT_ROOT / "data"
+DATASET_DIR  = DATA_DIR / "dataset"
+MODELS_DIR   = DATA_DIR / "models"
+
+RUNTIME_DIR   = PROJECT_ROOT / "runtime"
+CAPTURES_DIR  = RUNTIME_DIR / "captures"
+SNAPSHOTS_DIR = RUNTIME_DIR / "snapshots"
+RUNS_DIR      = RUNTIME_DIR / "runs"
+
+LOGS_DIR     = PROJECT_ROOT / "logs"
+
+RESULT_CSV   = CAPTURES_DIR / "result.csv"
+EVENTS_CSV   = CAPTURES_DIR / "events.csv"
diff --git a/core/pipeline.py b/core/pipeline.py
new file mode 100644
index 0000000..973fbc0
--- /dev/null
+++ b/core/pipeline.py
@@ -0,0 +1,79 @@
+"""Per-frame detect + group + depth-gate + track + capture + emit pipeline.
+
+Announcements are gated on two conditions:
+  * the track is currently stationary (``is_stationary()``), so walkers-by
+    never trigger a reject/TTS — only people planted at the checkpoint do;
+  * the current status differs from the last-announced status for the track,
+    so we never repeat the same announcement.
+"""
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Dict, Optional
+
+from core.camera import estimate_person_distance_m
+from core.capture import save_event_snapshot, save_track_image
+from core.detection import collect_detections
+from core.drawing import draw_track
+from core.events import emit_event, write_result_csv
+from core.grouping import group_detections_to_people
+from core.paths import RESULT_CSV
+from core.tracking import PersonTracker
+
+
+def _filter_by_depth(candidates, depth_frame, depth_scale: float, max_distance_m: float):
+    """Drop candidates whose median depth exceeds ``max_distance_m``.
+
+    Fail-open: candidates with no valid depth pass through. ``max_distance_m``
+    <= 0 disables the filter entirely.
+    """
+    if depth_frame is None or max_distance_m <= 0 or not candidates:
+        return candidates
+    kept = []
+    for cand in candidates:
+        d = estimate_person_distance_m(depth_frame, cand.bbox, depth_scale)
+        if d is None or d <= max_distance_m:
+            cand.distance_m = d
+            kept.append(cand)
+    return kept
+
+
+def process_frame(frame, model, tracker: PersonTracker, frame_idx: int, conf: float,
+                  capture_dirs: Dict[str, Path], write_csv: bool = True,
+                  *,
+                  snapshot_dirs: Optional[Dict[str, Path]] = None,
+                  depth_frame=None, depth_scale: float = 0.001,
+                  max_distance_m: float = 0.0):
+    annotated = frame.copy()
+    h, w = annotated.shape[:2]
+
+    detections = collect_detections(frame, model, conf)
+    candidates = group_detections_to_people(detections, w, h)
+    candidates = _filter_by_depth(candidates, depth_frame, depth_scale, max_distance_m)
+
+    tracker.update(candidates, frame_idx)
+    visible = tracker.visible_tracks()
+
+    # Pass 1: save crops + draw annotations so snapshots capture the full scene.
+    for track in visible:
+        save_track_image(frame, track, capture_dirs)
+        draw_track(annotated, track)
+
+    # Pass 2: emit events ONLY for stationary tracks whose current status
+    # hasn't been announced yet. Walkers-by never reach this branch because
+    # their bbox-center history shows movement above the tolerance.
+    for track in visible:
+        if track.announced_status == track.status:
+            continue
+        if not track.is_stationary():
+            continue
+        ev_type = "NEW" if track.announced_status is None else "STATUS_CHANGE"
+        snap_path = None
+        if snapshot_dirs is not None:
+            snap_path = save_event_snapshot(annotated, track, snapshot_dirs)
+        emit_event(track, tracker.event_logger, ev_type, snapshot_path=snap_path)
+
+    if write_csv:
+        write_result_csv(list(tracker.tracks.values()), RESULT_CSV)
+
+    return annotated, visible
diff --git a/saqr/core/streaming.py b/core/streaming.py
similarity index 81%
rename from saqr/core/streaming.py
rename to core/streaming.py
index a81ec78..4004d31 100644
--- a/saqr/core/streaming.py
+++ b/core/streaming.py
@@ -8,9 +8,11 @@ from typing import Optional
 
 import cv2
 
-from saqr.utils.logger import get_logger
+from utils.config import load_config
+from utils.logger import get_logger
 
 log = get_logger("Inference", "streaming")
+_CFG = load_config("core")["stream"]
 
 _stream_frame: Optional[bytes] = None
 _stream_lock = threading.Lock()
@@ -47,16 +49,18 @@ class MJPEGHandler(BaseHTTPRequestHandler):
         pass
 
 
-def start_stream_server(port: int = 8080):
-    server = HTTPServer(("0.0.0.0", port), MJPEGHandler)
+def start_stream_server(port: Optional[int] = None, host: Optional[str] = None):
+    h = host or _CFG["host"]
+    p = port if port is not None else _CFG["port"]
+    server = HTTPServer((h, p), MJPEGHandler)
     t = threading.Thread(target=server.serve_forever, daemon=True)
     t.start()
-    log.info(f"MJPEG stream server started on http://0.0.0.0:{port}")
+    log.info(f"MJPEG stream server started on http://{h}:{p}")
     return server
 
 
 def update_stream_frame(frame):
     global _stream_frame
-    _, jpeg = cv2.imencode(".jpg", frame, [cv2.IMWRITE_JPEG_QUALITY, 70])
+    _, jpeg = cv2.imencode(".jpg", frame, [cv2.IMWRITE_JPEG_QUALITY, _CFG["jpeg_quality"]])
     with _stream_lock:
         _stream_frame = jpeg.tobytes()
diff --git a/saqr/core/tracking.py b/core/tracking.py
similarity index 69%
rename from saqr/core/tracking.py
rename to core/tracking.py
index 59bb1f1..6565393 100644
--- a/saqr/core/tracking.py
+++ b/core/tracking.py
@@ -1,14 +1,24 @@
-"""Per-person Track dataclass and the greedy nearest-match PersonTracker."""
+"""Per-person Track dataclass and the greedy nearest-match PersonTracker.
+
+Tracks keep a rolling history of bbox centers so ``is_stationary()`` can gate
+announcements: the bridge only fires TTS + arm actions on people standing
+still at the checkpoint, not on anyone walking past.
+"""
 from __future__ import annotations
 
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Dict, List, Optional, Tuple
 
-from saqr.core.compliance import status_from_items
-from saqr.core.events import EventLogger, now_iso
-from saqr.core.geometry import box_distance
-from saqr.core.grouping import PersonCandidate
+from core.compliance import status_from_items
+from core.events import EventLogger, now_iso
+from core.geometry import box_distance
+from core.grouping import PersonCandidate
+from utils.config import load_config
+
+_TRK = load_config("core")["tracking"]
+STATIONARY_CHECK_FRAMES = int(_TRK.get("stationary_check_frames", 15))
+STATIONARY_TOLERANCE_PX = float(_TRK.get("stationary_tolerance_px", 40.0))
 
 
 @dataclass
@@ -26,6 +36,26 @@ class Track:
     event_count:      int = 0
     pending_status:   Optional[str] = None
     pending_count:    int = 0
+    # Rolling history of bbox centers (most recent N frames) used to decide
+    # whether the person is standing still at the checkpoint.
+    recent_centers:   List[Tuple[float, float]] = field(default_factory=list)
+
+    def record_center(self) -> None:
+        cx = (self.bbox[0] + self.bbox[2]) / 2.0
+        cy = (self.bbox[1] + self.bbox[3]) / 2.0
+        self.recent_centers.append((cx, cy))
+        if len(self.recent_centers) > STATIONARY_CHECK_FRAMES:
+            del self.recent_centers[0]
+
+    def is_stationary(self) -> bool:
+        """True when the bbox center hasn't moved more than the configured
+        tolerance across the last ``STATIONARY_CHECK_FRAMES`` frames."""
+        if len(self.recent_centers) < STATIONARY_CHECK_FRAMES:
+            return False
+        xs = [c[0] for c in self.recent_centers]
+        ys = [c[1] for c in self.recent_centers]
+        spread = max(max(xs) - min(xs), max(ys) - min(ys))
+        return spread <= STATIONARY_TOLERANCE_PX
 
 
 class PersonTracker:
@@ -101,6 +131,7 @@ class PersonTracker:
                     track.pending_status = None
                     track.pending_count = 0
 
+            track.record_center()
             used.add(track.track_id)
 
         stale = []
diff --git a/docs/DEPLOY.md b/docs/DEPLOY.md
index 3b17855..ac80bb9 100644
--- a/docs/DEPLOY.md
+++ b/docs/DEPLOY.md
@@ -21,15 +21,27 @@
 ## Repo layout
 
 ```
-saqr/           # python package (core/apps/gui/robot/utils)
-scripts/        # deploy.sh, start_saqr.sh, run_local.sh, run_robot.sh, saqr-bridge.service
-config/         # logging.json
-data/           # dataset/, models/        (gitignored)
-runtime/        # captures/, logs/, runs/  (gitignored)
-docs/           # this file, start.md, use_case_catalogue.pdf
-pyproject.toml  # installs the `saqr`, `saqr-bridge`, `saqr-gui`, ... scripts
+Saqr/
+├── core/           # detection + tracking + events (shared by CLI/GUI/bridge)
+├── apps/           # CLI entry points (saqr, detect, train, manager, view_stream)
+├── gui/            # PySide6 desktop GUI
+├── robot/          # G1 bridge + DDS controller
+├── utils/          # logger
+├── scripts/        # start_saqr.sh (entry), saqr-bridge.service (systemd), deploy.sh
+├── config/         # logging.json
+├── data/           # dataset/, models/        (gitignored)
+├── runtime/        # captures/, runs/         (gitignored)
+├── logs/           # per-module .log files    (gitignored)
+├── docs/           # this file, start.md, use_case_catalogue.pdf
+├── pyproject.toml  # installs the package (no console scripts — use start_saqr.sh)
+└── README.md
 ```
 
+`core/`, `apps/`, `gui/`, `robot/`, and `utils/` are the importable Python
+packages. Run them via `python -m apps.saqr_cli`, `python -m robot.bridge`,
+etc. The project root is auto-detected from `core/paths.py`; override with
+`SAQR_ROOT=/custom/path` if the tree lives elsewhere.
+
 ---
 
 ## Step 1: Train the Model (Dev Machine)
@@ -38,7 +50,7 @@ pyproject.toml  # installs the `saqr`, `saqr-bridge`, `saqr-gui`, ... scripts
 cd ~/Robotics_workspace/yslootahtech/Project/Saqr
 conda activate AI_MSI_yolo
 pip install -e .
-saqr-train --dataset data/dataset --epochs 100 --batch 16
+python -m apps.train_cli --dataset data/dataset --epochs 100 --batch 16
 ```
 
 Verify model exists:
@@ -58,10 +70,9 @@ scripts/deploy.sh --run              # ...and start the bridge
 scripts/deploy.sh --ip 10.0.0.5      # custom robot IP
 ```
 
-The script rsyncs `saqr/`, `scripts/`, `config/`, `docs/`,
-`pyproject.toml`, `requirements.txt`, and `README.md` to
-`~/Saqr` on the robot, then runs `pip install -e .` inside the
-target conda env (default `marcus`).
+The script rsyncs `core/`, `apps/`, `gui/`, `robot/`, `utils/`, `scripts/`,
+`config/`, `docs/`, and `pyproject.toml` to `~/Saqr` on the robot, then runs
+`pip install -e .` inside the target conda env (default `marcus`).
 
 ---
 
@@ -100,56 +111,49 @@ sudo date -s "2026-04-10 15:00:00"
 python -c "from ultralytics import YOLO; print('ultralytics OK')"
 python -c "import torch; print('CUDA:', torch.cuda.is_available())"
 python -c "import cv2; print('opencv OK')"
-python -c "import saqr; print('saqr', saqr.__version__)"
+python -c "import core.paths; print('project root:', core.paths.PROJECT_ROOT)"
 ```
 
 ---
 
 ## Step 4: Run Saqr (Robot)
 
-### Production: bridge with R2+X / R2+Y
-
-The bridge owns the DDS clients and spawns `saqr` on demand. On the robot:
+**The project has one entry point: [scripts/start_saqr.sh](../scripts/start_saqr.sh).**
+Everything — conda activation, DDS iface, camera source, stream port — is
+read from `config/robot_config.json.start_saqr` (env vars override).
 
 ```bash
 cd ~/Saqr
-scripts/start_saqr.sh                 # manual launch
-sudo systemctl restart saqr-bridge    # systemd-managed (see start.md)
+scripts/start_saqr.sh                 # manual (foreground, Ctrl+C to stop)
+sudo systemctl restart saqr-bridge    # systemd-owned (see start.md)
 ```
 
-Or without the helper script:
+The launcher execs `python -m robot.bridge` with the production flags.
+The bridge owns the DDS clients and spawns `python -m apps.saqr_cli` as a
+subprocess when R2+X is pressed.
+
+### Overrides (no config edit)
 
 ```bash
-conda activate marcus
-python -m saqr.robot.bridge --iface eth0 --source realsense --headless -- --stream 8080
+SAQR_SOURCE=/dev/video2 scripts/start_saqr.sh   # V4L2 fallback
+CONDA_ENV=teleimager    scripts/start_saqr.sh
+DDS_IFACE=wlan0         scripts/start_saqr.sh
+STREAM_PORT=9090        scripts/start_saqr.sh
 ```
 
-### Plain saqr (no bridge)
+### Companion utilities (off the run path)
+
+Operators occasionally need these on the robot. They're not part of the
+normal run flow:
 
 ```bash
-# With display
-scripts/run_robot.sh --stream 8080
-
-# Headless
-scripts/run_robot.sh --headless --stream 8080
-
-# V4L2 fallback if RealSense SDK won't enumerate
-scripts/run_robot.sh --source /dev/video2 --headless
+python -m apps.manager_cli --export      # dump a CSV report
+python -m apps.manager_cli               # interactive photo CRUD
 ```
 
-Equivalent `python -m` forms:
-
+Training (dev machine only):
 ```bash
-python -m saqr.apps.saqr_cli --source realsense --model saqr_best.pt --headless
-python -m saqr.apps.detect_cli --source /dev/video2 --model saqr_best.pt
-python -m saqr.apps.manager_cli --export
-```
-
-### Dev machine GUI
-
-```bash
-pip install -e ".[gui]"
-python -m saqr.gui.app --source 0
+python -m apps.train_cli --epochs 100
 ```
 
 ---
@@ -205,7 +209,7 @@ cat runtime/captures/result.csv     # current state per tracked person
 cat runtime/captures/events.csv     # audit log
 ls runtime/captures/{SAFE,PARTIAL,UNSAFE}/
 
-saqr-manager --export               # quick CSV export
+python -m apps.manager_cli --export # quick CSV export
 
 # Download to dev machine
 scp -r unitree@192.168.123.164:~/Saqr/runtime/captures ./captures_from_robot
@@ -268,7 +272,7 @@ saqr --source realsense --headless \
 | `result.csv` | `runtime/captures/` | Current state of tracked persons |
 | `events.csv` | `runtime/captures/` | Audit log (NEW / STATUS_CHANGE) |
 | Person crops | `runtime/captures/{SAFE,PARTIAL,UNSAFE}/*.jpg` | Latest crop per track |
-| Logs | `runtime/logs/Inference/*.log` | Module log output |
+| Logs | `logs/Inference/*.log` | Module log output |
 
 ---
 
@@ -276,17 +280,17 @@ saqr --source realsense --headless \
 
 | Path | Purpose |
 |------|---------|
-| [saqr/apps/saqr_cli.py](../saqr/apps/saqr_cli.py) | Main PPE tracking entry (`saqr`) |
-| [saqr/robot/bridge.py](../saqr/robot/bridge.py) | Saqr → G1 bridge (R2+X/R2+Y) |
-| [saqr/robot/robot_controller.py](../saqr/robot/robot_controller.py) | G1 arm + audio + TTS worker |
-| [saqr/robot/controller.py](../saqr/robot/controller.py) | G1 wireless-remote DDS reader |
-| [saqr/core/pipeline.py](../saqr/core/pipeline.py) | Per-frame detect + track + emit |
-| [saqr/core/tracking.py](../saqr/core/tracking.py) | `PersonTracker`, `Track` |
-| [saqr/core/events.py](../saqr/core/events.py) | Event-line format (contract with bridge) |
-| [saqr/apps/detect_cli.py](../saqr/apps/detect_cli.py) | Simple detection without tracking |
-| [saqr/apps/train_cli.py](../saqr/apps/train_cli.py) | YOLO11n training |
-| [saqr/apps/manager_cli.py](../saqr/apps/manager_cli.py) | Capture CRUD + CSV export |
-| [saqr/gui/app.py](../saqr/gui/app.py) | PySide6 desktop GUI |
+| [apps/saqr_cli.py](../apps/saqr_cli.py) | Main PPE tracking entry (`saqr`) |
+| [robot/bridge.py](../robot/bridge.py) | Saqr → G1 bridge (R2+X/R2+Y) |
+| [robot/robot_controller.py](../robot/robot_controller.py) | G1 arm + audio + TTS worker |
+| [robot/controller.py](../robot/controller.py) | G1 wireless-remote DDS reader |
+| [core/pipeline.py](../core/pipeline.py) | Per-frame detect + track + emit |
+| [core/tracking.py](../core/tracking.py) | `PersonTracker`, `Track` |
+| [core/events.py](../core/events.py) | Event-line format (contract with bridge) |
+| [apps/detect_cli.py](../apps/detect_cli.py) | Simple detection without tracking |
+| [apps/train_cli.py](../apps/train_cli.py) | YOLO11n training |
+| [apps/manager_cli.py](../apps/manager_cli.py) | Capture CRUD + CSV export |
+| [gui/app.py](../gui/app.py) | PySide6 desktop GUI |
 
 ---
 
diff --git a/docs/start.md b/docs/start.md
index d222874..7e31b1a 100644
--- a/docs/start.md
+++ b/docs/start.md
@@ -1,6 +1,6 @@
 # Saqr — Auto-start on boot
 
-How to auto-start `saqr.robot.bridge` on every boot of the Unitree G1
+How to auto-start `robot.bridge` on every boot of the Unitree G1
 (Jetson), via `systemd` + `scripts/start_saqr.sh`.
 
 ---
@@ -9,7 +9,7 @@ How to auto-start `saqr.robot.bridge` on every boot of the Unitree G1
 
 | File | Role |
 |------|------|
-| `~/Saqr/saqr/robot/bridge.py` | The bridge process (DDS + TTS + R2+X/R2+Y trigger loop). Entry point: `python -m saqr.robot.bridge`. |
+| `~/Saqr/robot/bridge.py` | The bridge process (DDS + TTS + R2+X/R2+Y trigger loop). Entry point: `python -m robot.bridge`. |
 | `~/Saqr/scripts/start_saqr.sh` | Bash launcher: sources conda, activates `marcus`, `cd ~/Saqr`, exec the bridge with the production flags. |
 | `~/Saqr/scripts/saqr-bridge.service` | systemd unit that runs `start_saqr.sh` as user `unitree` on every boot, restarts on failure, logs to journalctl. |
 
@@ -164,6 +164,6 @@ sudo systemctl restart saqr-bridge
 ### Bridge is running twice
 
 ```bash
-ps -ef | grep "saqr.robot.bridge"
+ps -ef | grep "robot.bridge"
 sudo systemctl restart saqr-bridge
 ```
diff --git a/saqr/utils/__init__.py b/gui/__init__.py
similarity index 100%
rename from saqr/utils/__init__.py
rename to gui/__init__.py
diff --git a/saqr/gui/app.py b/gui/app.py
similarity index 91%
rename from saqr/gui/app.py
rename to gui/app.py
index 147eba4..f3b78b2 100644
--- a/saqr/gui/app.py
+++ b/gui/app.py
@@ -18,18 +18,20 @@ from PySide6.QtWidgets import (
 )
 from ultralytics import YOLO
 
-from saqr.core.capture import save_track_image, setup_capture_dirs
-from saqr.core.compliance import split_wearing_missing
-from saqr.core.detection import STATUSES, collect_detections
-from saqr.core.drawing import draw_counters, draw_track
-from saqr.core.events import EventLogger, emit_event, write_result_csv
-from saqr.core.grouping import group_detections_to_people
-from saqr.core.model import resolve_model_path
-from saqr.core.paths import EVENTS_CSV, MODELS_DIR, PROJECT_ROOT, RESULT_CSV
-from saqr.core.tracking import PersonTracker
-from saqr.utils.logger import get_logger
+from core.capture import save_track_image, setup_capture_dirs
+from core.compliance import split_wearing_missing
+from core.detection import STATUSES, collect_detections
+from core.drawing import draw_counters, draw_track
+from core.events import EventLogger, emit_event, write_result_csv
+from core.grouping import group_detections_to_people
+from core.model import resolve_model_path
+from core.paths import EVENTS_CSV, MODELS_DIR, PROJECT_ROOT, RESULT_CSV
+from core.tracking import PersonTracker
+from utils.config import load_config
+from utils.logger import get_logger
 
 log = get_logger("Inference", "gui")
+_CORE = load_config("core")
 
 
 def list_cameras(max_idx: int = 10) -> List[str]:
@@ -47,7 +49,10 @@ def list_cameras(max_idx: int = 10) -> List[str]:
     return sources if sources else ["0"]
 
 
-def open_camera(source: str, width: int = 640, height: int = 480, fps: int = 30):
+def open_camera(source: str,
+                width:  int = _CORE["camera"]["width"],
+                height: int = _CORE["camera"]["height"],
+                fps:    int = _CORE["camera"]["fps"]):
     if source.startswith("/dev/video"):
         cap = cv2.VideoCapture(source, cv2.CAP_V4L2)
     elif source.isdigit():
@@ -191,7 +196,9 @@ def cv_to_qpixmap(frame: np.ndarray, max_w: int = 960, max_h: int = 720) -> QPix
 
 
 class SaqrWindow(QMainWindow):
-    def __init__(self, default_model: str = "saqr_best.pt", default_source: str = "0"):
+    def __init__(self,
+                 default_model:  str = _CORE["detection"]["default_model"],
+                 default_source: str = _CORE["camera"]["default_source"]):
         super().__init__()
         self.setWindowTitle("Saqr - PPE Safety Tracking")
         self.setMinimumSize(1200, 700)
@@ -235,26 +242,26 @@ class SaqrWindow(QMainWindow):
         self.conf_spin = QDoubleSpinBox()
         self.conf_spin.setRange(0.1, 0.9)
         self.conf_spin.setSingleStep(0.05)
-        self.conf_spin.setValue(0.35)
+        self.conf_spin.setValue(_CORE["detection"]["conf"])
         param_lay.addWidget(self.conf_spin, 0, 1)
 
         param_lay.addWidget(QLabel("Max Missing:"), 1, 0)
         self.missing_spin = QSpinBox()
         self.missing_spin.setRange(10, 300)
-        self.missing_spin.setValue(90)
+        self.missing_spin.setValue(_CORE["tracking"]["max_missing"])
         param_lay.addWidget(self.missing_spin, 1, 1)
 
         param_lay.addWidget(QLabel("Match Dist:"), 2, 0)
         self.dist_spin = QDoubleSpinBox()
         self.dist_spin.setRange(50, 500)
         self.dist_spin.setSingleStep(10)
-        self.dist_spin.setValue(250)
+        self.dist_spin.setValue(_CORE["tracking"]["match_distance"])
         param_lay.addWidget(self.dist_spin, 2, 1)
 
         param_lay.addWidget(QLabel("Confirm Frames:"), 3, 0)
         self.confirm_spin = QSpinBox()
         self.confirm_spin.setRange(1, 20)
-        self.confirm_spin.setValue(5)
+        self.confirm_spin.setValue(_CORE["tracking"]["status_confirm_frames"])
         param_lay.addWidget(self.confirm_spin, 3, 1)
 
         left.addWidget(param_grp)
@@ -437,7 +444,7 @@ class SaqrWindow(QMainWindow):
             "CSV Files (*.csv)"
         )
         if path:
-            from saqr.apps.manager_cli import export_csv, load_photos
+            from apps.manager_cli import export_csv, load_photos
             export_csv(load_photos(), Path(path))
             self._on_event(f"Exported: {path}")
 
@@ -449,8 +456,8 @@ class SaqrWindow(QMainWindow):
 def main():
     import argparse
     parser = argparse.ArgumentParser(description="Saqr PPE GUI")
-    parser.add_argument("--model",  default="saqr_best.pt")
-    parser.add_argument("--source", default="0")
+    parser.add_argument("--model",  default=_CORE["detection"]["default_model"])
+    parser.add_argument("--source", default=_CORE["camera"]["default_source"])
     args = parser.parse_args()
 
     app = QApplication(sys.argv)
diff --git a/pyproject.toml b/pyproject.toml
index 074d70b..c01c09b 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -18,14 +18,9 @@ dependencies = [
 gui = ["PySide6>=6.5.0"]
 realsense = ["pyrealsense2"]
 
-[project.scripts]
-saqr            = "saqr.apps.saqr_cli:main"
-saqr-detect     = "saqr.apps.detect_cli:main"
-saqr-train      = "saqr.apps.train_cli:main"
-saqr-manager    = "saqr.apps.manager_cli:main"
-saqr-view-stream = "saqr.apps.view_stream:main"
-saqr-gui        = "saqr.gui.app:main"
-saqr-bridge     = "saqr.robot.bridge:main"
+# No console scripts — the project is launched through scripts/start_saqr.sh
+# (which calls `python -m robot.bridge`). That bridge then spawns
+# `python -m apps.saqr_cli` as a subprocess. Everything else is a module.
 
 [tool.setuptools.packages.find]
-include = ["saqr*"]
+include = ["core*", "apps*", "gui*", "robot*", "utils*"]
diff --git a/robot/__init__.py b/robot/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/robot/audio_player.py b/robot/audio_player.py
new file mode 100644
index 0000000..6bcd55f
--- /dev/null
+++ b/robot/audio_player.py
@@ -0,0 +1,181 @@
+"""Plays pre-recorded WAV clips via AudioClient.PlayStream — bypasses TtsMaker.
+
+Expected WAV format (required by the G1 audio channel):
+    16 kHz, mono, 16-bit signed PCM.
+
+Library layout under assets/audio/:
+    fixed/<key>.wav              e.g. safe.wav, unsafe_generic.wav, ready.wav
+    unsafe_missing/<key>.wav     e.g. helmet.wav, vest.wav, helmet_vest.wav
+                                  (key is sorted-joined PPE names, "_" separator)
+
+Callers look up by (category, key). Missing clip → returns False so the caller
+can fall back to TtsMaker.
+"""
+from __future__ import annotations
+
+import datetime
+import json
+import time
+import wave
+from pathlib import Path
+from typing import Dict, Optional, Tuple
+
+from core.paths import PROJECT_ROOT
+
+AUDIO_ROOT = PROJECT_ROOT / "assets" / "audio"
+
+
+def _ts() -> str:
+    return datetime.datetime.now().strftime("%H:%M:%S.%f")[:-3]
+
+EXPECTED_RATE     = 16000
+EXPECTED_CHANNELS = 1
+EXPECTED_WIDTH    = 2          # bytes per sample (int16)
+PLAY_CHUNK_BYTES  = 96000      # 3 s per PlayStream call (matches the Unitree example)
+PLAY_APP_NAME     = "saqr_audio"
+
+# The G1 arm action and audio stack share a firmware busy state. If we try to
+# PlayStream while an arm action is still being processed, chunk 0 is often
+# rejected with rc=3104. Later chunks can fail too if the firmware hasn't
+# fully processed the previous chunk. Retry both cases with back-off.
+CHUNK0_RETRIES     = 4
+CHUNK0_BACKOFF_S   = 1.0    # 1,2,3,4 s — total ≈ 10 s, covers a full arm cycle
+CHUNKN_RETRIES     = 2
+CHUNKN_BACKOFF_S   = 1.0    # 1,2 s — firmware usually clears within 1-2 s
+PRE_STREAM_SLEEP   = 0.1    # pause after PlayStop before first PlayStream
+INTER_CHUNK_MARGIN = 0.1    # extra sleep after each chunk's audio duration
+
+
+def _read_wav_pcm(path: Path) -> Optional[bytes]:
+    """Return the raw PCM bytes if the WAV matches the expected format, else None."""
+    try:
+        with wave.open(str(path), "rb") as wf:
+            ch = wf.getnchannels()
+            sw = wf.getsampwidth()
+            fr = wf.getframerate()
+            if ch != EXPECTED_CHANNELS or sw != EXPECTED_WIDTH or fr != EXPECTED_RATE:
+                print(
+                    f"[audio_player {_ts()}][WARN] {path}: expected "
+                    f"{EXPECTED_RATE} Hz mono 16-bit; got "
+                    f"{fr} Hz {ch}-ch {sw*8}-bit. Skipping.",
+                    flush=True,
+                )
+                return None
+            return wf.readframes(wf.getnframes())
+    except Exception as e:
+        print(f"[audio_player {_ts()}][WARN] failed to load {path}: {e}", flush=True)
+        return None
+
+
+class AudioPlayer:
+    """Loads WAVs under ``assets/audio/<category>/<key>.wav`` and plays them on the G1."""
+
+    def __init__(self, audio_client):
+        self.audio_client = audio_client
+        self._clips: Dict[Tuple[str, str], bytes] = {}
+        self._load_all()
+        if self._clips:
+            print(f"[audio_player {_ts()}] loaded {len(self._clips)} clip(s): "
+                  f"{sorted(self._clips.keys())}", flush=True)
+        else:
+            print(f"[audio_player {_ts()}] no clips found under {AUDIO_ROOT}", flush=True)
+
+    # ── library ─────────────────────────────────────────────────────────────
+    def _load_all(self) -> None:
+        if not AUDIO_ROOT.exists():
+            return
+        for category_dir in sorted(AUDIO_ROOT.iterdir()):
+            if not category_dir.is_dir():
+                continue
+            for wav_path in sorted(category_dir.glob("*.wav")):
+                pcm = _read_wav_pcm(wav_path)
+                if pcm is not None:
+                    self._clips[(category_dir.name, wav_path.stem)] = pcm
+
+    def has(self, category: str, key: str) -> bool:
+        return (category, key) in self._clips
+
+    # ── playback ────────────────────────────────────────────────────────────
+    def play(self, category: str, key: str) -> bool:
+        """Blocking play. Returns True on success, False if clip missing / failed."""
+        pcm = self._clips.get((category, key))
+        if pcm is None:
+            return False
+        if self.audio_client is None:
+            return False
+
+        # Re-assert max volume on every play; the firmware sometimes resets
+        # between sessions or after certain events. Matches the Unitree
+        # reference script's pattern.
+        try:
+            self.audio_client.SetVolume(100)
+        except Exception:
+            pass
+
+        # Cancel any in-flight stream from a previous call.
+        try:
+            from unitree_sdk2py.g1.audio.g1_audio_api import ROBOT_API_ID_AUDIO_STOP_PLAY
+            self.audio_client._Call(
+                ROBOT_API_ID_AUDIO_STOP_PLAY,
+                json.dumps({"app_name": PLAY_APP_NAME}),
+            )
+        except Exception:
+            pass
+        time.sleep(PRE_STREAM_SLEEP)
+
+        sid = f"saqr_{int(time.time() * 1000)}"
+        offset = 0
+        chunk0_attempts = 0
+        chunkn_attempts = 0
+        while offset < len(pcm):
+            chunk = pcm[offset:offset + PLAY_CHUNK_BYTES]
+            code, _ = self.audio_client.PlayStream(PLAY_APP_NAME, sid, chunk)
+
+            if code != 0:
+                # chunk 0: likely firmware busy (arm motion colliding with audio).
+                # retry with linear back-off; new sid so firmware sees a fresh stream.
+                if offset == 0 and chunk0_attempts < CHUNK0_RETRIES:
+                    chunk0_attempts += 1
+                    delay = CHUNK0_BACKOFF_S * chunk0_attempts
+                    print(f"[audio_player {_ts()}][WARN] PlayStream rc={code} at byte 0; "
+                          f"retry {chunk0_attempts}/{CHUNK0_RETRIES} in {delay:.1f}s",
+                          flush=True)
+                    time.sleep(delay)
+                    sid = f"saqr_{int(time.time() * 1000)}"
+                    continue
+
+                # later chunks: firmware is still processing the previous chunk.
+                # Retry a couple of times with back-off, keeping the same sid.
+                if offset > 0 and chunkn_attempts < CHUNKN_RETRIES:
+                    chunkn_attempts += 1
+                    delay = CHUNKN_BACKOFF_S * chunkn_attempts
+                    print(f"[audio_player {_ts()}][WARN] PlayStream rc={code} at byte {offset}; "
+                          f"mid-stream retry {chunkn_attempts}/{CHUNKN_RETRIES} in {delay:.1f}s",
+                          flush=True)
+                    time.sleep(delay)
+                    continue
+
+                print(f"[audio_player {_ts()}][WARN] PlayStream rc={code} at byte {offset} "
+                      f"(retries exhausted)", flush=True)
+                return False
+
+            if offset == 0 and chunk0_attempts > 0:
+                print(f"[audio_player {_ts()}] chunk 0 succeeded after "
+                      f"{chunk0_attempts} retry/retries", flush=True)
+            elif offset > 0 and chunkn_attempts > 0:
+                print(f"[audio_player {_ts()}] chunk at byte {offset} succeeded after "
+                      f"{chunkn_attempts} retry/retries", flush=True)
+                chunkn_attempts = 0   # reset for any subsequent chunk
+
+            offset += len(chunk)
+            # Wait for the chunk to finish playing before sending the next.
+            chunk_seconds = len(chunk) / (EXPECTED_RATE * EXPECTED_WIDTH)
+            time.sleep(chunk_seconds + INTER_CHUNK_MARGIN)
+
+        # Short settle time after the final chunk before PlayStop.
+        time.sleep(0.2)
+        try:
+            self.audio_client.PlayStop(PLAY_APP_NAME)
+        except Exception:
+            pass
+        return True
diff --git a/saqr/robot/bridge.py b/robot/bridge.py
similarity index 67%
rename from saqr/robot/bridge.py
rename to robot/bridge.py
index 75022d9..000241d 100644
--- a/saqr/robot/bridge.py
+++ b/robot/bridge.py
@@ -11,7 +11,7 @@ R2+Y stops it. While saqr is running the bridge parses its event stream and:
 
 See docs/DEPLOY.md for wireless-remote workflow and systemd deploy notes.
 
-Saqr event line format (from saqr.core.events.emit_event):
+Saqr event line format (from core.events.emit_event):
     ID 0001 | NEW | UNSAFE | wearing: ... | missing: ... | unknown: ...
     ID 0001 | STATUS_CHANGE | SAFE  | wearing: ... | missing: ... | unknown: ...
 """
@@ -27,32 +27,42 @@ import threading
 import time
 from typing import Dict, Optional
 
-from saqr.core.paths import PROJECT_ROOT
-from saqr.robot.robot_controller import RobotController
+from core.paths import PROJECT_ROOT
+from robot.robot_controller import RobotController
+from utils.config import load_config
+
+import datetime
+
+
+def _ts() -> str:
+    return datetime.datetime.now().strftime("%H:%M:%S.%f")[:-3]
+
+_ROBOT   = load_config("robot")
+_BRIDGE  = _ROBOT["bridge"]
+_PHRASES = _ROBOT["phrases"]
 
 DANGER_STATUS = "UNSAFE"
 SAFE_STATUS   = "SAFE"
 
-# speaker_id=2 was confirmed English on current G1 firmware via
-# Project/Sanad/voice_example.py mode 6. speaker_id=0 is Chinese.
-TTS_SPEAKER_ID = 2
+# speaker_id is locked to a language by G1 firmware: 2=English, 0=Chinese.
+# Confirmed via Project/Sanad/voice_example.py mode 6.
+TTS_SPEAKER_ID = _ROBOT["tts"]["speaker_id"]
 
-TTS_TEXT_SAFE           = "Safe to enter. Have a good day."
-TTS_UNSAFE_WITH_MISSING = (
-    "Please stop. Wear your proper safety equipment. You are missing {items}."
-)
-TTS_UNSAFE_GENERIC      = "Please stop. Wear your proper safety equipment."
-TTS_BRIDGE_DEACTIVATED  = "Saqr deactivated."
-TTS_BRIDGE_READY        = "Saqr is running. Press R2 plus X to start."
-TTS_BRIDGE_NO_CAMERA    = (
-    "Camera not connected. Please plug in the camera and try again."
-)
+TTS_TEXT_SAFE           = _PHRASES["safe"]
+TTS_UNSAFE_WITH_MISSING = _PHRASES["unsafe_missing"]
+TTS_UNSAFE_GENERIC      = _PHRASES["unsafe_generic"]
+TTS_BRIDGE_DEACTIVATED  = _PHRASES["deactivated"]
+TTS_BRIDGE_READY        = _PHRASES["ready"]
+TTS_BRIDGE_NO_CAMERA    = _PHRASES["no_camera"]
 
-QUICK_FAIL_WINDOW_S = 8.0
+QUICK_FAIL_WINDOW_S = _BRIDGE["quick_fail_window"]
+TRIGGER_POLL_HZ     = _BRIDGE["trigger_poll_hz"]
 
-# ID NNNN | EVENT_TYPE | STATUS | wearing: ... | missing: ... | unknown: ...
+# [HH:MM:SS.fff] ID NNNN | EVENT_TYPE | STATUS | wearing: ... | missing: ... | unknown: ...
+# The leading timestamp is optional for backwards compatibility with old logs.
 EVENT_RE = re.compile(
-    r"^ID\s+(?P<id>\d+)\s*\|\s*"
+    r"^(?:\[[\d:.]+\]\s+)?"
+    r"ID\s+(?P<id>\d+)\s*\|\s*"
     r"(?P<event>NEW|STATUS_CHANGE)\s*\|\s*"
     r"(?P<status>SAFE|PARTIAL|UNSAFE)\s*\|\s*"
     r"wearing:\s*(?P<wearing>[^|]*?)\s*\|\s*"
@@ -86,7 +96,7 @@ def build_unsafe_tts(missing: list) -> str:
 
 def build_saqr_cmd(saqr_extra_args: list) -> list:
     """Invoke the saqr CLI via ``python -m`` so it picks up the package layout."""
-    return [sys.executable, "-u", "-m", "saqr.apps.saqr_cli", *saqr_extra_args]
+    return [sys.executable, "-u", "-m", "apps.saqr_cli", *saqr_extra_args]
 
 
 def split_argv(argv):
@@ -101,10 +111,12 @@ class Bridge:
 
     def __init__(self, robot: RobotController, cooldown_s: float,
                  release_after_s: float, saqr_args: list,
-                 env: Dict[str, str], cwd: str):
+                 env: Dict[str, str], cwd: str,
+                 audio_lead_s: float = 0.3):
         self.robot = robot
         self.cooldown_s = cooldown_s
         self.release_after_s = release_after_s
+        self.audio_lead_s = audio_lead_s
         self.saqr_args = saqr_args
         self.env = env
         self.cwd = cwd
@@ -125,11 +137,11 @@ class Bridge:
     def start_saqr(self):
         with self._proc_lock:
             if self.proc is not None and self.proc.poll() is None:
-                print("[BRIDGE] start ignored — saqr already running", flush=True)
+                print(f"[BRIDGE {_ts()}] start ignored — saqr already running", flush=True)
                 return
 
             cmd = build_saqr_cmd(self.saqr_args)
-            print(f"[BRIDGE] starting saqr: {' '.join(cmd)}", flush=True)
+            print(f"[BRIDGE {_ts()}] starting saqr: {' '.join(cmd)}", flush=True)
             self.proc = subprocess.Popen(
                 cmd,
                 cwd=self.cwd,
@@ -154,10 +166,10 @@ class Bridge:
         with self._proc_lock:
             proc = self.proc
             if proc is None or proc.poll() is not None:
-                print("[BRIDGE] stop ignored — saqr not running", flush=True)
+                print(f"[BRIDGE {_ts()}] stop ignored — saqr not running", flush=True)
                 self.proc = None
                 return
-            print("[BRIDGE] stopping saqr (SIGINT)", flush=True)
+            print(f"[BRIDGE {_ts()}] stopping saqr (SIGINT)", flush=True)
             try:
                 proc.send_signal(signal.SIGINT)
             except Exception:
@@ -166,12 +178,12 @@ class Bridge:
         try:
             proc.wait(timeout=3.0)
         except subprocess.TimeoutExpired:
-            print("[BRIDGE] saqr did not exit in 3s, sending SIGTERM", flush=True)
+            print(f"[BRIDGE {_ts()}] saqr did not exit in 3s, sending SIGTERM", flush=True)
             try:
                 proc.terminate()
                 proc.wait(timeout=2.0)
             except subprocess.TimeoutExpired:
-                print("[BRIDGE] saqr unresponsive, sending SIGKILL", flush=True)
+                print(f"[BRIDGE {_ts()}] saqr unresponsive, sending SIGKILL", flush=True)
                 proc.kill()
                 proc.wait()
 
@@ -182,7 +194,7 @@ class Bridge:
             self.proc = None
             self.reader_thread = None
 
-        self.robot.speak(TTS_BRIDGE_DEACTIVATED)
+        self.robot.speak(TTS_BRIDGE_DEACTIVATED, category="fixed", key="deactivated")
 
     def _read_stdout(self, proc: subprocess.Popen):
         start_t = self._proc_start_t
@@ -191,17 +203,17 @@ class Bridge:
             for line in proc.stdout:
                 self.handle_line(line)
         except Exception as e:
-            print(f"[BRIDGE][ERR] reader thread: {e}", flush=True)
+            print(f"[BRIDGE {_ts()}][ERR] reader thread: {e}", flush=True)
         rc = proc.wait()
         lifetime = time.time() - start_t if start_t > 0 else 0.0
-        print(f"[BRIDGE] saqr exited rc={rc} (lifetime={lifetime:.1f}s)",
+        print(f"[BRIDGE {_ts()}] saqr exited rc={rc} (lifetime={lifetime:.1f}s)",
               flush=True)
 
         if rc not in (0, -2) and 0 < lifetime < QUICK_FAIL_WINDOW_S:
             try:
-                self.robot.speak(TTS_BRIDGE_NO_CAMERA)
+                self.robot.speak(TTS_BRIDGE_NO_CAMERA, category="fixed", key="no_camera")
             except Exception as e:
-                print(f"[BRIDGE][ERR] no-camera tts failed: {e}", flush=True)
+                print(f"[BRIDGE {_ts()}][ERR] no-camera tts failed: {e}", flush=True)
 
     def handle_line(self, line: str):
         line = line.rstrip()
@@ -235,21 +247,33 @@ class Bridge:
 
         try:
             if status == DANGER_STATUS:
-                self.robot.speak(build_unsafe_tts(missing))
+                # Fire audio first, give the worker thread a head start so
+                # PlayStream reaches the firmware BEFORE the arm command.
+                # Once audio is actively playing, the arm command queues
+                # behind it without blocking playback → audible overlap.
+                if missing:
+                    key = "_".join(sorted(missing))
+                    self.robot.speak(build_unsafe_tts(missing),
+                                     category="unsafe_missing", key=key)
+                else:
+                    self.robot.speak(TTS_UNSAFE_GENERIC,
+                                     category="fixed", key="unsafe_generic")
+                time.sleep(self.audio_lead_s)
                 self.robot.reject(release_after=self.release_after_s)
             else:
-                self.robot.speak(TTS_TEXT_SAFE)
+                self.robot.speak(TTS_TEXT_SAFE,
+                                 category="fixed", key="safe")
         except Exception as e:
-            print(f"[BRIDGE][ERR] robot action failed: {e}", flush=True)
+            print(f"[BRIDGE {_ts()}][ERR] robot action failed: {e}", flush=True)
 
 
 def trigger_loop(bridge: Bridge, hub, stop_event: threading.Event,
-                 poll_hz: float = 50.0):
+                 poll_hz: float = TRIGGER_POLL_HZ):
     """Watch the wireless remote for R2+X (start) and R2+Y (stop)."""
     period = 1.0 / max(poll_hz, 1.0)
     waiting_release_x = False
     waiting_release_y = False
-    print("[BRIDGE] trigger loop ready — press R2+X to start, R2+Y to stop.",
+    print(f"[BRIDGE {_ts()}] trigger loop ready — press R2+X to start, R2+Y to stop.",
           flush=True)
     while not stop_event.is_set():
         time.sleep(period)
@@ -264,22 +288,22 @@ def trigger_loop(bridge: Bridge, hub, stop_event: threading.Event,
                 waiting_release_x = False
         elif r2x:
             waiting_release_x = True
-            print("[BRIDGE] R2+X pressed -> start saqr", flush=True)
+            print(f"[BRIDGE {_ts()}] R2+X pressed -> start saqr", flush=True)
             try:
                 bridge.start_saqr()
             except Exception as e:
-                print(f"[BRIDGE][ERR] start_saqr failed: {e}", flush=True)
+                print(f"[BRIDGE {_ts()}][ERR] start_saqr failed: {e}", flush=True)
 
         if waiting_release_y:
             if not r2y:
                 waiting_release_y = False
         elif r2y:
             waiting_release_y = True
-            print("[BRIDGE] R2+Y pressed -> stop saqr", flush=True)
+            print(f"[BRIDGE {_ts()}] R2+Y pressed -> stop saqr", flush=True)
             try:
                 bridge.stop_saqr()
             except Exception as e:
-                print(f"[BRIDGE][ERR] stop_saqr failed: {e}", flush=True)
+                print(f"[BRIDGE {_ts()}][ERR] stop_saqr failed: {e}", flush=True)
 
 
 def main():
@@ -288,14 +312,14 @@ def main():
     ap = argparse.ArgumentParser(
         description="Bridge Saqr PPE events to the G1 arm 'reject' action."
     )
-    ap.add_argument("--iface", default=None,
+    ap.add_argument("--iface",         default=_BRIDGE["iface"],
                     help="DDS network interface (e.g. eth0).")
-    ap.add_argument("--timeout", type=float, default=10.0)
-    ap.add_argument("--cooldown", type=float, default=8.0)
-    ap.add_argument("--release-after", type=float, default=2.0)
-    ap.add_argument("--dry-run", action="store_true")
-    ap.add_argument("--speaker-id", type=int, default=TTS_SPEAKER_ID)
-    ap.add_argument("--no-trigger", action="store_true")
+    ap.add_argument("--timeout",       type=float, default=_BRIDGE["timeout"])
+    ap.add_argument("--cooldown",      type=float, default=_BRIDGE["cooldown"])
+    ap.add_argument("--release-after", type=float, default=_BRIDGE["release_after"])
+    ap.add_argument("--dry-run",       action="store_true")
+    ap.add_argument("--speaker-id",    type=int,   default=TTS_SPEAKER_ID)
+    ap.add_argument("--no-trigger",    action="store_true")
 
     ap.add_argument("--source", default=None)
     ap.add_argument("--headless", action="store_true")
@@ -335,18 +359,19 @@ def main():
         robot=robot,
         cooldown_s=args.cooldown,
         release_after_s=args.release_after,
+        audio_lead_s=float(_BRIDGE.get("audio_lead_s", 0.3)),
         saqr_args=saqr_args,
         env=env,
         cwd=str(PROJECT_ROOT),
     )
-    print(f"[BRIDGE] saqr cmd template: {' '.join(build_saqr_cmd(saqr_args))}",
+    print(f"[BRIDGE {_ts()}] saqr cmd template: {' '.join(build_saqr_cmd(saqr_args))}",
           flush=True)
-    print(f"[BRIDGE] cwd: {PROJECT_ROOT}", flush=True)
+    print(f"[BRIDGE {_ts()}] cwd: {PROJECT_ROOT}", flush=True)
 
     stop_event = threading.Event()
 
     def _forward_signal(signum, _frame):
-        print(f"[BRIDGE] signal {signum} -> shutting down", flush=True)
+        print(f"[BRIDGE {_ts()}] signal {signum} -> shutting down", flush=True)
         stop_event.set()
 
     signal.signal(signal.SIGINT, _forward_signal)
@@ -354,16 +379,16 @@ def main():
 
     have_hub = use_trigger and robot.hub is not None
     if use_trigger and not have_hub:
-        print("[BRIDGE][WARN] --no-trigger not set, but no LowState hub is "
+        print(f"[BRIDGE {_ts()}][WARN] --no-trigger not set, but no LowState hub is "
               "available. Falling back to legacy auto-start mode.", flush=True)
 
     trigger_thread: Optional[threading.Thread] = None
     try:
         if have_hub:
             try:
-                robot.speak(TTS_BRIDGE_READY)
+                robot.speak(TTS_BRIDGE_READY, category="fixed", key="ready")
             except Exception as e:
-                print(f"[BRIDGE][WARN] startup announce failed: {e}", flush=True)
+                print(f"[BRIDGE {_ts()}][WARN] startup announce failed: {e}", flush=True)
 
             trigger_thread = threading.Thread(
                 target=trigger_loop,
@@ -388,7 +413,7 @@ def main():
             robot.shutdown_tts()
         except Exception:
             pass
-        print("[BRIDGE] bye.", flush=True)
+        print(f"[BRIDGE {_ts()}] bye.", flush=True)
         sys.exit(0)
 
 
diff --git a/saqr/robot/controller.py b/robot/controller.py
similarity index 100%
rename from saqr/robot/controller.py
rename to robot/controller.py
diff --git a/saqr/robot/robot_controller.py b/robot/robot_controller.py
similarity index 51%
rename from saqr/robot/robot_controller.py
rename to robot/robot_controller.py
index 03014bc..afe19be 100644
--- a/saqr/robot/robot_controller.py
+++ b/robot/robot_controller.py
@@ -1,8 +1,10 @@
 """G1 arm + audio + LowState DDS client owned by the bridge.
 
-A dedicated TTS worker thread paces ``TtsMaker`` calls so overlapping phrases
-don't trip the SDK's "device busy" error (3104). The busy multiplier adapts
-up on 3104s and decays on clean calls.
+Announcements run on a dedicated worker thread. Each queue item is a tuple
+``(text, category, key)``. The worker picks WAV playback via
+``AudioClient.PlayStream`` when the clip exists under ``assets/audio/`` and
+``tts.mode`` allows, otherwise falls back to ``TtsMaker`` with the adaptive
+busy-factor backoff for 3104 ("device busy") errors.
 """
 from __future__ import annotations
 
@@ -10,20 +12,28 @@ import collections
 import datetime
 import threading
 import time
-from typing import Deque, Optional
+from typing import Deque, Optional, Tuple
 
-TTS_VOLUME = 100
+from utils.config import load_config
 
-TTS_SECONDS_PER_CHAR = 0.12
-TTS_MIN_SECONDS      = 2.5
-TTS_QUEUE_MAX        = 4
-TTS_BUSY_FACTOR_MIN  = 1.0
-TTS_BUSY_FACTOR_MAX  = 2.5
-TTS_BUSY_FACTOR_UP   = 1.20
-TTS_BUSY_FACTOR_DOWN = 0.97
+_ROBOT = load_config("robot")
+_TTS = _ROBOT["tts"]
+_ARM = _ROBOT["arm"]
 
-REJECT_ACTION  = "reject"
-RELEASE_ACTION = "release arm"
+TTS_VOLUME           = _TTS["volume"]
+TTS_SECONDS_PER_CHAR = _TTS["seconds_per_char"]
+TTS_MIN_SECONDS      = _TTS["min_seconds"]
+TTS_QUEUE_MAX        = _TTS["queue_max"]
+TTS_BUSY_FACTOR_MIN  = _TTS["busy_factor"]["min"]
+TTS_BUSY_FACTOR_MAX  = _TTS["busy_factor"]["max"]
+TTS_BUSY_FACTOR_UP   = _TTS["busy_factor"]["up"]
+TTS_BUSY_FACTOR_DOWN = _TTS["busy_factor"]["down"]
+TTS_MODE             = _TTS.get("mode", "tts_only")   # tts_only | recorded_or_tts | recorded_only
+
+REJECT_ACTION  = _ARM["reject_action"]
+RELEASE_ACTION = _ARM["release_action"]
+
+QueueItem = Tuple[str, Optional[str], Optional[str]]   # (text, category, key)
 
 
 def _ts() -> str:
@@ -31,11 +41,7 @@ def _ts() -> str:
 
 
 class RobotController:
-    """Owns both the G1 arm action client and the G1 audio (TTS) client.
-
-    A single ``ChannelFactoryInitialize`` call is shared by both clients and
-    the optional ``rt/lowstate`` subscriber used by the wireless-remote loop.
-    """
+    """Owns both the G1 arm action client and the G1 audio (TTS + PlayStream) client."""
 
     def __init__(self, iface: Optional[str], timeout: float, dry_run: bool,
                  tts_speaker_id: int, want_lowstate: bool = True):
@@ -46,8 +52,9 @@ class RobotController:
         self._action_map = None
         self.hub = None
         self._lowstate_sub = None
+        self._player = None     # AudioPlayer, lazily initialised
 
-        self._tts_queue: Deque[str] = collections.deque(maxlen=TTS_QUEUE_MAX)
+        self._tts_queue: Deque[QueueItem] = collections.deque(maxlen=TTS_QUEUE_MAX)
         self._tts_event = threading.Event()
         self._tts_worker_stop = threading.Event()
         self._tts_worker_thread: Optional[threading.Thread] = None
@@ -57,7 +64,7 @@ class RobotController:
         self._tts_busy_count: int = 0
 
         if dry_run:
-            print("[BRIDGE] DRY RUN — G1 SDK will not be loaded.", flush=True)
+            print(f"[BRIDGE {_ts()}] DRY RUN — G1 SDK will not be loaded.", flush=True)
             return
 
         from unitree_sdk2py.core.channel import ChannelFactoryInitialize
@@ -77,7 +84,7 @@ class RobotController:
         self.arm_client = G1ArmActionClient()
         self.arm_client.SetTimeout(timeout)
         self.arm_client.Init()
-        print(f"[BRIDGE] G1ArmActionClient ready (iface={iface or 'default'})",
+        print(f"[BRIDGE {_ts()}] G1ArmActionClient ready (iface={iface or 'default'})",
               flush=True)
 
         self.audio_client = AudioClient()
@@ -86,9 +93,13 @@ class RobotController:
         try:
             self.audio_client.SetVolume(TTS_VOLUME)
         except Exception as e:
-            print(f"[BRIDGE][WARN] AudioClient.SetVolume failed: {e}", flush=True)
-        print(f"[BRIDGE] G1 AudioClient ready (speaker_id={tts_speaker_id})",
-              flush=True)
+            print(f"[BRIDGE {_ts()}][WARN] AudioClient.SetVolume failed: {e}", flush=True)
+        print(f"[BRIDGE {_ts()}] G1 AudioClient ready (speaker_id={tts_speaker_id}, "
+              f"tts_mode={TTS_MODE})", flush=True)
+
+        # Pre-recorded clip library (WAVs under assets/audio/).
+        from robot.audio_player import AudioPlayer
+        self._player = AudioPlayer(self.audio_client)
 
         self._tts_worker_thread = threading.Thread(
             target=self._tts_worker_loop, name="TtsWorker", daemon=True,
@@ -99,35 +110,40 @@ class RobotController:
             try:
                 from unitree_sdk2py.core.channel import ChannelSubscriber
                 from unitree_sdk2py.idl.unitree_hg.msg.dds_ import LowState_
-                from saqr.robot.controller import LowStateHub
+                from robot.controller import LowStateHub
 
                 self.hub = LowStateHub(watchdog_timeout=0.25)
                 self._lowstate_sub = ChannelSubscriber("rt/lowstate", LowState_)
                 self._lowstate_sub.Init(self.hub.handler, 10)
-                print("[BRIDGE] Subscribed to rt/lowstate (wireless remote)",
+                print(f"[BRIDGE {_ts()}] Subscribed to rt/lowstate (wireless remote)",
                       flush=True)
             except Exception as e:
-                print(f"[BRIDGE][WARN] LowState subscribe failed: {e}", flush=True)
-                print("[BRIDGE][WARN] Trigger keys (R2+X / R2+Y) will not work.",
+                print(f"[BRIDGE {_ts()}][WARN] LowState subscribe failed: {e}", flush=True)
+                print(f"[BRIDGE {_ts()}][WARN] Trigger keys (R2+X / R2+Y) will not work.",
                       flush=True)
                 self.hub = None
 
-    # ── TTS ─────────────────────────────────────────────────────────────────
-    def _estimate_tts_seconds(self, text: str) -> float:
-        base = max(TTS_MIN_SECONDS, len(text) * TTS_SECONDS_PER_CHAR)
-        return base * self._tts_busy_factor
+    # ── Public API ──────────────────────────────────────────────────────────
+    def speak(self, text: str,
+              category: Optional[str] = None, key: Optional[str] = None):
+        """Non-blocking — enqueue an announcement for the worker thread.
 
-    def speak(self, text: str):
-        """Non-blocking — enqueue the phrase for the worker thread."""
+        If ``category`` + ``key`` are provided and a matching WAV exists under
+        ``assets/audio/<category>/<key>.wav``, the worker plays the WAV via
+        PlayStream (no firmware TTS latency). Otherwise falls back per
+        ``tts.mode`` setting.
+        """
         if self.dry_run:
-            print(f"[BRIDGE] (dry) would TtsMaker({text!r}, "
-                  f"speaker_id={self.tts_speaker_id})", flush=True)
+            print(f"[BRIDGE {_ts()}] (dry) would speak({text!r}, category={category!r}, "
+                  f"key={key!r})", flush=True)
             return
         if self.audio_client is None:
             return
-        if self._tts_queue and self._tts_queue[-1] == text:
+        item: QueueItem = (text, category, key)
+        # Drop adjacent duplicates (same text + routing).
+        if self._tts_queue and self._tts_queue[-1] == item:
             return
-        self._tts_queue.append(text)
+        self._tts_queue.append(item)
         self._tts_event.set()
 
     def shutdown_tts(self):
@@ -136,6 +152,7 @@ class RobotController:
         if self._tts_worker_thread is not None:
             self._tts_worker_thread.join(timeout=1.0)
 
+    # ── Worker thread ───────────────────────────────────────────────────────
     def _tts_worker_loop(self):
         while not self._tts_worker_stop.is_set():
             if not self._tts_queue:
@@ -143,10 +160,49 @@ class RobotController:
                 self._tts_event.clear()
                 continue
             try:
-                text = self._tts_queue.popleft()
+                item = self._tts_queue.popleft()
             except IndexError:
                 continue
-            self._speak_blocking(text)
+            self._dispatch(*item)
+
+    def _dispatch(self, text: str, category: Optional[str], key: Optional[str]):
+        """Route one queue item to PlayStream or TtsMaker per ``tts.mode``."""
+        wants_clip = (
+            TTS_MODE in ("recorded_or_tts", "recorded_only")
+            and category is not None and key is not None
+            and self._player is not None
+            and self._player.has(category, key)
+        )
+        if wants_clip:
+            print(f"[BRIDGE {_ts()}] play -> {category}/{key!r}  "
+                  f"(text={text!r})", flush=True)
+            call_t0 = time.monotonic()
+            ok = self._player.play(category, key)
+            dt = time.monotonic() - call_t0
+            if ok:
+                print(f"[BRIDGE {_ts()}] play done ({dt*1000:.0f} ms)", flush=True)
+                return
+            # Play failed. Decide by mode whether to fall back to TtsMaker.
+            if TTS_MODE == "recorded_only":
+                print(f"[BRIDGE {_ts()}][WARN] play failed and tts.mode=recorded_only "
+                      f"— dropping phrase silently", flush=True)
+                return
+            print(f"[BRIDGE {_ts()}][WARN] play failed; falling back to TtsMaker",
+                  flush=True)
+            # fall through to TtsMaker
+
+        if TTS_MODE == "recorded_only":
+            # No clip exists for this phrase and user opted out of TtsMaker.
+            print(f"[BRIDGE {_ts()}] skip (recorded_only, no clip for "
+                  f"{category}/{key!r}): {text!r}", flush=True)
+            return
+
+        self._speak_blocking(text)
+
+    # ── TtsMaker path (fallback + legacy) ───────────────────────────────────
+    def _estimate_tts_seconds(self, text: str) -> float:
+        base = max(TTS_MIN_SECONDS, len(text) * TTS_SECONDS_PER_CHAR)
+        return base * self._tts_busy_factor
 
     def _speak_blocking(self, text: str):
         if self.audio_client is None:
@@ -199,18 +255,18 @@ class RobotController:
     # ── Arm ─────────────────────────────────────────────────────────────────
     def reject(self, release_after: float):
         if self.dry_run:
-            print(f"[BRIDGE] (dry) would run '{REJECT_ACTION}' "
+            print(f"[BRIDGE {_ts()}] (dry) would run '{REJECT_ACTION}' "
                   f"then release after {release_after:.1f}s", flush=True)
             return
         if self.arm_client is None or self._action_map is None:
             return
         if REJECT_ACTION not in self._action_map:
-            print(f"[BRIDGE][ERR] '{REJECT_ACTION}' not in SDK action_map",
+            print(f"[BRIDGE {_ts()}][ERR] '{REJECT_ACTION}' not in SDK action_map",
                   flush=True)
             return
-        print(f"[BRIDGE] -> {REJECT_ACTION}", flush=True)
+        print(f"[BRIDGE {_ts()}] -> {REJECT_ACTION}", flush=True)
         self.arm_client.ExecuteAction(self._action_map[REJECT_ACTION])
         if release_after > 0:
             time.sleep(release_after)
-            print(f"[BRIDGE] -> {RELEASE_ACTION}", flush=True)
+            print(f"[BRIDGE {_ts()}] -> {RELEASE_ACTION}", flush=True)
             self.arm_client.ExecuteAction(self._action_map[RELEASE_ACTION])
diff --git a/saqr/__init__.py b/saqr/__init__.py
deleted file mode 100644
index 200f45d..0000000
--- a/saqr/__init__.py
+++ /dev/null
@@ -1,3 +0,0 @@
-"""Saqr — PPE safety detection and G1 humanoid integration."""
-
-__version__ = "0.1.0"
diff --git a/saqr/core/camera.py b/saqr/core/camera.py
deleted file mode 100644
index 5ed53bd..0000000
--- a/saqr/core/camera.py
+++ /dev/null
@@ -1,88 +0,0 @@
-"""Camera backends: RealSense SDK and OpenCV V4L2."""
-from __future__ import annotations
-
-from typing import Optional
-
-import cv2
-import numpy as np
-
-from saqr.utils.logger import get_logger
-
-log = get_logger("Inference", "camera")
-
-try:
-    import pyrealsense2 as rs
-    HAS_REALSENSE = True
-except ImportError:
-    HAS_REALSENSE = False
-
-
-class RealSenseCapture:
-    """pyrealsense2 pipeline with an OpenCV-like read() interface."""
-
-    def __init__(self, width: int = 640, height: int = 480, fps: int = 30,
-                 serial: Optional[str] = None):
-        if not HAS_REALSENSE:
-            raise RuntimeError("pyrealsense2 not installed")
-        self.pipeline = rs.pipeline()
-        cfg = rs.config()
-        if serial:
-            cfg.enable_device(serial)
-        cfg.enable_stream(rs.stream.color, width, height, rs.format.bgr8, fps)
-        self.profile = self.pipeline.start(cfg)
-        self._open = True
-        dev = self.profile.get_device()
-        log.info(f"RealSense opened | {dev.get_info(rs.camera_info.name)} "
-                 f"serial={dev.get_info(rs.camera_info.serial_number)} "
-                 f"{width}x{height}@{fps}")
-
-    def isOpened(self) -> bool:
-        return self._open
-
-    def read(self):
-        if not self._open:
-            return False, None
-        try:
-            frames = self.pipeline.wait_for_frames(timeout_ms=3000)
-            color = frames.get_color_frame()
-            if not color:
-                return False, None
-            return True, np.asanyarray(color.get_data())
-        except Exception:
-            return False, None
-
-    def release(self):
-        if self._open:
-            self.pipeline.stop()
-            self._open = False
-
-
-def open_capture(source: str):
-    if source.lower().startswith("realsense"):
-        serial = None
-        if ":" in source:
-            serial = source.split(":", 1)[1]
-        return RealSenseCapture(width=640, height=480, fps=30, serial=serial)
-
-    if str(source).isdigit():
-        idx = int(source)
-        cap = cv2.VideoCapture(idx)
-        if cap.isOpened():
-            return cap
-        cap = cv2.VideoCapture(idx, cv2.CAP_ANY)
-        if cap.isOpened():
-            return cap
-        cap = cv2.VideoCapture(idx, cv2.CAP_V4L2)
-        return cap
-
-    if source.startswith("/dev/video"):
-        cap = cv2.VideoCapture(source, cv2.CAP_V4L2)
-        if cap.isOpened():
-            cap.set(cv2.CAP_PROP_BUFFERSIZE, 1)
-            cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc(*"MJPG"))
-            cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
-            cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)
-            cap.set(cv2.CAP_PROP_FPS, 30)
-        return cap
-
-    return cv2.VideoCapture(source)
diff --git a/saqr/core/capture.py b/saqr/core/capture.py
deleted file mode 100644
index f3c6a64..0000000
--- a/saqr/core/capture.py
+++ /dev/null
@@ -1,41 +0,0 @@
-"""Per-track image cropping + capture directory setup."""
-from __future__ import annotations
-
-from pathlib import Path
-from typing import Dict, Optional
-
-import cv2
-
-from saqr.core.detection import STATUSES
-from saqr.core.geometry import clamp_bbox
-from saqr.core.paths import CAPTURES_DIR
-
-
-def setup_capture_dirs() -> Dict[str, Path]:
-    dirs: Dict[str, Path] = {}
-    for s in STATUSES:
-        d = CAPTURES_DIR / s
-        d.mkdir(parents=True, exist_ok=True)
-        dirs[s] = d
-    return dirs
-
-
-def save_track_image(frame, track, capture_dirs: Dict[str, Path]) -> Optional[Path]:
-    h, w = frame.shape[:2]
-    x1, y1, x2, y2 = clamp_bbox(track.bbox, w, h)
-    if x2 <= x1 or y2 <= y1:
-        return None
-    crop = frame[y1:y2, x1:x2]
-    if crop.size == 0:
-        return None
-
-    target = capture_dirs[track.status] / f"track_{track.track_id:04d}.jpg"
-    if track.photo_path and track.photo_path != target and track.photo_path.exists():
-        try:
-            track.photo_path.unlink()
-        except OSError:
-            pass
-
-    cv2.imwrite(str(target), crop)
-    track.photo_path = target
-    return target
diff --git a/saqr/core/compliance.py b/saqr/core/compliance.py
deleted file mode 100644
index 6d82560..0000000
--- a/saqr/core/compliance.py
+++ /dev/null
@@ -1,36 +0,0 @@
-"""SAFE / PARTIAL / UNSAFE classification (helmet + vest focus)."""
-from __future__ import annotations
-
-from typing import Dict, List, Tuple
-
-from saqr.core.detection import POSITIVE_TO_NEGATIVE, PPE_DISPLAY_ORDER
-
-
-def status_from_items(items: Dict[str, float]) -> str:
-    has_helmet = items.get("helmet", 0.0) > items.get("no-helmet", 0.0) and items.get("helmet", 0.0) > 0
-    has_vest   = items.get("vest", 0.0) > items.get("no-vest", 0.0) and items.get("vest", 0.0) > 0
-    no_helmet  = items.get("no-helmet", 0.0) > 0
-    no_vest    = items.get("no-vest", 0.0) > 0
-
-    if no_helmet or no_vest:
-        return "UNSAFE"
-    if has_helmet and has_vest:
-        return "SAFE"
-    if has_helmet or has_vest:
-        return "PARTIAL"
-    return "UNSAFE"
-
-
-def split_wearing_missing(items: Dict[str, float]) -> Tuple[List[str], List[str], List[str]]:
-    wearing, missing, unknown = [], [], []
-    for pos in PPE_DISPLAY_ORDER:
-        neg = POSITIVE_TO_NEGATIVE[pos]
-        pos_conf = items.get(pos, 0.0)
-        neg_conf = items.get(neg, 0.0)
-        if pos_conf > neg_conf and pos_conf > 0:
-            wearing.append(pos)
-        elif neg_conf >= pos_conf and neg_conf > 0:
-            missing.append(pos)
-        else:
-            unknown.append(pos)
-    return wearing, missing, unknown
diff --git a/saqr/core/paths.py b/saqr/core/paths.py
deleted file mode 100644
index 7cb3a4f..0000000
--- a/saqr/core/paths.py
+++ /dev/null
@@ -1,19 +0,0 @@
-"""Canonical project paths, resolved from the saqr package location."""
-from __future__ import annotations
-
-from pathlib import Path
-
-PROJECT_ROOT = Path(__file__).resolve().parents[2]
-
-CONFIG_DIR   = PROJECT_ROOT / "config"
-DATA_DIR     = PROJECT_ROOT / "data"
-DATASET_DIR  = DATA_DIR / "dataset"
-MODELS_DIR   = DATA_DIR / "models"
-
-RUNTIME_DIR  = PROJECT_ROOT / "runtime"
-CAPTURES_DIR = RUNTIME_DIR / "captures"
-LOGS_DIR     = RUNTIME_DIR / "logs"
-RUNS_DIR     = RUNTIME_DIR / "runs"
-
-RESULT_CSV   = CAPTURES_DIR / "result.csv"
-EVENTS_CSV   = CAPTURES_DIR / "events.csv"
diff --git a/saqr/core/pipeline.py b/saqr/core/pipeline.py
deleted file mode 100644
index bba3e81..0000000
--- a/saqr/core/pipeline.py
+++ /dev/null
@@ -1,40 +0,0 @@
-"""Per-frame detect + group + track + capture + emit pipeline."""
-from __future__ import annotations
-
-from pathlib import Path
-from typing import Dict
-
-from saqr.core.capture import save_track_image
-from saqr.core.detection import collect_detections
-from saqr.core.drawing import draw_track
-from saqr.core.events import emit_event, write_result_csv
-from saqr.core.grouping import group_detections_to_people
-from saqr.core.paths import RESULT_CSV
-from saqr.core.tracking import PersonTracker
-
-
-def process_frame(frame, model, tracker: PersonTracker, frame_idx: int, conf: float,
-                  capture_dirs: Dict[str, Path], write_csv: bool = True):
-    annotated = frame.copy()
-    h, w = annotated.shape[:2]
-
-    detections = collect_detections(frame, model, conf)
-    candidates = group_detections_to_people(detections, w, h)
-    created, changed = tracker.update(candidates, frame_idx)
-    visible = tracker.visible_tracks()
-
-    created_ids = {t.track_id for t in created}
-    changed_ids = {t.track_id for t in changed}
-    event_ids = created_ids | changed_ids
-
-    for track in visible:
-        save_track_image(frame, track, capture_dirs)
-        if track.track_id in event_ids:
-            ev_type = "NEW" if track.track_id in created_ids else "STATUS_CHANGE"
-            emit_event(track, tracker.event_logger, ev_type)
-        draw_track(annotated, track)
-
-    if write_csv:
-        write_result_csv(list(tracker.tracks.values()), RESULT_CSV)
-
-    return annotated, visible
diff --git a/scripts/deploy.sh b/scripts/deploy.sh
index c6786b4..ed92f9a 100755
--- a/scripts/deploy.sh
+++ b/scripts/deploy.sh
@@ -15,11 +15,22 @@ set -e
 HERE="$(cd "$(dirname "$0")" && pwd)"
 PROJECT_ROOT="$(cd "$HERE/.." && pwd)"
 
-ROBOT_IP="${ROBOT_IP:-192.168.123.164}"
-ROBOT_USER="${ROBOT_USER:-unitree}"
-ROBOT_ENV="${ROBOT_ENV:-marcus}"
-REMOTE_DIR="/home/${ROBOT_USER}/Saqr"
-SSH_OPTS="-o StrictHostKeyChecking=no -o ConnectTimeout=10"
+config_get() {
+    python3 -c "
+import json, os, sys
+with open('$PROJECT_ROOT/config/robot_config.json') as f:
+    c = json.load(f)
+for k in sys.argv[1].split('.'):
+    c = c[k]
+print(os.path.expandvars(str(c)))
+" "$1"
+}
+
+ROBOT_IP="${ROBOT_IP:-$(config_get deploy.robot_ip)}"
+ROBOT_USER="${ROBOT_USER:-$(config_get deploy.robot_user)}"
+ROBOT_ENV="${ROBOT_ENV:-$(config_get deploy.robot_env)}"
+REMOTE_DIR="${REMOTE_DIR:-$(config_get deploy.remote_dir)}"
+SSH_OPTS="${SSH_OPTS:-$(config_get deploy.ssh_opts)}"
 
 RUN_AFTER=false
 
@@ -50,7 +61,10 @@ ssh ${SSH_OPTS} ${ROBOT_USER}@${ROBOT_IP} "echo 'Connected OK'" || {
 
 echo "[2/4] Creating remote directory..."
 ssh ${SSH_OPTS} ${ROBOT_USER}@${ROBOT_IP} \
-    "mkdir -p ${REMOTE_DIR}/runtime/{captures/{SAFE,PARTIAL,UNSAFE},logs,runs} ${REMOTE_DIR}/data/models"
+    "mkdir -p ${REMOTE_DIR}/runtime/captures/{SAFE,PARTIAL,UNSAFE} \
+             ${REMOTE_DIR}/runtime/runs \
+             ${REMOTE_DIR}/logs \
+             ${REMOTE_DIR}/data/models"
 
 echo "[3/4] Rsyncing package..."
 RSYNC_OPTS=(
@@ -60,12 +74,16 @@ RSYNC_OPTS=(
     --exclude="*.egg-info" --exclude="build/" --exclude="dist/"
 )
 rsync "${RSYNC_OPTS[@]}" \
-    "${PROJECT_ROOT}/saqr" \
+    "${PROJECT_ROOT}/core" \
+    "${PROJECT_ROOT}/apps" \
+    "${PROJECT_ROOT}/gui" \
+    "${PROJECT_ROOT}/robot" \
+    "${PROJECT_ROOT}/utils" \
     "${PROJECT_ROOT}/scripts" \
     "${PROJECT_ROOT}/config" \
     "${PROJECT_ROOT}/docs" \
+    "${PROJECT_ROOT}/assets" \
     "${PROJECT_ROOT}/pyproject.toml" \
-    "${PROJECT_ROOT}/requirements.txt" \
     "${PROJECT_ROOT}/README.md" \
     ${ROBOT_USER}@${ROBOT_IP}:${REMOTE_DIR}/
 
@@ -84,8 +102,8 @@ ssh ${SSH_OPTS} ${ROBOT_USER}@${ROBOT_IP} bash -s <<INSTALL_EOF
     source ~/miniconda3/etc/profile.d/conda.sh
     conda activate ${ROBOT_ENV}
     cd ${REMOTE_DIR}
-    pip install -q -e . 2>/dev/null || pip install -e .
-    chmod +x scripts/start_saqr.sh scripts/run_robot.sh scripts/run_local.sh 2>/dev/null || true
+    pip install -e .
+    chmod +x scripts/start_saqr.sh
     echo "  Install OK"
 INSTALL_EOF
 
diff --git a/scripts/generate_phrases.py b/scripts/generate_phrases.py
new file mode 100644
index 0000000..07592f1
--- /dev/null
+++ b/scripts/generate_phrases.py
@@ -0,0 +1,178 @@
+#!/usr/bin/env python3
+"""
+generate_phrases.py — synthesize Saqr's 8 phrases to WAV using piper-tts.
+
+This produces the audio library locally (no G1 mic, no TtsMaker capture,
+no PulseAudio) at exactly the format robot.audio_player expects:
+16 kHz mono int16 WAV under assets/audio/<category>/<key>.wav.
+
+Setup (once):
+    pip install piper-tts
+    python -m piper.download_voices en_US-amy-medium
+
+Usage:
+    python scripts/generate_phrases.py
+    python scripts/generate_phrases.py --voice en_US-lessac-medium
+    python scripts/generate_phrases.py --only safe,helmet_vest
+    python scripts/generate_phrases.py --voice-dir /custom/path
+
+Common voices (run the download command above, swap the name):
+    en_US-amy-medium      female, neutral         (~60 MB)
+    en_US-lessac-medium   female, friendly        (~60 MB)
+    en_US-ryan-high       male, clear             (~120 MB)
+    en_GB-alan-medium     male, British           (~60 MB)
+"""
+from __future__ import annotations
+
+import argparse
+import io
+import sys
+import wave
+from pathlib import Path
+
+PROJECT_ROOT = Path(__file__).resolve().parent.parent
+ASSETS_ROOT  = PROJECT_ROOT / "assets" / "audio"
+
+PHRASES = [
+    ("fixed",          "ready",
+     "Saqr is running. Press R2 plus X to start."),
+    ("fixed",          "deactivated",
+     "Saqr deactivated."),
+    ("fixed",          "no_camera",
+     "Camera not connected. Please plug in the camera and try again."),
+    ("fixed",          "safe",
+     "Safe to enter. Have a good day."),
+    ("fixed",          "unsafe_generic",
+     "Please stop. Wear your proper safety equipment."),
+    ("unsafe_missing", "helmet",
+     "Please stop. Wear your proper safety equipment. You are missing helmet."),
+    ("unsafe_missing", "vest",
+     "Please stop. Wear your proper safety equipment. You are missing vest."),
+    ("unsafe_missing", "helmet_vest",
+     "Please stop. Wear your proper safety equipment. You are missing helmet and vest."),
+]
+
+
+def find_voice_files(voice_name: str, override_dir: Path = None):
+    """Locate <voice_name>.onnx + .onnx.json in piper's standard dirs."""
+    search_dirs = []
+    if override_dir is not None:
+        search_dirs.append(Path(override_dir))
+    search_dirs += [
+        Path.home() / ".local" / "share" / "piper-voices",
+        Path.home() / ".cache" / "piper" / "voices",
+        Path.cwd(),
+    ]
+    for d in search_dirs:
+        onnx = d / f"{voice_name}.onnx"
+        js   = d / f"{voice_name}.onnx.json"
+        if onnx.exists() and js.exists():
+            return onnx, js
+    return None, None
+
+
+def resample_to_16k_mono_int16(pcm: bytes, src_rate: int, src_channels: int) -> bytes:
+    """Linear-resample int16 PCM to 16 kHz mono."""
+    import numpy as np
+    a = np.frombuffer(pcm, dtype=np.int16)
+    if src_channels > 1:
+        a = a.reshape(-1, src_channels).mean(axis=1).astype(np.int16)
+    if src_rate == 16000:
+        return a.tobytes()
+    target_len = int(round(len(a) * 16000 / src_rate))
+    idx = np.linspace(0, len(a) - 1, target_len).astype(np.int64)
+    return a[idx].astype(np.int16).tobytes()
+
+
+def save_wav_16k_mono(pcm: bytes, path: Path) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with wave.open(str(path), "wb") as wf:
+        wf.setnchannels(1)
+        wf.setsampwidth(2)
+        wf.setframerate(16000)
+        wf.writeframes(pcm)
+
+
+def main():
+    ap = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    ap.add_argument("--voice",      default="en_US-amy-medium",
+                    help="piper voice name (default en_US-amy-medium)")
+    ap.add_argument("--voice-dir",  default=None,
+                    help="override directory to search for the voice .onnx files")
+    ap.add_argument("--only",       default=None,
+                    help="comma-separated keys to generate (e.g. safe,helmet)")
+    ap.add_argument("--length-scale", type=float, default=None,
+                    help="speaking rate (piper default ~1.0; 0.8 faster, 1.2 slower)")
+    args = ap.parse_args()
+
+    only = set(args.only.split(",")) if args.only else None
+    targets = [(c, k, t) for (c, k, t) in PHRASES if only is None or k in only]
+    if not targets:
+        print(f"No phrases match --only={args.only!r}.")
+        sys.exit(1)
+
+    # Import piper
+    try:
+        from piper.voice import PiperVoice
+    except ImportError:
+        print("ERROR: piper-tts is not installed in this Python environment.")
+        print()
+        print("Install it:")
+        print("  pip install piper-tts")
+        print(f"  python -m piper.download_voices {args.voice}")
+        sys.exit(1)
+
+    onnx, js = find_voice_files(args.voice, args.voice_dir)
+    if onnx is None:
+        print(f"ERROR: voice files for {args.voice!r} not found.")
+        print()
+        print("Download:")
+        print(f"  python -m piper.download_voices {args.voice}")
+        print()
+        print("Or pass --voice-dir pointing at a directory that contains")
+        print(f"  {args.voice}.onnx + {args.voice}.onnx.json")
+        sys.exit(1)
+
+    print(f"Loading voice: {onnx}")
+    voice = PiperVoice.load(str(onnx), config_path=str(js))
+    src_rate = int(voice.config.sample_rate)
+    print(f"Native rate: {src_rate} Hz")
+    print(f"Target:      16000 Hz mono int16 under {ASSETS_ROOT}")
+    print()
+
+    synth_kwargs = {}
+    if args.length_scale is not None:
+        synth_kwargs["length_scale"] = args.length_scale
+
+    for cat, key, text in targets:
+        out = ASSETS_ROOT / cat / f"{key}.wav"
+        print(f"[{cat}/{key}]")
+        print(f"   text : {text}")
+
+        # Synthesize into an in-memory WAV buffer.
+        buf = io.BytesIO()
+        with wave.open(buf, "wb") as wf:
+            voice.synthesize(text, wf, **synth_kwargs)
+        buf.seek(0)
+        with wave.open(buf, "rb") as wf:
+            pcm       = wf.readframes(wf.getnframes())
+            rate_in   = wf.getframerate()
+            channels  = wf.getnchannels()
+
+        resampled = resample_to_16k_mono_int16(pcm, rate_in, channels)
+        save_wav_16k_mono(resampled, out)
+
+        dur = len(resampled) / 2.0 / 16000
+        print(f"   saved → {out.relative_to(PROJECT_ROOT)}   ({dur:.1f}s)")
+        print()
+
+    print("Done. Deploy + restart:")
+    print("   scripts/deploy.sh")
+    print("   ssh unitree@192.168.123.164 'sudo systemctl restart saqr-bridge && "
+          "journalctl -u saqr-bridge -n 20 | grep audio_player'")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/import_voices.py b/scripts/import_voices.py
new file mode 100644
index 0000000..0262ec5
--- /dev/null
+++ b/scripts/import_voices.py
@@ -0,0 +1,206 @@
+#!/usr/bin/env python3
+"""
+import_voices.py — Convert TTS MP3s to Saqr's audio library format.
+
+Reads MP3s from a source directory, maps each to its correct
+``assets/audio/<category>/<key>.wav`` path, and transcodes to
+16 kHz mono int16 WAV using ffmpeg.
+
+Matching rules:
+  1. If the filename (without extension, trailing dots stripped) exactly
+     equals one of the known phrases, that's a direct map.
+  2. Any file that doesn't match — typical case is ``ttsmaker-file-*.mp3``
+     where the site didn't rename it — is assigned to whichever phrase
+     key is still missing. If exactly ONE phrase is unmapped and ONE
+     file is unmatched, they're paired. Otherwise you're prompted.
+
+Requires ffmpeg on PATH:  sudo apt install -y ffmpeg
+
+Usage:
+    python scripts/import_voices.py
+    python scripts/import_voices.py --src "/home/zedx/Downloads/voices saqr"
+    python scripts/import_voices.py --dry-run       # just show mapping
+"""
+from __future__ import annotations
+
+import argparse
+import shutil
+import subprocess
+import sys
+from pathlib import Path
+
+PROJECT_ROOT  = Path(__file__).resolve().parent.parent
+ASSETS_ROOT   = PROJECT_ROOT / "assets" / "audio"
+DEFAULT_SRC   = Path("/home/zedx/Downloads/voices saqr")
+
+# (category, key, text) — filename match uses the text (with or without
+# trailing period).
+PHRASES = [
+    ("fixed",          "ready",
+     "Saqr is running. Press R2 plus X to start."),
+    ("fixed",          "deactivated",
+     "Saqr deactivated."),
+    ("fixed",          "no_camera",
+     "Camera not connected. Please plug in the camera and try again."),
+    ("fixed",          "safe",
+     "Safe to enter. Have a good day."),
+    ("fixed",          "unsafe_generic",
+     "Please stop. Wear your proper safety equipment."),
+    ("unsafe_missing", "helmet",
+     "Please stop. Wear your proper safety equipment. You are missing helmet."),
+    ("unsafe_missing", "vest",
+     "Please stop. Wear your proper safety equipment. You are missing vest."),
+    ("unsafe_missing", "helmet_vest",
+     "Please stop. Wear your proper safety equipment. You are missing helmet and vest."),
+]
+
+# Text → (category, key)  and  normalised-text → (category, key)
+TEXT_TO_KEY = {text: (cat, key) for (cat, key, text) in PHRASES}
+
+
+def normalise(stem: str) -> str:
+    """Strip trailing dots so 'Saqr deactivated.' matches 'Saqr deactivated..mp3'."""
+    return stem.rstrip(".").strip()
+
+
+def match_by_filename(stem: str):
+    candidates = {stem, stem + ".", normalise(stem), normalise(stem) + "."}
+    for c in candidates:
+        if c in TEXT_TO_KEY:
+            return TEXT_TO_KEY[c]
+    return None
+
+
+def convert_to_wav(src: Path, dst: Path, dry_run: bool = False) -> None:
+    dst.parent.mkdir(parents=True, exist_ok=True)
+    cmd = [
+        "ffmpeg", "-y", "-loglevel", "error",
+        "-i", str(src),
+        "-ac", "1",            # mono
+        "-ar", "16000",        # 16 kHz
+        "-sample_fmt", "s16",  # 16-bit int
+        str(dst),
+    ]
+    if dry_run:
+        return
+    subprocess.run(cmd, check=True)
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__,
+                                 formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--src",     default=str(DEFAULT_SRC),
+                    help=f"source directory of MP3s (default: {DEFAULT_SRC!s})")
+    ap.add_argument("--dry-run", action="store_true",
+                    help="print the mapping but don't convert")
+    args = ap.parse_args()
+
+    if shutil.which("ffmpeg") is None and not args.dry_run:
+        print("ERROR: ffmpeg not on PATH.  Install it first:")
+        print("    sudo apt install -y ffmpeg")
+        sys.exit(1)
+
+    src_dir = Path(args.src)
+    if not src_dir.exists():
+        print(f"ERROR: source directory not found: {src_dir}")
+        sys.exit(1)
+
+    mp3s = sorted(src_dir.glob("*.mp3"))
+    if not mp3s:
+        print(f"No .mp3 files under {src_dir}")
+        sys.exit(1)
+
+    print(f"Source: {src_dir}")
+    print(f"Target: {ASSETS_ROOT}")
+    print(f"Files:  {len(mp3s)} mp3  (expected 8)")
+    print()
+
+    # Pass 1: match by filename.
+    assigned: dict[tuple, Path] = {}  # (cat,key) → src mp3
+    unmatched: list[Path] = []
+    for mp3 in mp3s:
+        key = match_by_filename(mp3.stem)
+        if key is None:
+            unmatched.append(mp3)
+        elif key in assigned:
+            print(f"  [WARN] duplicate match for {key}: keeping {assigned[key].name},"
+                  f" ignoring {mp3.name}")
+        else:
+            assigned[key] = mp3
+
+    # Pass 2: assign unmatched files to still-missing phrases.
+    missing_keys = [(cat, key) for (cat, key, _) in PHRASES if (cat, key) not in assigned]
+    if len(unmatched) == 1 and len(missing_keys) == 1:
+        only_key = missing_keys[0]
+        only_file = unmatched[0]
+        print(f"  [AUTO] assigning '{only_file.name}' → {only_key[0]}/{only_key[1]} "
+              f"(the only phrase still unmapped)")
+        assigned[only_key] = only_file
+        unmatched = []
+        missing_keys = []
+    elif unmatched:
+        # Prompt user
+        print(f"  [WARN] {len(unmatched)} unmatched file(s) and "
+              f"{len(missing_keys)} unassigned phrase(s). Manual mapping needed:")
+        for f in unmatched:
+            print(f"    unmatched: {f.name}")
+        for (cat, key) in missing_keys:
+            print(f"    missing:   {cat}/{key}")
+        print()
+        print("Resolve manually by renaming the source files to one of:")
+        for (_, _, text) in PHRASES:
+            print(f"    '{text}.mp3'")
+        sys.exit(1)
+
+    # Show the plan
+    print("Mapping:")
+    for (cat, key, _) in PHRASES:
+        src = assigned.get((cat, key))
+        if src is None:
+            print(f"  [MISSING] no source for {cat}/{key}")
+            continue
+        rel = src.name if len(src.name) <= 60 else src.name[:57] + "..."
+        print(f"  {rel:<60s} → assets/audio/{cat}/{key}.wav")
+    print()
+
+    if args.dry_run:
+        print("(dry-run — nothing written)")
+        return
+
+    # Convert
+    print("Converting…")
+    converted = 0
+    for (cat, key, _) in PHRASES:
+        src = assigned.get((cat, key))
+        if src is None:
+            continue
+        dst = ASSETS_ROOT / cat / f"{key}.wav"
+        try:
+            convert_to_wav(src, dst)
+        except subprocess.CalledProcessError as e:
+            print(f"  [FAIL] {src.name} → {dst.name}: ffmpeg exited {e.returncode}")
+            continue
+
+        # Probe the result
+        import wave
+        with wave.open(str(dst), "rb") as wf:
+            ok = (wf.getframerate() == 16000
+                  and wf.getnchannels() == 1
+                  and wf.getsampwidth() == 2)
+            dur = wf.getnframes() / wf.getframerate()
+        mark = "✓" if ok else "⚠ wrong format"
+        print(f"  {mark}  {dst.relative_to(PROJECT_ROOT)}   ({dur:.1f}s)")
+        if ok:
+            converted += 1
+
+    print()
+    print(f"Done — {converted}/{len(PHRASES)} clip(s) converted.")
+    print()
+    print("Deploy + restart:")
+    print("  scripts/deploy.sh")
+    print("  ssh unitree@192.168.123.164 "
+          "'sudo systemctl restart saqr-bridge && journalctl -u saqr-bridge -n 20 | grep audio_player'")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/record_phrases.py b/scripts/record_phrases.py
new file mode 100755
index 0000000..9ed6021
--- /dev/null
+++ b/scripts/record_phrases.py
@@ -0,0 +1,408 @@
+#!/usr/bin/env python3
+"""
+record_phrases.py — Build Saqr's audio library for AudioClient.PlayStream.
+
+Two modes:
+
+  DEFAULT (TTS capture) — the script calls AudioClient.TtsMaker for each
+                         phrase and simultaneously records the G1 speaker
+                         output via the mic. Zero speaking required. Voice
+                         is the G1's own TTS voice, but at runtime
+                         PlayStream skips the firmware synthesis buffer so
+                         it plays back ~200–700 ms faster.
+
+  --mic (your voice)    — the script shows each phrase, counts down, and
+                         records whatever the mic hears. You speak each
+                         line yourself.
+
+Both modes save 16 kHz mono int16 WAVs under
+``assets/audio/<category>/<key>.wav`` — exactly what ``robot/audio_player``
+expects.
+
+Usage (run on the robot or any machine on the G1 subnet):
+
+    python3 scripts/record_phrases.py                      # TTS capture, all 8
+    python3 scripts/record_phrases.py --only safe,helmet   # just those two
+    python3 scripts/record_phrases.py --iface enp3s0       # custom DDS iface
+    python3 scripts/record_phrases.py --mic                # your voice instead
+    python3 scripts/record_phrases.py --play               # verify each clip by replaying
+
+Requires ``unitree_sdk2py`` in the active conda env (TTS-capture and --play).
+"""
+from __future__ import annotations
+
+import argparse
+import socket
+import struct
+import subprocess
+import sys
+import threading
+import time
+import wave
+from pathlib import Path
+
+PROJECT_ROOT = Path(__file__).resolve().parent.parent
+ASSETS_ROOT  = PROJECT_ROOT / "assets" / "audio"
+
+MCAST_GRP  = "239.168.123.161"
+MCAST_PORT = 5555
+RATE       = 16000
+
+# (category, key, text).
+PHRASES = [
+    ("fixed",          "ready",
+     "Saqr is running. Press R2 plus X to start."),
+    ("fixed",          "deactivated",
+     "Saqr deactivated."),
+    ("fixed",          "no_camera",
+     "Camera not connected. Please plug in the camera and try again."),
+    ("fixed",          "safe",
+     "Safe to enter. Have a good day."),
+    ("fixed",          "unsafe_generic",
+     "Please stop. Wear your proper safety equipment."),
+    ("unsafe_missing", "helmet",
+     "Please stop. Wear your proper safety equipment. You are missing helmet."),
+    ("unsafe_missing", "vest",
+     "Please stop. Wear your proper safety equipment. You are missing vest."),
+    ("unsafe_missing", "helmet_vest",
+     "Please stop. Wear your proper safety equipment. You are missing helmet and vest."),
+]
+
+
+# ── mic capture ─────────────────────────────────────────────────────────────
+def find_local_ip() -> str:
+    """Pick the first 192.168.123.x address on this machine."""
+    out = subprocess.run(
+        ["ip", "-4", "-o", "addr"], capture_output=True, text=True
+    ).stdout
+    for line in out.splitlines():
+        for tok in line.split():
+            if tok.startswith("192.168.123."):
+                return tok.split("/")[0]
+    raise RuntimeError(
+        "No 192.168.123.x address on this machine — connect to the G1 network first."
+    )
+
+
+def _open_mcast_socket(local_ip: str) -> socket.socket:
+    s = socket.socket(socket.AF_INET, socket.SOCK_DGRAM)
+    s.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+    s.bind(("", MCAST_PORT))
+    mreq = struct.pack("4s4s", socket.inet_aton(MCAST_GRP), socket.inet_aton(local_ip))
+    s.setsockopt(socket.IPPROTO_IP, socket.IP_ADD_MEMBERSHIP, mreq)
+    return s
+
+
+def record_multicast_fixed(seconds: float) -> bytes:
+    """Blocking mic capture for exactly ``seconds`` (used by --mic mode)."""
+    local_ip = find_local_ip()
+    sock = _open_mcast_socket(local_ip)
+    sock.settimeout(2.0)
+    target_bytes = int(RATE * 2 * seconds)
+    buf = bytearray()
+    t0 = time.time()
+    try:
+        while len(buf) < target_bytes and time.time() - t0 < seconds + 5:
+            try:
+                data, _ = sock.recvfrom(4096)
+                buf.extend(data)
+            except socket.timeout:
+                continue
+    finally:
+        sock.close()
+    return bytes(buf)
+
+
+def record_while_tts_plays(ac, text: str, speaker_id: int,
+                           capture_seconds: float) -> bytes:
+    """Start mic capture, call TtsMaker, capture for ``capture_seconds``, stop."""
+    local_ip = find_local_ip()
+    buf = bytearray()
+    stop_flag = threading.Event()
+
+    def _mic():
+        sock = _open_mcast_socket(local_ip)
+        sock.settimeout(0.5)
+        try:
+            while not stop_flag.is_set():
+                try:
+                    data, _ = sock.recvfrom(4096)
+                    buf.extend(data)
+                except socket.timeout:
+                    continue
+        finally:
+            sock.close()
+
+    t = threading.Thread(target=_mic, daemon=True)
+    t.start()
+    time.sleep(0.3)   # let the mic socket settle before TTS fires
+
+    try:
+        code = ac.TtsMaker(text, speaker_id)
+    except Exception as e:
+        print(f"   [FAIL] TtsMaker raised: {e}")
+        stop_flag.set()
+        t.join(timeout=1.0)
+        return b""
+    if code != 0:
+        print(f"   [FAIL] TtsMaker rc={code} — retry in 2s…")
+        time.sleep(2.0)
+        try:
+            code = ac.TtsMaker(text, speaker_id)
+        except Exception as e:
+            print(f"   [FAIL] TtsMaker retry raised: {e}")
+            stop_flag.set()
+            t.join(timeout=1.0)
+            return b""
+        if code != 0:
+            print(f"   [FAIL] TtsMaker retry rc={code}")
+            stop_flag.set()
+            t.join(timeout=1.0)
+            return b""
+
+    time.sleep(capture_seconds)
+    stop_flag.set()
+    t.join(timeout=1.0)
+    return bytes(buf)
+
+
+# ── processing ──────────────────────────────────────────────────────────────
+def rms_and_peak(pcm: bytes):
+    import numpy as np
+    a = np.frombuffer(pcm, dtype=np.int16)
+    if a.size == 0:
+        return 0.0, 0
+    rms = float(np.sqrt(np.mean(a.astype(np.float32) ** 2)))
+    peak = int(np.abs(a).max())
+    return rms, peak
+
+
+def trim_leading_silence(pcm: bytes, threshold: int = 500,
+                         head_ms: int = 80) -> bytes:
+    """Cut silence before the first sample above threshold; keep ``head_ms`` lead-in."""
+    import numpy as np
+    a = np.frombuffer(pcm, dtype=np.int16)
+    above = np.where(np.abs(a) > threshold)[0]
+    if above.size == 0:
+        return pcm
+    start = max(0, int(above[0]) - int(head_ms / 1000.0 * RATE))
+    return a[start:].tobytes()
+
+
+def trim_trailing_silence(pcm: bytes, threshold: int = 500,
+                          tail_ms: int = 150) -> bytes:
+    import numpy as np
+    a = np.frombuffer(pcm, dtype=np.int16)
+    above = np.where(np.abs(a) > threshold)[0]
+    if above.size == 0:
+        return pcm
+    end = int(above[-1]) + int(tail_ms / 1000.0 * RATE)
+    end = min(end, a.size)
+    return a[:end].tobytes()
+
+
+def save_wav(pcm: bytes, path: Path) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with wave.open(str(path), "wb") as wf:
+        wf.setnchannels(1)
+        wf.setsampwidth(2)
+        wf.setframerate(RATE)
+        wf.writeframes(pcm)
+
+
+def estimate_tts_duration(text: str) -> float:
+    """Match RobotController's pacing estimate (0.12 s/char, min 2.5 s)."""
+    return max(2.5, len(text) * 0.12)
+
+
+# ── optional playback verification ──────────────────────────────────────────
+def play_on_g1(ac, path: Path) -> None:
+    with wave.open(str(path), "rb") as wf:
+        pcm = wf.readframes(wf.getnframes())
+    CHUNK = 96000
+    sid = f"verify_{int(time.time() * 1000)}"
+    offset = 0
+    while offset < len(pcm):
+        chunk = pcm[offset:offset + CHUNK]
+        ac.PlayStream("saqr_verify", sid, chunk)
+        offset += len(chunk)
+        time.sleep(len(chunk) / (RATE * 2) / 2)
+    time.sleep(len(pcm) / (RATE * 2) + 0.3)
+    try:
+        ac.PlayStop("saqr_verify")
+    except Exception:
+        pass
+
+
+def countdown(seconds: int) -> None:
+    for i in range(seconds, 0, -1):
+        print(f"  starting in {i}...", end="\r", flush=True)
+        time.sleep(1)
+    print("  " + " " * 30, end="\r")
+
+
+# ── main ────────────────────────────────────────────────────────────────────
+def main():
+    ap = argparse.ArgumentParser(
+        description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    ap.add_argument("--mic",         action="store_true",
+                    help="record YOUR voice via the G1 mic instead of TtsMaker")
+    ap.add_argument("--duration",    type=float, default=None,
+                    help="seconds to record per phrase (default: estimate per phrase)")
+    ap.add_argument("--pause",       type=float, default=1.0,
+                    help="seconds of pause between phrases (default 1)")
+    ap.add_argument("--only",        default=None,
+                    help="comma-separated keys to record (e.g. 'safe,helmet,vest')")
+    ap.add_argument("--iface",       default="eth0",
+                    help="DDS iface for TtsMaker / PlayStream (default eth0)")
+    ap.add_argument("--speaker-id",  type=int, default=2,
+                    help="TtsMaker speaker_id (2 = English on current firmware)")
+    ap.add_argument("--volume",      type=int, default=100,
+                    help="G1 speaker volume 0–100 (default 100)")
+    ap.add_argument("--play",        action="store_true",
+                    help="play each saved clip back on the G1 via PlayStream")
+    ap.add_argument("--countdown",   type=int, default=2,
+                    help="countdown seconds before --mic recordings (default 2)")
+    ap.add_argument("--no-trim",     action="store_true",
+                    help="don't auto-trim leading/trailing silence")
+    ap.add_argument("--threshold",   type=int, default=500,
+                    help="amplitude threshold for silence trim (default 500)")
+    args = ap.parse_args()
+
+    only = set(args.only.split(",")) if args.only else None
+    targets = [(c, k, t) for (c, k, t) in PHRASES if only is None or k in only]
+    if not targets:
+        print(f"No phrases match --only={args.only!r}. Known keys: "
+              f"{', '.join(k for _, k, _ in PHRASES)}")
+        sys.exit(1)
+
+    try:
+        local_ip = find_local_ip()
+    except RuntimeError as e:
+        print(f"[FATAL] {e}")
+        sys.exit(1)
+
+    mode = "mic" if args.mic else "tts"
+
+    # Init AudioClient for TTS mode or --play.
+    ac = None
+    if mode == "tts" or args.play:
+        try:
+            from unitree_sdk2py.core.channel import ChannelFactoryInitialize
+            from unitree_sdk2py.g1.audio.g1_audio_client import AudioClient
+        except ImportError as e:
+            print(f"[FATAL] unitree_sdk2py not available in this env: {e}")
+            sys.exit(1)
+        print(f"[init] ChannelFactoryInitialize(0, {args.iface!r})")
+        ChannelFactoryInitialize(0, args.iface)
+        ac = AudioClient()
+        ac.SetTimeout(10.0)
+        ac.Init()
+        try:
+            ac.SetVolume(args.volume)
+        except Exception as e:
+            print(f"[init][WARN] SetVolume failed: {e}")
+        print(f"[init] AudioClient ready (speaker_id={args.speaker_id} volume={args.volume})")
+        print()
+
+    print("=" * 60)
+    print(f" Saqr phrase recorder — {len(targets)} clip(s)")
+    print(f" Mode:      {mode.upper()}  "
+          f"({'TtsMaker → mic' if mode == 'tts' else 'your voice → mic'})")
+    print(f" Duration:  {('auto (per phrase)' if args.duration is None else f'{args.duration}s')}")
+    print(f" Pause:     {args.pause}s between phrases")
+    print(f" Trim:      {'off' if args.no_trim else f'on (threshold={args.threshold})'}")
+    print(f" Output:    {ASSETS_ROOT}")
+    print(f" Mic feed:  {MCAST_GRP}:{MCAST_PORT} (local IP: {local_ip})")
+    print("=" * 60)
+    print()
+
+    if mode == "mic":
+        print("Stand within ~1 m of the G1. When you see '>>> SPEAK NOW <<<' say")
+        print("the SAY: line clearly. Target rms 500–5000.")
+    else:
+        print("The G1 will speak each phrase via TtsMaker while the script records")
+        print("through the mic. Stay quiet during capture; any room noise you hear")
+        print("will end up baked into the clip.")
+    print()
+    try:
+        input("Press Enter to start…")
+    except EOFError:
+        return
+    print()
+
+    saved = []
+    for i, (cat, key, text) in enumerate(targets, 1):
+        out = ASSETS_ROOT / cat / f"{key}.wav"
+        print(f"── [{i}/{len(targets)}  {cat}/{key}] ──")
+        print(f"   SAY: {text}")
+
+        if mode == "mic":
+            if args.countdown > 0:
+                countdown(args.countdown)
+            dur = args.duration if args.duration is not None else 5.0
+            print(f"   >>> SPEAK NOW — {dur}s <<<", flush=True)
+            pcm = record_multicast_fixed(dur)
+        else:
+            dur = args.duration if args.duration is not None else \
+                  estimate_tts_duration(text) + 1.5
+            print(f"   TtsMaker playing… (capturing {dur:.1f}s)", flush=True)
+            pcm = record_while_tts_plays(ac, text, args.speaker_id, dur)
+
+        if not pcm:
+            print(f"   [FAIL] no audio captured")
+            print()
+            continue
+
+        if not args.no_trim:
+            pcm = trim_leading_silence(pcm, threshold=args.threshold)
+            pcm = trim_trailing_silence(pcm, threshold=args.threshold)
+
+        rms, peak = rms_and_peak(pcm)
+        dur_s = len(pcm) / 2.0 / RATE
+        save_wav(pcm, out)
+
+        if rms < 400:
+            marker = f"  ⚠ TOO QUIET — re-record with --only {key}"
+        elif peak > 30000:
+            marker = "  ⚠ CLIPPING"
+        else:
+            marker = "  ✓"
+        print(f"   saved → {out.relative_to(PROJECT_ROOT)}  "
+              f"({dur_s:.1f}s  rms={rms:.0f}  peak={peak}){marker}")
+        saved.append((cat, key, out, dur_s, rms, peak))
+
+        if args.play and ac is not None:
+            print(f"   playing back…")
+            try:
+                play_on_g1(ac, out)
+            except Exception as e:
+                print(f"   [play] failed: {e}")
+
+        if i < len(targets) and args.pause > 0:
+            time.sleep(args.pause)
+        print()
+
+    # Summary
+    print("=" * 60)
+    print(f" Done — {len(saved)}/{len(targets)} clip(s) saved")
+    print("=" * 60)
+    bad = []
+    for cat, key, path, dur, rms, peak in saved:
+        warn = " ⚠ low level" if rms < 400 else (" ⚠ clipping" if peak > 30000 else "")
+        if warn:
+            bad.append(key)
+        print(f"  {cat}/{key:20s}  {dur:4.1f}s  rms={rms:5.0f}  peak={peak:5d}{warn}")
+    print()
+    if bad:
+        print(f"Re-record the flagged ones with:")
+        print(f"  python3 scripts/record_phrases.py --only {','.join(bad)}")
+        print()
+    print("Next:")
+    print("  1. (on robot) sudo systemctl restart saqr-bridge")
+    print("  2. expect:   [audio_player] loaded N clip(s): ...")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/scripts/run_local.sh b/scripts/run_local.sh
deleted file mode 100755
index 4393321..0000000
--- a/scripts/run_local.sh
+++ /dev/null
@@ -1,87 +0,0 @@
-#!/bin/bash
-# ============================================================================
-#  Saqr PPE Detection - Run on Local Laptop
-# ============================================================================
-#
-#  Usage:
-#    scripts/run_local.sh                          # webcam 0
-#    scripts/run_local.sh --source 1               # webcam 1
-#    scripts/run_local.sh --source video.mp4       # video file
-#    scripts/run_local.sh --gui                    # PySide6 GUI
-#    scripts/run_local.sh --detect                 # simple detection (no tracking)
-#
-# ============================================================================
-
-set -e
-
-HERE="$(cd "$(dirname "$0")" && pwd)"
-PROJECT_ROOT="$(cd "$HERE/.." && pwd)"
-cd "$PROJECT_ROOT"
-
-SOURCE="0"
-MODEL="saqr_best.pt"
-CONF="0.35"
-MODE="saqr"      # saqr | gui | detect
-HEADLESS=false
-MAX_MISSING=90
-MATCH_DIST=250
-CONFIRM=5
-
-while [[ $# -gt 0 ]]; do
-    case $1 in
-        --source)          SOURCE="$2"; shift 2 ;;
-        --model)           MODEL="$2"; shift 2 ;;
-        --conf)            CONF="$2"; shift 2 ;;
-        --gui)             MODE="gui"; shift ;;
-        --detect)          MODE="detect"; shift ;;
-        --headless)        HEADLESS=true; shift ;;
-        --max-missing)     MAX_MISSING="$2"; shift 2 ;;
-        --match-distance)  MATCH_DIST="$2"; shift 2 ;;
-        --confirm)         CONFIRM="$2"; shift 2 ;;
-        *)                 echo "Unknown arg: $1"; exit 1 ;;
-    esac
-done
-
-if command -v conda &>/dev/null; then
-    source "$(conda info --base)/etc/profile.d/conda.sh" 2>/dev/null || true
-    conda activate AI_MSI_yolo 2>/dev/null || true
-fi
-
-echo "============================================"
-echo "  Saqr PPE Detection - Local Laptop"
-echo "============================================"
-echo "  Mode   : $MODE"
-echo "  Source : $SOURCE"
-echo "  Model  : $MODEL"
-echo "  Conf   : $CONF"
-echo "============================================"
-echo ""
-
-HEADLESS_FLAG=""
-if [ "$HEADLESS" = true ]; then
-    HEADLESS_FLAG="--headless"
-fi
-
-case $MODE in
-    saqr)
-        python -m saqr.apps.saqr_cli \
-            --source "$SOURCE" \
-            --model "$MODEL" \
-            --conf "$CONF" \
-            --max-missing "$MAX_MISSING" \
-            --match-distance "$MATCH_DIST" \
-            --status-confirm-frames "$CONFIRM" \
-            $HEADLESS_FLAG
-        ;;
-    gui)
-        python -m saqr.gui.app \
-            --source "$SOURCE" \
-            --model "$MODEL"
-        ;;
-    detect)
-        python -m saqr.apps.detect_cli \
-            --source "$SOURCE" \
-            --model "$MODEL" \
-            --conf "$CONF"
-        ;;
-esac
diff --git a/scripts/run_robot.sh b/scripts/run_robot.sh
deleted file mode 100755
index 16083ab..0000000
--- a/scripts/run_robot.sh
+++ /dev/null
@@ -1,102 +0,0 @@
-#!/bin/bash
-# ============================================================================
-#  Saqr PPE Detection - Run on Unitree G1 Robot (no bridge, direct saqr run)
-# ============================================================================
-#
-#  Run on the robot's physical terminal (with monitor) or via ssh -X:
-#    scripts/run_robot.sh
-#    scripts/run_robot.sh --headless              # no display
-#    scripts/run_robot.sh --source /dev/video2    # V4L2 fallback
-#
-#  For the production R2+X / R2+Y workflow, use scripts/start_saqr.sh instead.
-# ============================================================================
-
-set -e
-
-HERE="$(cd "$(dirname "$0")" && pwd)"
-PROJECT_ROOT="$(cd "$HERE/.." && pwd)"
-cd "$PROJECT_ROOT"
-
-SOURCE="realsense"
-MODEL="saqr_best.pt"
-CONF="0.35"
-HEADLESS=false
-MAX_MISSING=120
-MATCH_DIST=300
-CONFIRM=7
-DEVICE="0"
-IMGSZ=320
-HALF=true
-STREAM_PORT=0
-
-while [[ $# -gt 0 ]]; do
-    case $1 in
-        --source)         SOURCE="$2"; shift 2 ;;
-        --model)          MODEL="$2"; shift 2 ;;
-        --conf)           CONF="$2"; shift 2 ;;
-        --headless)       HEADLESS=true; shift ;;
-        --max-missing)    MAX_MISSING="$2"; shift 2 ;;
-        --match-distance) MATCH_DIST="$2"; shift 2 ;;
-        --confirm)        CONFIRM="$2"; shift 2 ;;
-        --device)         DEVICE="$2"; shift 2 ;;
-        --imgsz)          IMGSZ="$2"; shift 2 ;;
-        --no-half)        HALF=false; shift ;;
-        --stream)         STREAM_PORT="$2"; shift 2 ;;
-        --cpu)            DEVICE="cpu"; HALF=false; shift ;;
-        *)                echo "Unknown arg: $1"; exit 1 ;;
-    esac
-done
-
-source ~/miniconda3/etc/profile.d/conda.sh 2>/dev/null || true
-conda activate marcus 2>/dev/null || conda activate teleimager 2>/dev/null || true
-
-YEAR=$(date +%Y)
-if [ "$YEAR" -lt 2025 ]; then
-    echo "[WARN] System clock is wrong (year=$YEAR). Fixing..."
-    echo "123" | sudo -S date -s "2026-04-10 16:00:00" 2>/dev/null || true
-fi
-
-if [ "$HEADLESS" = true ]; then
-    export QT_QPA_PLATFORM=offscreen
-    HEADLESS_FLAG="--headless"
-    echo "Mode: HEADLESS (no display, results saved to runtime/captures/)"
-else
-    xhost + >/dev/null 2>&1 || true
-    export DISPLAY=:0
-    HEADLESS_FLAG=""
-    echo "Mode: DISPLAY (OpenCV window on monitor)"
-fi
-
-HALF_FLAG=""
-if [ "$HALF" = true ]; then
-    HALF_FLAG="--half"
-fi
-
-STREAM_FLAG=""
-if [ "$STREAM_PORT" -gt 0 ]; then
-    STREAM_FLAG="--stream $STREAM_PORT"
-fi
-
-echo "============================================"
-echo "  Saqr PPE Detection - Unitree G1 Robot"
-echo "============================================"
-echo "  Source : $SOURCE"
-echo "  Model  : $MODEL"
-echo "  Device : $DEVICE (half=$HALF, imgsz=$IMGSZ)"
-echo "  Conf   : $CONF"
-echo "  Stream : ${STREAM_PORT:-disabled}"
-echo "============================================"
-echo ""
-
-python -m saqr.apps.saqr_cli \
-    --source "$SOURCE" \
-    --model "$MODEL" \
-    --conf "$CONF" \
-    --max-missing "$MAX_MISSING" \
-    --match-distance "$MATCH_DIST" \
-    --status-confirm-frames "$CONFIRM" \
-    --device "$DEVICE" \
-    --imgsz "$IMGSZ" \
-    $HALF_FLAG \
-    $STREAM_FLAG \
-    $HEADLESS_FLAG
diff --git a/scripts/start_saqr.sh b/scripts/start_saqr.sh
index 0f65038..37bf460 100755
--- a/scripts/start_saqr.sh
+++ b/scripts/start_saqr.sh
@@ -6,7 +6,7 @@
 #  What it does:
 #    1. Sources miniconda and activates the target env (default: marcus).
 #    2. cd to the project root (parent of this scripts/ dir).
-#    3. Execs `python -m saqr.robot.bridge` with the production flags.
+#    3. Execs `python -m robot.bridge` with the production flags.
 #
 #  The bridge will:
 #    - init the G1 arm + audio + LowState DDS clients
@@ -22,19 +22,33 @@ set -u
 HERE="$(cd "$(dirname "$0")" && pwd)"
 SAQR_DIR="${SAQR_DIR:-$(cd "$HERE/.." && pwd)}"
 
-CONDA_ROOT="${CONDA_ROOT:-$HOME/miniconda3}"
-CONDA_ENV="${CONDA_ENV:-marcus}"
-DDS_IFACE="${DDS_IFACE:-eth0}"
-SAQR_SOURCE="${SAQR_SOURCE:-realsense}"
-STREAM_PORT="${STREAM_PORT:-8080}"
+# Read defaults from config/robot_config.json (env vars override).
+config_get() {
+    # config_get dotted.key
+    python3 -c "
+import json, os, sys
+with open('$SAQR_DIR/config/robot_config.json') as f:
+    c = json.load(f)
+for k in sys.argv[1].split('.'):
+    c = c[k]
+print(os.path.expandvars(str(c)))
+" "$1"
+}
+
+CONDA_ROOT="${CONDA_ROOT:-$(config_get start_saqr.conda_root)}"
+CONDA_ENV="${CONDA_ENV:-$(config_get start_saqr.conda_env)}"
+DDS_IFACE="${DDS_IFACE:-$(config_get start_saqr.dds_iface)}"
+SAQR_SOURCE="${SAQR_SOURCE:-$(config_get start_saqr.saqr_source)}"
+STREAM_PORT="${STREAM_PORT:-$(config_get start_saqr.stream_port)}"
 
 if [ ! -d "$SAQR_DIR" ]; then
     echo "[start_saqr] FATAL: SAQR_DIR not found: $SAQR_DIR" >&2
     exit 1
 fi
 
-if [ ! -d "$SAQR_DIR/saqr" ]; then
-    echo "[start_saqr] FATAL: saqr/ package not found in $SAQR_DIR" >&2
+if [ ! -f "$SAQR_DIR/robot/bridge.py" ]; then
+    echo "[start_saqr] FATAL: robot/bridge.py not found in $SAQR_DIR" >&2
+    echo "                    (expected $SAQR_DIR to contain core/ apps/ robot/ utils/)" >&2
     exit 1
 fi
 
@@ -58,7 +72,7 @@ cd "$SAQR_DIR" || {
 echo "[start_saqr] env=$CONDA_ENV cwd=$PWD iface=$DDS_IFACE source=$SAQR_SOURCE stream=$STREAM_PORT"
 echo "[start_saqr] launching bridge..."
 
-exec python3 -m saqr.robot.bridge \
+exec python3 -m robot.bridge \
     --iface "$DDS_IFACE" \
     --source "$SAQR_SOURCE" \
     --headless \
diff --git a/utils/__init__.py b/utils/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/utils/config.py b/utils/config.py
new file mode 100644
index 0000000..526bde5
--- /dev/null
+++ b/utils/config.py
@@ -0,0 +1,45 @@
+"""Cached JSON config loader for config/<name>_config.json files.
+
+Usage:
+    from utils.config import load_config
+    cfg = load_config("core")           # reads config/core_config.json
+    conf_threshold = cfg["detection"]["conf"]
+
+String values pass through ``os.path.expandvars`` on load, so JSON like
+``"$HOME/miniconda3"`` resolves at read time.
+"""
+from __future__ import annotations
+
+import json
+import os
+from typing import Any, Dict
+
+from core.paths import CONFIG_DIR
+
+_CACHE: Dict[str, Dict[str, Any]] = {}
+
+
+def _expand(value):
+    if isinstance(value, str):
+        return os.path.expandvars(value)
+    if isinstance(value, list):
+        return [_expand(v) for v in value]
+    if isinstance(value, dict):
+        return {k: _expand(v) for k, v in value.items()}
+    return value
+
+
+def load_config(name: str) -> Dict[str, Any]:
+    """Return the parsed JSON config for ``config/<name>_config.json`` (cached)."""
+    if name in _CACHE:
+        return _CACHE[name]
+
+    path = CONFIG_DIR / f"{name}_config.json"
+    if not path.exists():
+        raise FileNotFoundError(f"Config not found: {path}")
+
+    with open(path, "r") as f:
+        raw = json.load(f)
+    cfg = _expand(raw)
+    _CACHE[name] = cfg
+    return cfg
diff --git a/saqr/utils/logger.py b/utils/logger.py
similarity index 91%
rename from saqr/utils/logger.py
rename to utils/logger.py
index 06d6d9e..3910fc5 100644
--- a/saqr/utils/logger.py
+++ b/utils/logger.py
@@ -4,7 +4,7 @@ import json
 import logging
 from typing import Dict
 
-from saqr.core.paths import CONFIG_DIR, LOGS_DIR
+from core.paths import CONFIG_DIR, LOGS_DIR
 
 _LOGGER_CACHE: Dict[str, logging.Logger] = {}
 
@@ -23,7 +23,7 @@ def _level_from_name(name: str) -> int:
 
 
 def get_logger(category: str, name: str) -> logging.Logger:
-    """Return a cached logger that writes to runtime/logs/<category>/<name>.log."""
+    """Return a cached logger that writes to logs/<category>/<name>.log."""
     key = f"{category}.{name}"
     if key in _LOGGER_CACHE:
         return _LOGGER_CACHE[key]