diff --git a/scripts/speed_limit_vision/build_track_proposal_detector_dataset.py b/scripts/speed_limit_vision/build_track_proposal_detector_dataset.py
new file mode 100644
index 000000000..3edd9e99b
--- /dev/null
+++ b/scripts/speed_limit_vision/build_track_proposal_detector_dataset.py
@@ -0,0 +1,242 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import csv
+import hashlib
+import json
+import os
+import shutil
+
+from collections import Counter
+from pathlib import Path
+
+import cv2
+import yaml
+
+if __package__ in (None, ""):
+  import sys
+  sys.path.insert(0, str(Path(__file__).resolve().parent))
+  from import_manual_review_queue import merged_review_rows, parse_speed  # type: ignore
+else:
+  from .import_manual_review_queue import merged_review_rows, parse_speed
+
+
+POSITIVE_STATUSES = frozenset(("accepted", "corrected"))
+IMPORTANT_SPEEDS = frozenset(range(30, 70, 5))
+
+
+def parse_args() -> argparse.Namespace:
+  parser = argparse.ArgumentParser(description="Add clean reviewed sign tracks to the single-class runtime proposal detector.")
+  parser.add_argument("--base-yaml", type=Path, required=True, help="Existing single-class detector dataset YAML to preserve.")
+  parser.add_argument("--queue", type=Path, required=True, help="Reviewed manual_review_queue.csv used to create the tracks.")
+  parser.add_argument("--labels", type=Path, help="Defaults to manual_review_labels.csv beside the queue.")
+  parser.add_argument("--track-samples", type=Path, required=True, help="track_samples.csv from mine_reviewed_sign_tracks.py.")
+  parser.add_argument("--output", type=Path, required=True, help="Output directory for added images, labels, and dataset.yaml.")
+  parser.add_argument("--train-ratio", type=float, default=0.85, help="Route-level train split ratio.")
+  parser.add_argument("--min-growth", type=float, default=1.0, help="Minimum tracked box area relative to its reviewed anchor.")
+  parser.add_argument("--min-exact-confidence", type=float, default=0.80)
+  parser.add_argument("--min-detector-confidence", type=float, default=0.30)
+  parser.add_argument("--max-track-rank", type=int, default=4)
+  parser.add_argument("--important-repeat", type=int, default=2, help="Train repeats for accepted 30-65 mph samples.")
+  parser.add_argument("--other-repeat", type=int, default=1, help="Train repeats for other accepted speed samples.")
+  return parser.parse_args()
+
+
+def parse_bbox(value: str) -> tuple[int, int, int, int] | None:
+  try:
+    values = tuple(int(round(float(part.strip()))) for part in value.split(","))
+  except ValueError:
+    return None
+  if len(values) != 4:
+    return None
+  x1, y1, x2, y2 = values
+  return values if x2 > x1 and y2 > y1 else None
+
+
+def split_for_key(key: str, train_ratio: float) -> str:
+  fraction = int(hashlib.sha1(key.encode()).hexdigest()[:8], 16) / 0xFFFFFFFF
+  return "train" if fraction < train_ratio else "val"
+
+
+def link_or_copy(source: Path, destination: Path) -> None:
+  destination.parent.mkdir(parents=True, exist_ok=True)
+  if destination.exists():
+    return
+  try:
+    os.link(source, destination)
+  except OSError:
+    shutil.copy2(source, destination)
+
+
+def yolo_label(bbox: tuple[int, int, int, int], image_path: Path) -> str | None:
+  image = cv2.imread(str(image_path))
+  if image is None:
+    return None
+  height, width = image.shape[:2]
+  x1, y1, x2, y2 = bbox
+  x1 = max(min(x1, width - 1), 0)
+  y1 = max(min(y1, height - 1), 0)
+  x2 = max(min(x2, width), 0)
+  y2 = max(min(y2, height), 0)
+  if x2 <= x1 or y2 <= y1:
+    return None
+  values = (
+    (x1 + x2) / (2 * width),
+    (y1 + y2) / (2 * height),
+    (x2 - x1) / width,
+    (y2 - y1) / height,
+  )
+  return f"0 {values[0]:.8f} {values[1]:.8f} {values[2]:.8f} {values[3]:.8f}\n"
+
+
+def reviewed_positive_rows(queue_path: Path, labels_path: Path) -> dict[str, dict[str, str]]:
+  return {
+    row.get("record_key", ""): row
+    for row in merged_review_rows(queue_path, labels_path)
+    if row.get("record_key") and row.get("review_status") in POSITIVE_STATUSES and parse_speed(row.get("review_speed_limit_mph", ""))
+  }
+
+
+def trusted_track_row(row: dict[str, str], reviewed_row: dict[str, str], args: argparse.Namespace) -> bool:
+  original_bbox = parse_bbox(reviewed_row.get("bbox", ""))
+  corrected_bbox = parse_bbox(reviewed_row.get("review_bbox", ""))
+  if corrected_bbox is not None and corrected_bbox != original_bbox:
+    # Existing tracks predate manual box correction, so their propagated boxes are stale.
+    return False
+  try:
+    expected = int(row.get("expected_speed_limit_mph", ""))
+    reviewed_speed = int(parse_speed(reviewed_row.get("review_speed_limit_mph", "")) or 0)
+    predicted = int(row.get("predicted_speed_limit_mph", "") or 0)
+    read_confidence = float(row.get("read_confidence", "") or 0.0)
+    detector_confidence = float(row.get("detector_confidence", "") or 0.0)
+    growth = float(row.get("area_ratio_to_anchor", "") or 0.0)
+    rank = int(row.get("rank", "") or 999)
+  except ValueError:
+    return False
+  exact_read = predicted == expected and read_confidence >= args.min_exact_confidence
+  detector_snap = detector_confidence >= args.min_detector_confidence
+  return expected == reviewed_speed and growth >= args.min_growth and rank <= args.max_track_rank and (exact_read or detector_snap)
+
+
+def add_sample(
+  output: Path,
+  split: str,
+  stem: str,
+  image_path: Path,
+  bbox: tuple[int, int, int, int],
+  repeats: int,
+) -> int:
+  label = yolo_label(bbox, image_path)
+  if label is None:
+    return 0
+  created = 0
+  for repeat in range(max(repeats, 1) if split == "train" else 1):
+    suffix = f"_r{repeat:02d}" if split == "train" and repeats > 1 else ""
+    destination_stem = f"{stem}{suffix}"
+    destination_image = output / "images" / split / f"{destination_stem}{image_path.suffix.lower()}"
+    destination_label = output / "labels" / split / f"{destination_stem}.txt"
+    link_or_copy(image_path, destination_image)
+    destination_label.parent.mkdir(parents=True, exist_ok=True)
+    destination_label.write_text(label, encoding="ascii")
+    created += 1
+  return created
+
+
+def add_reviewed_anchors(
+  reviewed_rows: dict[str, dict[str, str]],
+  output: Path,
+  args: argparse.Namespace,
+) -> Counter[str]:
+  counts: Counter[str] = Counter()
+  for record_key, row in reviewed_rows.items():
+    speed = int(parse_speed(row.get("review_speed_limit_mph", "")) or 0)
+    bbox = parse_bbox(row.get("review_bbox") or row.get("bbox", ""))
+    image_path = Path(row.get("frame_path", "")).expanduser().resolve()
+    if bbox is None or not image_path.is_file():
+      counts["anchor_rejected"] += 1
+      continue
+    split = split_for_key(row.get("route") or record_key, args.train_ratio)
+    repeats = args.important_repeat if speed in IMPORTANT_SPEEDS else args.other_repeat
+    created = add_sample(output, split, f"anchor_{record_key}", image_path, bbox, repeats)
+    counts[f"anchor_{split}"] += created
+    counts[f"speed_{speed}"] += created
+  return counts
+
+
+def add_track_samples(
+  reviewed_rows: dict[str, dict[str, str]],
+  output: Path,
+  args: argparse.Namespace,
+) -> Counter[str]:
+  counts: Counter[str] = Counter()
+  with args.track_samples.expanduser().resolve().open(encoding="utf-8", newline="") as input_file:
+    for row in csv.DictReader(input_file):
+      reviewed_row = reviewed_rows.get(row.get("track_key", ""))
+      if reviewed_row is None or not trusted_track_row(row, reviewed_row, args):
+        counts["track_rejected"] += 1
+        continue
+      bbox = parse_bbox(row.get("bbox", ""))
+      image_path = Path(row.get("frame_path", "")).expanduser().resolve()
+      if bbox is None or not image_path.is_file():
+        counts["track_rejected"] += 1
+        continue
+      speed = int(row["expected_speed_limit_mph"])
+      split = split_for_key(row.get("route") or row.get("track_key", ""), args.train_ratio)
+      repeats = args.important_repeat if speed in IMPORTANT_SPEEDS else args.other_repeat
+      stem = f"track_{row.get('track_key', '')}_{row.get('rank', '')}_{row.get('frame_time_s', '').replace('.', 'p')}"
+      created = add_sample(output, split, stem, image_path, bbox, repeats)
+      counts[f"track_{split}"] += created
+      counts[f"speed_{speed}"] += created
+  return counts
+
+
+def resolved_dataset_paths(base_yaml: Path, key: str) -> list[str]:
+  data = yaml.safe_load(base_yaml.read_text(encoding="utf-8")) or {}
+  base_root = Path(data.get("path", base_yaml.parent)).expanduser()
+  if not base_root.is_absolute():
+    base_root = (base_yaml.parent / base_root).resolve()
+  entries = data.get(key, [])
+  if isinstance(entries, str):
+    entries = [entries]
+  return [str((base_root / entry).resolve()) if not Path(entry).is_absolute() else str(Path(entry).resolve()) for entry in entries]
+
+
+def write_dataset_yaml(base_yaml: Path, output: Path) -> Path:
+  train_paths = (*resolved_dataset_paths(base_yaml, "train"), str((output / "images" / "train").resolve()))
+  val_paths = (*resolved_dataset_paths(base_yaml, "val"), str((output / "images" / "val").resolve()))
+  lines = ["train:", *(f"  - {path}" for path in train_paths), "val:", *(f"  - {path}" for path in val_paths), "names:", "  0: speed_limit_sign"]
+  dataset_yaml = output / "dataset.yaml"
+  dataset_yaml.write_text("\n".join(lines) + "\n", encoding="ascii")
+  return dataset_yaml
+
+
+def main() -> int:
+  args = parse_args()
+  if not 0.0 < args.train_ratio < 1.0:
+    raise ValueError("--train-ratio must be between zero and one")
+  queue_path = args.queue.expanduser().resolve()
+  labels_path = args.labels.expanduser().resolve() if args.labels else queue_path.with_name("manual_review_labels.csv")
+  output = args.output.expanduser().resolve()
+  for split in ("train", "val"):
+    (output / "images" / split).mkdir(parents=True, exist_ok=True)
+    (output / "labels" / split).mkdir(parents=True, exist_ok=True)
+
+  reviewed_rows = reviewed_positive_rows(queue_path, labels_path)
+  counts = add_reviewed_anchors(reviewed_rows, output, args)
+  counts.update(add_track_samples(reviewed_rows, output, args))
+  dataset_yaml = write_dataset_yaml(args.base_yaml.expanduser().resolve(), output)
+  summary = {
+    "accepted_source_records": len(reviewed_rows),
+    "base_yaml": str(args.base_yaml.expanduser().resolve()),
+    "dataset_yaml": str(dataset_yaml),
+    "output": str(output),
+    "counts": dict(sorted(counts.items())),
+  }
+  (output / "track_proposal_dataset_summary.json").write_text(json.dumps(summary, indent=2, sort_keys=True) + "\n", encoding="ascii")
+  print(json.dumps(summary, indent=2, sort_keys=True))
+  return 0
+
+
+if __name__ == "__main__":
+  raise SystemExit(main())
diff --git a/scripts/speed_limit_vision/evaluate_route_ground_truth.py b/scripts/speed_limit_vision/evaluate_route_ground_truth.py
new file mode 100644
index 000000000..9d921bda4
--- /dev/null
+++ b/scripts/speed_limit_vision/evaluate_route_ground_truth.py
@@ -0,0 +1,241 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+
+from collections import Counter, defaultdict
+from dataclasses import dataclass
+from pathlib import Path
+
+import starpilot.system.speed_limit_vision as slv
+
+if __package__ in (None, ""):
+  import sys
+  sys.path.insert(0, str(Path(__file__).resolve().parent))
+  from replay_route_runtime import (  # type: ignore
+    build_runtime_context,
+    configure_models,
+    qlog_paths,
+    replay_route,
+    route_log_id,
+    segment_paths,
+  )
+else:
+  from .replay_route_runtime import build_runtime_context, configure_models, qlog_paths, replay_route, route_log_id, segment_paths
+
+
+@dataclass(frozen=True)
+class GroundTruthEvent:
+  route: str
+  time_s: float
+  speed_limit_mph: int
+  sign_type: str
+  initial_speed_limit_mph: int
+  notes: str
+
+
+def parse_args() -> argparse.Namespace:
+  parser = argparse.ArgumentParser(description="Score continuous route replay against independently labeled speed-sign events.")
+  parser.add_argument("--ground-truth", type=Path, required=True, help="CSV with route,time_s,speed_limit_mph and optional metadata.")
+  parser.add_argument("--clip-root", type=Path, default=Path("/Volumes/T5/starpilot_speed_limit/realdata"))
+  parser.add_argument("--models-dir", type=Path, default=Path("starpilot/assets/vision_models"))
+  parser.add_argument("--output-csv", type=Path, required=True)
+  parser.add_argument("--window-before", type=float, default=2.0)
+  parser.add_argument("--window-after", type=float, default=4.0)
+  parser.add_argument("--event-dedupe-seconds", type=float, default=2.0)
+  parser.add_argument("--qlog-context", action="store_true")
+  parser.add_argument("--measured-base-inference-seconds", type=float, default=0.44)
+  parser.add_argument("--measured-classifier-forward-seconds", type=float, default=0.066)
+  parser.add_argument("--crop-ocr", action="store_true")
+  return parser.parse_args()
+
+
+def int_value(text: str, default: int = 0) -> int:
+  try:
+    return int(float(text))
+  except (TypeError, ValueError):
+    return default
+
+
+def load_ground_truth(path: Path) -> list[GroundTruthEvent]:
+  events: list[GroundTruthEvent] = []
+  with path.open(encoding="utf-8", newline="") as input_file:
+    for row in csv.DictReader(input_file):
+      route = route_log_id(row.get("route", ""))
+      try:
+        time_s = float(row.get("time_s", ""))
+      except ValueError:
+        continue
+      speed = int_value(row.get("speed_limit_mph", ""))
+      applicable = row.get("applicable", "true").strip().lower() not in ("0", "false", "no")
+      if not route or time_s < 0 or speed <= 0 or not applicable:
+        continue
+      events.append(GroundTruthEvent(
+        route=route,
+        time_s=time_s,
+        speed_limit_mph=speed,
+        sign_type=row.get("sign_type", "regulatory").strip() or "regulatory",
+        initial_speed_limit_mph=int_value(row.get("initial_speed_limit_mph", "")),
+        notes=row.get("notes", "").strip(),
+      ))
+  return sorted(events, key=lambda event: (event.route, event.time_s))
+
+
+def runtime_value(event: dict[str, str]) -> int:
+  key = "candidateSpeedLimitMph" if event.get("event") == "candidate" else "speedLimitMph"
+  return int_value(event.get(key, ""))
+
+
+def clustered_runtime_events(events: list[dict[str, str]], event_type: str, dedupe_seconds: float) -> list[dict[str, str]]:
+  clusters: list[dict[str, str]] = []
+  for event in events:
+    if event.get("event") != event_type or runtime_value(event) <= 0:
+      continue
+    if (
+      clusters and
+      runtime_value(clusters[-1]) == runtime_value(event) and
+      float(event["time_s"]) - float(clusters[-1]["time_s"]) <= dedupe_seconds
+    ):
+      continue
+    clusters.append(event)
+  return clusters
+
+
+def events_in_window(events: list[dict[str, str]], truth: GroundTruthEvent, before: float, after: float) -> list[dict[str, str]]:
+  return [event for event in events if truth.time_s - before <= float(event["time_s"]) <= truth.time_s + after]
+
+
+def first_exact(events: list[dict[str, str]], speed: int) -> dict[str, str] | None:
+  return next((event for event in events if runtime_value(event) == speed), None)
+
+
+def score_route(
+  truth_events: list[GroundTruthEvent],
+  runtime_events: list[dict[str, str]],
+  window_before: float,
+  window_after: float,
+  dedupe_seconds: float,
+) -> tuple[list[dict[str, object]], Counter[str]]:
+  candidates = clustered_runtime_events(runtime_events, "candidate", dedupe_seconds)
+  publishes = clustered_runtime_events(runtime_events, "publish", dedupe_seconds)
+  rows: list[dict[str, object]] = []
+  totals: Counter[str] = Counter(total=len(truth_events))
+
+  for truth in truth_events:
+    candidate_window = events_in_window(candidates, truth, window_before, window_after)
+    publish_window = events_in_window(publishes, truth, window_before, window_after)
+    exact_candidate = first_exact(candidate_window, truth.speed_limit_mph)
+    exact_publish = first_exact(publish_window, truth.speed_limit_mph)
+    wrong_candidates = sorted({runtime_value(event) for event in candidate_window if runtime_value(event) != truth.speed_limit_mph})
+    wrong_publishes = sorted({runtime_value(event) for event in publish_window if runtime_value(event) != truth.speed_limit_mph})
+    totals.update(
+      candidate_hit=int(exact_candidate is not None),
+      publish_hit=int(exact_publish is not None),
+      wrong_candidate=int(bool(wrong_candidates)),
+      wrong_publish=int(bool(wrong_publishes)),
+    )
+    rows.append({
+      "route": truth.route,
+      "time_s": f"{truth.time_s:.3f}",
+      "speed_limit_mph": truth.speed_limit_mph,
+      "sign_type": truth.sign_type,
+      "candidate_hit": exact_candidate is not None,
+      "publish_hit": exact_publish is not None,
+      "candidate_latency_s": f"{float(exact_candidate['time_s']) - truth.time_s:.3f}" if exact_candidate else "",
+      "publish_latency_s": f"{float(exact_publish['time_s']) - truth.time_s:.3f}" if exact_publish else "",
+      "candidate_values": "|".join(str(runtime_value(event)) for event in candidate_window),
+      "publish_values": "|".join(str(runtime_value(event)) for event in publish_window),
+      "wrong_candidate_values": "|".join(map(str, wrong_candidates)),
+      "wrong_publish_values": "|".join(map(str, wrong_publishes)),
+      "notes": truth.notes,
+    })
+
+  unmatched_publishes = [
+    event for event in publishes
+    if not any(
+      truth.time_s - window_before <= float(event["time_s"]) <= truth.time_s + window_after and
+      runtime_value(event) == truth.speed_limit_mph
+      for truth in truth_events
+    )
+  ]
+  totals["publish_bursts"] = len(publishes)
+  totals["unmatched_publish_bursts"] = len(unmatched_publishes)
+  return rows, totals
+
+
+def main() -> int:
+  args = parse_args()
+  configure_models(args.models_dir)
+  slv.DETECTOR_CLASSIFIER_CROP_OCR_ENABLED = args.crop_ocr
+  clip_root = args.clip_root.expanduser().resolve()
+  truth_by_route: dict[str, list[GroundTruthEvent]] = defaultdict(list)
+  for event in load_ground_truth(args.ground_truth.expanduser().resolve()):
+    truth_by_route[event.route].append(event)
+  if not truth_by_route:
+    raise ValueError("Ground-truth manifest contains no applicable speed-sign events")
+
+  output_rows: list[dict[str, object]] = []
+  aggregate: Counter[str] = Counter()
+  by_speed: dict[int, Counter[str]] = defaultdict(Counter)
+  route_summaries: dict[str, dict[str, int]] = {}
+  for route, truth_events in truth_by_route.items():
+    segments = segment_paths(clip_root, route)
+    if not segments:
+      raise FileNotFoundError(f"No camera segments for {route} under {clip_root}")
+    runtime_context = None
+    if args.qlog_context:
+      qlogs = qlog_paths(clip_root, route)
+      runtime_context = build_runtime_context(qlogs) if qlogs else None
+    initial_speed = next((event.initial_speed_limit_mph for event in truth_events if event.initial_speed_limit_mph > 0), 0)
+    _summary, runtime_events = replay_route(
+      route,
+      segments,
+      runtime_context,
+      0.0,
+      None,
+      False,
+      False,
+      0.0,
+      args.measured_base_inference_seconds,
+      args.measured_classifier_forward_seconds,
+      initial_speed,
+    )
+    rows, totals = score_route(truth_events, runtime_events, args.window_before, args.window_after, args.event_dedupe_seconds)
+    output_rows.extend(rows)
+    aggregate.update(totals)
+    route_summaries[route] = dict(totals)
+    for row in rows:
+      speed_counts = by_speed[int(row["speed_limit_mph"])]
+      speed_counts.update(
+        total=1,
+        candidate_hit=int(bool(row["candidate_hit"])),
+        publish_hit=int(bool(row["publish_hit"])),
+        wrong_publish=int(bool(row["wrong_publish_values"])),
+      )
+
+  output_path = args.output_csv.expanduser().resolve()
+  output_path.parent.mkdir(parents=True, exist_ok=True)
+  with output_path.open("w", encoding="utf-8", newline="") as output_file:
+    writer = csv.DictWriter(output_file, fieldnames=list(output_rows[0]))
+    writer.writeheader()
+    writer.writerows(output_rows)
+  summary = {
+    "models_dir": str(args.models_dir.expanduser().resolve()),
+    "ground_truth": str(args.ground_truth.expanduser().resolve()),
+    "continuous_route_replay": True,
+    "detector_selected_inputs": False,
+    "measured_base_inference_seconds": args.measured_base_inference_seconds,
+    "measured_classifier_forward_seconds": args.measured_classifier_forward_seconds,
+    "totals": dict(aggregate),
+    "by_speed": {str(speed): dict(counts) for speed, counts in sorted(by_speed.items())},
+    "by_route": route_summaries,
+  }
+  output_path.with_suffix(".json").write_text(json.dumps(summary, indent=2, sort_keys=True) + "\n", encoding="ascii")
+  print(json.dumps(summary, indent=2, sort_keys=True))
+  return 0
+
+
+if __name__ == "__main__":
+  raise SystemExit(main())
diff --git a/scripts/speed_limit_vision/test_route_ground_truth.py b/scripts/speed_limit_vision/test_route_ground_truth.py
new file mode 100644
index 000000000..029f78112
--- /dev/null
+++ b/scripts/speed_limit_vision/test_route_ground_truth.py
@@ -0,0 +1,44 @@
+from __future__ import annotations
+
+import importlib.util
+import sys
+
+from pathlib import Path
+
+
+def load_local_module(name: str):
+  path = Path(__file__).resolve().with_name(f"{name}.py")
+  spec = importlib.util.spec_from_file_location(f"test_local_{name}", path)
+  assert spec is not None and spec.loader is not None
+  module = importlib.util.module_from_spec(spec)
+  sys.modules[spec.name] = module
+  spec.loader.exec_module(module)
+  return module
+
+
+route_eval = load_local_module("evaluate_route_ground_truth")
+GroundTruthEvent = route_eval.GroundTruthEvent
+
+
+def test_clustered_runtime_events_collapses_same_value_burst():
+  events = [
+    {"event": "publish", "time_s": "10.0", "speedLimitMph": "35"},
+    {"event": "publish", "time_s": "10.5", "speedLimitMph": "35"},
+    {"event": "publish", "time_s": "11.0", "speedLimitMph": "45"},
+  ]
+  clusters = route_eval.clustered_runtime_events(events, "publish", 2.0)
+  assert [route_eval.runtime_value(event) for event in clusters] == [35, 45]
+
+
+def test_score_route_separates_candidate_publish_and_wrong_value():
+  truth = [GroundTruthEvent("route", 10.0, 35, "regulatory", 40, "")]
+  events = [
+    {"event": "candidate", "time_s": "10.2", "candidateSpeedLimitMph": "35"},
+    {"event": "publish", "time_s": "10.4", "speedLimitMph": "45"},
+  ]
+  rows, totals = route_eval.score_route(truth, events, 1.0, 2.0, 0.5)
+  assert totals["candidate_hit"] == 1
+  assert totals["publish_hit"] == 0
+  assert totals["wrong_publish"] == 1
+  assert totals["unmatched_publish_bursts"] == 1
+  assert rows[0]["wrong_publish_values"] == "45"
diff --git a/scripts/speed_limit_vision/test_track_proposal_detector_dataset.py b/scripts/speed_limit_vision/test_track_proposal_detector_dataset.py
new file mode 100644
index 000000000..c33d0a587
--- /dev/null
+++ b/scripts/speed_limit_vision/test_track_proposal_detector_dataset.py
@@ -0,0 +1,82 @@
+from __future__ import annotations
+
+import importlib.util
+
+from argparse import Namespace
+from pathlib import Path
+
+
+def load_local_module(name: str):
+  path = Path(__file__).resolve().with_name(f"{name}.py")
+  spec = importlib.util.spec_from_file_location(f"test_local_{name}", path)
+  assert spec is not None and spec.loader is not None
+  module = importlib.util.module_from_spec(spec)
+  spec.loader.exec_module(module)
+  return module
+
+
+track_dataset = load_local_module("build_track_proposal_detector_dataset")
+split_for_key = track_dataset.split_for_key
+trusted_track_row = track_dataset.trusted_track_row
+
+
+def args() -> Namespace:
+  return Namespace(min_exact_confidence=0.8, min_detector_confidence=0.3, min_growth=1.0, max_track_rank=4)
+
+
+def test_split_for_key_keeps_route_samples_together():
+  assert split_for_key("route-a", 0.85) == split_for_key("route-a", 0.85)
+
+
+def test_trusted_track_requires_current_review_speed_to_match():
+  track = {
+    "expected_speed_limit_mph": "45",
+    "predicted_speed_limit_mph": "45",
+    "read_confidence": "0.99",
+    "detector_confidence": "0.0",
+    "area_ratio_to_anchor": "1.2",
+    "rank": "1",
+  }
+  assert trusted_track_row(track, {"review_speed_limit_mph": "45"}, args())
+  assert not trusted_track_row(track, {"review_speed_limit_mph": "55"}, args())
+
+
+def test_trusted_track_accepts_detector_snap_without_classifier_read():
+  track = {
+    "expected_speed_limit_mph": "35",
+    "predicted_speed_limit_mph": "",
+    "read_confidence": "",
+    "detector_confidence": "0.7",
+    "area_ratio_to_anchor": "1.0",
+    "rank": "4",
+  }
+  assert trusted_track_row(track, {"review_speed_limit_mph": "35"}, args())
+
+
+def test_trusted_track_rejects_low_growth_and_rank():
+  track = {
+    "expected_speed_limit_mph": "35",
+    "predicted_speed_limit_mph": "35",
+    "read_confidence": "0.99",
+    "detector_confidence": "0.7",
+    "area_ratio_to_anchor": "0.9",
+    "rank": "5",
+  }
+  assert not trusted_track_row(track, {"review_speed_limit_mph": "35"}, args())
+
+
+def test_trusted_track_rejects_tracks_from_redrawn_anchor():
+  track = {
+    "expected_speed_limit_mph": "35",
+    "predicted_speed_limit_mph": "35",
+    "read_confidence": "0.99",
+    "detector_confidence": "0.7",
+    "area_ratio_to_anchor": "1.2",
+    "rank": "1",
+  }
+  review = {
+    "review_speed_limit_mph": "35",
+    "bbox": "10,10,40,50",
+    "review_bbox": "12,12,35,45",
+  }
+  assert not trusted_track_row(track, review, args())
diff --git a/starpilot/assets/vision_models/speed_limit_us_detector.onnx b/starpilot/assets/vision_models/speed_limit_us_detector.onnx
index 275c2c27b..2f1f1cc24 100755
Binary files a/starpilot/assets/vision_models/speed_limit_us_detector.onnx and b/starpilot/assets/vision_models/speed_limit_us_detector.onnx differ