diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..362a3ad
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,7 @@
+__pycache__/
+.claude/
+crops/
+dataset/
+debug/
+*.tgz
+nul
diff --git a/eval.py b/eval.py
new file mode 100644
index 0000000..25ecf41
--- /dev/null
+++ b/eval.py
@@ -0,0 +1,148 @@
+"""Evaluate the saved accidental classifier on the held-out test set.
+
+Uses the same seed (42), split ratios, and data loading as train.py
+so the test set is identical to what training used.
+"""
+
+import os
+import random
+from pathlib import Path
+from collections import defaultdict
+
+import numpy as np
+import torch
+from PIL import Image
+
+from train import (
+    AccidentalCNN,
+    CROPS_ROOT,
+    OUTPUT_DIR,
+    LABELS,
+    IMG_SIZE,
+    SEED,
+    VAL_SPLIT,
+    TEST_SPLIT,
+    MAX_PER_CLASS,
+    InMemoryDataset,
+    load_all_into_ram,
+    split_data,
+)
+from torch.utils.data import DataLoader
+
+MODEL_PATH = OUTPUT_DIR / "accidental_classifier.pt"
+
+
+def collect_file_paths():
+    """Collect file paths in the same order as load_all_into_ram.
+
+    Must be called after seeding so random.shuffle matches train.py.
+    """
+    paths = []
+    for label_idx, label_name in enumerate(LABELS):
+        label_dir = CROPS_ROOT / label_name
+        files = [f for f in os.listdir(label_dir) if f.endswith(".png")]
+        random.shuffle(files)
+        if MAX_PER_CLASS and len(files) > MAX_PER_CLASS:
+            files = files[:MAX_PER_CLASS]
+        for fname in files:
+            paths.append(str(label_dir / fname))
+    return paths
+
+
+def main():
+    # Reproduce the exact same RNG state and splits as train.py
+    random.seed(SEED)
+    np.random.seed(SEED)
+    torch.manual_seed(SEED)
+
+    # Collect file paths (consumes same random state as load_all_into_ram)
+    file_paths = collect_file_paths()
+
+    # Re-seed and reload to get tensors (load_all_into_ram shuffles again internally)
+    random.seed(SEED)
+    np.random.seed(SEED)
+    torch.manual_seed(SEED)
+
+    images, labels = load_all_into_ram()
+    _, _, test_idx = split_data(images, labels)
+
+    print(f"\nTest set size: {len(test_idx)}")
+    for i, name in enumerate(LABELS):
+        n = (labels[test_idx] == i).sum().item()
+        print(f"  {name:8s}: {n}")
+
+    # Load model
+    device = torch.device("cpu")
+    model = AccidentalCNN().to(device)
+    model.load_state_dict(torch.load(MODEL_PATH, weights_only=True))
+    model.eval()
+    print(f"\nLoaded weights from {MODEL_PATH}")
+
+    # Run inference
+    test_ds = InMemoryDataset(images[test_idx], labels[test_idx], augment=False)
+    test_loader = DataLoader(test_ds, batch_size=256, shuffle=False, num_workers=0)
+
+    all_preds = []
+    all_labels = []
+
+    with torch.no_grad():
+        for imgs, lbls in test_loader:
+            imgs = imgs.to(device)
+            outputs = model(imgs)
+            _, predicted = outputs.max(1)
+            all_preds.append(predicted.cpu())
+            all_labels.append(lbls)
+
+    all_preds = torch.cat(all_preds)
+    all_labels = torch.cat(all_labels)
+
+    # Overall accuracy
+    correct = (all_preds == all_labels).sum().item()
+    total = len(all_labels)
+    print(f"\n=== Test Results ===")
+    print(f"Overall accuracy: {correct}/{total} ({correct/total:.1%})")
+
+    # Per-class accuracy
+    print(f"\nPer-class accuracy:")
+    for i, name in enumerate(LABELS):
+        mask = all_labels == i
+        cls_total = mask.sum().item()
+        cls_correct = (all_preds[mask] == i).sum().item()
+        acc = cls_correct / cls_total if cls_total > 0 else 0
+        print(f"  {name:8s}: {cls_correct}/{cls_total} ({acc:.1%})")
+
+    # Confusion matrix
+    n_classes = len(LABELS)
+    cm = torch.zeros(n_classes, n_classes, dtype=torch.long)
+    for t, p in zip(all_labels, all_preds):
+        cm[t][p] += 1
+
+    print(f"\nConfusion matrix (rows=true, cols=predicted):")
+    header = "           " + "".join(f"{name:>9s}" for name in LABELS)
+    print(header)
+    for i, name in enumerate(LABELS):
+        row = f"  {name:8s}" + "".join(f"{cm[i][j].item():9d}" for j in range(n_classes))
+        print(row)
+
+    # Sample misclassifications
+    test_indices = test_idx.tolist()
+    misclassified = []
+    for i in range(len(all_preds)):
+        if all_preds[i] != all_labels[i]:
+            orig_idx = test_indices[i]
+            misclassified.append((
+                file_paths[orig_idx],
+                LABELS[all_labels[i].item()],
+                LABELS[all_preds[i].item()],
+            ))
+
+    print(f"\nMisclassifications: {len(misclassified)}/{total}")
+    if misclassified:
+        print(f"\nSample misclassifications (up to 10):")
+        for path, true_lbl, pred_lbl in misclassified[:10]:
+            print(f"  {path}")
+            print(f"    true={true_lbl}  predicted={pred_lbl}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/explore_dataset.py b/explore_dataset.py
new file mode 100644
index 0000000..a493b41
--- /dev/null
+++ b/explore_dataset.py
@@ -0,0 +1,102 @@
+"""Explore the PrIMuS dataset to understand accidental distribution and image structure."""
+
+import glob
+import os
+from collections import Counter
+from PIL import Image
+import numpy as np
+
+DATASET_ROOT = r"C:\src\accidentals\dataset"
+
+
+def parse_agnostic(path: str) -> list[str]:
+    """Parse an agnostic encoding file into a list of tokens."""
+    with open(path, "r", encoding="utf-8") as f:
+        return f.read().strip().split("\t")
+
+
+def main():
+    # Find all agnostic files
+    patterns = [
+        os.path.join(DATASET_ROOT, "package_aa", "*", "*.agnostic"),
+        os.path.join(DATASET_ROOT, "package_ab", "*", "*.agnostic"),
+    ]
+    agnostic_files = []
+    for p in patterns:
+        agnostic_files.extend(glob.glob(p))
+
+    print(f"Total incipits: {len(agnostic_files)}")
+
+    # Count accidental tokens
+    accidental_type_counts = Counter()  # sharp/flat/natural
+    accidental_full_counts = Counter()  # full token like accidental.sharp-L5
+    incipits_with_accidentals = 0
+    incipits_with_inline_accidentals = 0  # accidentals that aren't in key sig
+    all_symbol_types = Counter()
+    total_accidentals = 0
+
+    for path in agnostic_files:
+        tokens = parse_agnostic(path)
+        has_any_accidental = False
+        has_inline = False
+        past_time_sig = False
+
+        for tok in tokens:
+            # Track symbol types (just the prefix)
+            base = tok.split("-")[0] if "-" in tok else tok
+            all_symbol_types[base] += 1
+
+            if tok.startswith("digit."):
+                past_time_sig = True
+
+            if tok.startswith("accidental."):
+                has_any_accidental = True
+                total_accidentals += 1
+                # Extract type: accidental.sharp, accidental.flat, etc.
+                acc_type = tok.split("-")[0]  # e.g. "accidental.sharp"
+                accidental_type_counts[acc_type] += 1
+                accidental_full_counts[tok] += 1
+
+                if past_time_sig:
+                    has_inline = True
+
+        if has_any_accidental:
+            incipits_with_accidentals += 1
+        if has_inline:
+            incipits_with_inline_accidentals += 1
+
+    print(f"\n=== Accidental Statistics ===")
+    print(f"Total accidental tokens: {total_accidentals}")
+    print(f"Incipits with any accidentals: {incipits_with_accidentals} / {len(agnostic_files)} ({100*incipits_with_accidentals/len(agnostic_files):.1f}%)")
+    print(f"Incipits with inline accidentals: {incipits_with_inline_accidentals} / {len(agnostic_files)} ({100*incipits_with_inline_accidentals/len(agnostic_files):.1f}%)")
+
+    print(f"\n=== Accidental Type Counts ===")
+    for acc_type, count in accidental_type_counts.most_common():
+        print(f"  {acc_type:25s} {count:7d}")
+
+    print(f"\n=== Top 20 Accidental Positions ===")
+    for tok, count in accidental_full_counts.most_common(20):
+        print(f"  {tok:30s} {count:7d}")
+
+    print(f"\n=== Top 30 Symbol Types ===")
+    for sym, count in all_symbol_types.most_common(30):
+        print(f"  {sym:30s} {count:7d}")
+
+    # Image statistics from a sample
+    print(f"\n=== Image Statistics (sample of 500) ===")
+    png_files = glob.glob(os.path.join(DATASET_ROOT, "package_aa", "*", "*.png"))[:500]
+    widths, heights = [], []
+    for f in png_files:
+        im = Image.open(f)
+        widths.append(im.size[0])
+        heights.append(im.size[1])
+
+    widths = np.array(widths)
+    heights = np.array(heights)
+    print(f"  Width:  min={widths.min()}, max={widths.max()}, mean={widths.mean():.0f}, std={widths.std():.0f}")
+    print(f"  Height: min={heights.min()}, max={heights.max()}, mean={heights.mean():.0f}, std={heights.std():.0f}")
+    print(f"  Modes:  {Counter(Image.open(f).mode for f in png_files[:50])}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/extract_accidentals.py b/extract_accidentals.py
new file mode 100644
index 0000000..8b72f67
--- /dev/null
+++ b/extract_accidentals.py
@@ -0,0 +1,214 @@
+"""Extract accidental crops from the PrIMuS dataset using Verovio re-rendering.
+
+For each MEI file:
+1. Render with Verovio to SVG (getting exact symbol positions)
+2. Rasterize SVG to PNG with cairosvg
+3. Parse SVG for accidental elements (sharp/flat/natural)
+4. Crop each accidental from the rasterized image
+5. Save organized by type: crops/{sharp,flat,natural}/
+
+SMuFL glyph IDs:
+  E262 = sharp
+  E260 = flat
+  E261 = natural
+"""
+
+import glob
+import os
+import re
+import sys
+import traceback
+from io import BytesIO
+
+import cairosvg
+import numpy as np
+from PIL import Image
+
+sys.stdout.reconfigure(encoding="utf-8")
+
+DATASET_ROOT = r"C:\src\accidentals\dataset"
+OUTPUT_ROOT = r"C:\src\accidentals\crops"
+
+GLYPH_MAP = {"E262": "sharp", "E260": "flat", "E261": "natural"}
+
+# Regex to find accidental <use> elements in Verovio SVG
+ACCID_RE = re.compile(
+    r'class="(keyAccid|accid)"[^>]*>\s*'
+    r'<use xlink:href="#(E\d+)[^"]*"\s+'
+    r'transform="translate\((\d+),\s*(\d+)\)\s*scale\(([^,]+),\s*([^)]+)\)"'
+)
+SVG_DIM_RE = re.compile(r'width="(\d+)px" height="(\d+)px"')
+
+# Crop parameters (in SVG pixel units, viewBox/10)
+CROP_PAD_LEFT = 3
+CROP_PAD_RIGHT = 22
+CROP_PAD_TOP = 25
+CROP_PAD_BOTTOM = 35
+
+
+def setup_verovio():
+    import verovio
+
+    tk = verovio.toolkit()
+    tk.setOptions(
+        {
+            "adjustPageHeight": True,
+            "adjustPageWidth": True,
+            "header": "none",
+            "footer": "none",
+            "breaks": "none",
+            "scale": 100,
+            "pageMarginLeft": 0,
+            "pageMarginRight": 0,
+            "pageMarginTop": 0,
+            "pageMarginBottom": 0,
+        }
+    )
+    return tk
+
+
+def extract_from_mei(tk, mei_path: str) -> list[dict]:
+    """Extract accidental crops from a single MEI file.
+
+    Returns list of dicts with 'type', 'image' (PIL Image), 'context' (keyAccid/accid).
+    """
+    with open(mei_path, "r", encoding="utf-8") as f:
+        mei_data = f.read()
+
+    tk.loadData(mei_data)
+    svg = tk.renderToSVG(1)
+
+    if not svg:
+        return []
+
+    # Get SVG pixel dimensions
+    dim_match = SVG_DIM_RE.search(svg)
+    if not dim_match:
+        return []
+    svg_w = int(dim_match.group(1))
+    svg_h = int(dim_match.group(2))
+
+    if svg_w == 0 or svg_h == 0:
+        return []
+
+    # Find accidentals in SVG
+    accid_positions = []
+    for match in ACCID_RE.finditer(svg):
+        cls = match.group(1)
+        glyph_id = match.group(2)
+        tx, ty = int(match.group(3)), int(match.group(4))
+
+        glyph_type = GLYPH_MAP.get(glyph_id)
+        if glyph_type is None:
+            continue
+
+        # viewBox coords -> pixel coords (viewBox is 10x pixel dimensions)
+        px = tx / 10.0
+        py = ty / 10.0
+
+        accid_positions.append(
+            {"type": glyph_type, "context": cls, "px": px, "py": py}
+        )
+
+    if not accid_positions:
+        return []
+
+    # Rasterize SVG to PNG
+    try:
+        png_bytes = cairosvg.svg2png(
+            bytestring=svg.encode("utf-8"),
+            output_width=svg_w,
+            output_height=svg_h,
+        )
+    except Exception:
+        return []
+
+    img = Image.open(BytesIO(png_bytes)).convert("L")
+
+    # Crop each accidental
+    results = []
+    for acc in accid_positions:
+        x1 = max(0, int(acc["px"] - CROP_PAD_LEFT))
+        y1 = max(0, int(acc["py"] - CROP_PAD_TOP))
+        x2 = min(svg_w, int(acc["px"] + CROP_PAD_RIGHT))
+        y2 = min(svg_h, int(acc["py"] + CROP_PAD_BOTTOM))
+
+        if x2 - x1 < 5 or y2 - y1 < 5:
+            continue
+
+        crop = img.crop((x1, y1, x2, y2))
+        results.append(
+            {
+                "type": acc["type"],
+                "context": acc["context"],
+                "image": crop,
+            }
+        )
+
+    return results
+
+
+def main():
+    # Find all MEI files
+    mei_files = []
+    for pkg in ["package_aa", "package_ab"]:
+        mei_files.extend(
+            glob.glob(os.path.join(DATASET_ROOT, pkg, "*", "*.mei"))
+        )
+    mei_files.sort()
+
+    print(f"Found {len(mei_files)} MEI files")
+
+    # Create output directories
+    for label in GLYPH_MAP.values():
+        os.makedirs(os.path.join(OUTPUT_ROOT, label), exist_ok=True)
+
+    tk = setup_verovio()
+
+    # Process with multiple Verovio fonts for variety
+    fonts = ["Leipzig", "Bravura", "Gootville"]
+    counts = {"sharp": 0, "flat": 0, "natural": 0}
+    errors = 0
+    total_processed = 0
+
+    for font_idx, font in enumerate(fonts):
+        tk.setOptions({"font": font})
+        print(f"\n=== Font: {font} (pass {font_idx+1}/{len(fonts)}) ===")
+
+        for i, mei_path in enumerate(mei_files):
+            if i % 5000 == 0:
+                print(
+                    f"  [{font}] {i}/{len(mei_files)}  "
+                    f"sharp={counts['sharp']} flat={counts['flat']} "
+                    f"natural={counts['natural']} errors={errors}"
+                )
+
+            try:
+                results = extract_from_mei(tk, mei_path)
+            except Exception:
+                errors += 1
+                if errors <= 5:
+                    traceback.print_exc()
+                continue
+
+            sample_id = os.path.basename(os.path.dirname(mei_path))
+            for j, res in enumerate(results):
+                label = res["type"]
+                fname = f"{sample_id}_f{font_idx}_{j}.png"
+                out_path = os.path.join(OUTPUT_ROOT, label, fname)
+                res["image"].save(out_path)
+                counts[label] += 1
+
+            total_processed += 1
+
+    print(f"\n=== Done ===")
+    print(f"Processed: {total_processed} MEI files across {len(fonts)} fonts")
+    print(f"Errors: {errors}")
+    print(f"Extracted crops:")
+    for label, count in sorted(counts.items()):
+        print(f"  {label:10s}: {count:7d}")
+    print(f"Output: {OUTPUT_ROOT}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/extract_fast.py b/extract_fast.py
new file mode 100644
index 0000000..2c805ca
--- /dev/null
+++ b/extract_fast.py
@@ -0,0 +1,161 @@
+"""Fast accidental extraction from PrIMuS using Verovio + cairosvg.
+
+Optimizations:
+- Pre-filter MEI files using agnostic encoding (skip files with no accidentals)
+- Single font pass (Leipzig) - augmentation can be done at training time
+- Flush output for progress monitoring
+"""
+
+import glob
+import os
+import re
+import sys
+import time
+import traceback
+from io import BytesIO
+
+import cairosvg
+import numpy as np
+from PIL import Image
+
+sys.stdout.reconfigure(encoding="utf-8")
+
+DATASET_ROOT = r"C:\src\accidentals\dataset"
+OUTPUT_ROOT = r"C:\src\accidentals\crops"
+
+GLYPH_MAP = {"E262": "sharp", "E260": "flat", "E261": "natural"}
+
+ACCID_RE = re.compile(
+    r'class="(keyAccid|accid)"[^>]*>\s*'
+    r'<use xlink:href="#(E\d+)[^"]*"\s+'
+    r'transform="translate\((\d+),\s*(\d+)\)\s*scale\(([^,]+),\s*([^)]+)\)"'
+)
+SVG_DIM_RE = re.compile(r'width="(\d+)px" height="(\d+)px"')
+
+
+def has_accidentals(agnostic_path: str) -> bool:
+    """Quick check if an agnostic file contains any accidental tokens."""
+    with open(agnostic_path, "r", encoding="utf-8") as f:
+        return "accidental." in f.read()
+
+
+def main():
+    import verovio
+
+    # Glob all MEI files at once (much faster than iterating dirs on Windows)
+    print("Finding MEI files...", flush=True)
+    mei_files = sorted(
+        glob.glob(os.path.join(DATASET_ROOT, "package_*", "*", "*.mei"))
+    )
+    print(f"Found {len(mei_files)} MEI files (skipping pre-filter)", flush=True)
+
+    # Create output directories
+    for label in GLYPH_MAP.values():
+        os.makedirs(os.path.join(OUTPUT_ROOT, label), exist_ok=True)
+
+    # Setup Verovio
+    tk = verovio.toolkit()
+    tk.setOptions({
+        "adjustPageHeight": True,
+        "adjustPageWidth": True,
+        "header": "none",
+        "footer": "none",
+        "breaks": "none",
+        "scale": 100,
+        "pageMarginLeft": 0,
+        "pageMarginRight": 0,
+        "pageMarginTop": 0,
+        "pageMarginBottom": 0,
+        "font": "Leipzig",
+    })
+
+    counts = {"sharp": 0, "flat": 0, "natural": 0}
+    errors = 0
+    t0 = time.time()
+
+    for i, mei_path in enumerate(mei_files):
+        if i % 1000 == 0:
+            elapsed = time.time() - t0
+            rate = i / elapsed if elapsed > 0 else 0
+            eta = (len(mei_files) - i) / rate if rate > 0 else 0
+            print(
+                f"[{i:6d}/{len(mei_files)}] "
+                f"sharp={counts['sharp']} flat={counts['flat']} "
+                f"natural={counts['natural']} "
+                f"err={errors} "
+                f"rate={rate:.0f}/s eta={eta/60:.0f}m",
+                flush=True,
+            )
+
+        try:
+            with open(mei_path, "r", encoding="utf-8") as f:
+                mei_data = f.read()
+
+            tk.loadData(mei_data)
+            svg = tk.renderToSVG(1)
+            if not svg:
+                continue
+
+            dim_match = SVG_DIM_RE.search(svg)
+            if not dim_match:
+                continue
+            svg_w = int(dim_match.group(1))
+            svg_h = int(dim_match.group(2))
+            if svg_w == 0 or svg_h == 0:
+                continue
+
+            # Find accidentals
+            accids = []
+            for match in ACCID_RE.finditer(svg):
+                glyph_type = GLYPH_MAP.get(match.group(2))
+                if glyph_type:
+                    accids.append({
+                        "type": glyph_type,
+                        "px": int(match.group(3)) / 10.0,
+                        "py": int(match.group(4)) / 10.0,
+                    })
+
+            if not accids:
+                continue
+
+            # Rasterize
+            png_bytes = cairosvg.svg2png(
+                bytestring=svg.encode("utf-8"),
+                output_width=svg_w,
+                output_height=svg_h,
+            )
+            img_rgba = Image.open(BytesIO(png_bytes))
+            white_bg = Image.new("RGBA", img_rgba.size, (255, 255, 255, 255))
+            img = Image.alpha_composite(white_bg, img_rgba).convert("L")
+
+            sample_id = os.path.basename(os.path.dirname(mei_path))
+            for j, acc in enumerate(accids):
+                x1 = max(0, int(acc["px"] - 3))
+                y1 = max(0, int(acc["py"] - 25))
+                x2 = min(svg_w, int(acc["px"] + 22))
+                y2 = min(svg_h, int(acc["py"] + 35))
+                if x2 - x1 < 5 or y2 - y1 < 5:
+                    continue
+
+                crop = img.crop((x1, y1, x2, y2))
+                fname = f"{sample_id}_{j}.png"
+                crop.save(os.path.join(OUTPUT_ROOT, acc["type"], fname))
+                counts[acc["type"]] += 1
+
+        except Exception:
+            errors += 1
+            if errors <= 5:
+                traceback.print_exc()
+
+    elapsed = time.time() - t0
+    print(f"\n=== Done in {elapsed/60:.1f} minutes ===", flush=True)
+    print(f"Errors: {errors}")
+    print(f"Crops extracted:")
+    for label, count in sorted(counts.items()):
+        print(f"  {label:10s}: {count:7d}")
+    print(f"Total: {sum(counts.values())}")
+    print(f"Output: {OUTPUT_ROOT}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/model/accidental_classifier.pt b/model/accidental_classifier.pt
new file mode 100644
index 0000000..37a211d
Binary files /dev/null and b/model/accidental_classifier.pt differ
diff --git a/segment_test.py b/segment_test.py
new file mode 100644
index 0000000..7c29067
--- /dev/null
+++ b/segment_test.py
@@ -0,0 +1,122 @@
+"""Test segmenting symbols from a PrIMuS image using connected components."""
+
+import cv2
+import numpy as np
+from PIL import Image
+
+
+def segment_symbols(png_path: str):
+    """Segment a PrIMuS image into individual symbols."""
+    # Load as grayscale
+    img = cv2.imread(png_path, cv2.IMREAD_GRAYSCALE)
+    if img is None:
+        # PrIMuS uses palette mode PNGs, load via PIL first
+        pil_img = Image.open(png_path).convert("L")
+        img = np.array(pil_img)
+
+    h, w = img.shape
+    print(f"Image size: {w} x {h}")
+
+    # Binarize (black ink on white background)
+    _, binary = cv2.threshold(img, 128, 255, cv2.THRESH_BINARY_INV)
+
+    # Remove staff lines using horizontal morphology
+    # Detect staff lines
+    horizontal_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (w // 4, 1))
+    staff_lines = cv2.morphologyEx(binary, cv2.MORPH_OPEN, horizontal_kernel)
+
+    # Remove staff lines from binary image
+    no_lines = binary - staff_lines
+
+    # Clean up with small morphological close to reconnect broken symbols
+    close_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 2))
+    no_lines = cv2.morphologyEx(no_lines, cv2.MORPH_CLOSE, close_kernel)
+
+    # Find connected components
+    num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(
+        no_lines, connectivity=8
+    )
+
+    # Filter out tiny noise (area < 10 pixels)
+    symbols = []
+    for i in range(1, num_labels):  # skip background
+        x, y, sw, sh, area = stats[i]
+        if area < 10:
+            continue
+        symbols.append({
+            "x": x, "y": y, "w": sw, "h": sh,
+            "area": area, "cx": centroids[i][0],
+        })
+
+    # Sort by x position (left to right)
+    symbols.sort(key=lambda s: s["x"])
+
+    # Group nearby components into symbol clusters
+    # (a single symbol like a sharp may have multiple disconnected parts)
+    clusters = []
+    for s in symbols:
+        if clusters and s["x"] < clusters[-1]["x2"] + 5:
+            # Merge into previous cluster
+            c = clusters[-1]
+            c["x1"] = min(c["x1"], s["x"])
+            c["y1"] = min(c["y1"], s["y"])
+            c["x2"] = max(c["x2"], s["x"] + s["w"])
+            c["y2"] = max(c["y2"], s["y"] + s["h"])
+            c["components"].append(s)
+        else:
+            clusters.append({
+                "x1": s["x"], "y1": s["y"],
+                "x2": s["x"] + s["w"], "y2": s["y"] + s["h"],
+                "components": [s],
+            })
+
+    print(f"Found {len(symbols)} components -> {len(clusters)} symbol clusters")
+    for i, c in enumerate(clusters):
+        w = c["x2"] - c["x1"]
+        h = c["y2"] - c["y1"]
+        print(f"  Cluster {i:3d}: x={c['x1']:5d}-{c['x2']:5d}  y={c['y1']:3d}-{c['y2']:3d}  size={w:3d}x{h:3d}")
+
+    return clusters, img, no_lines
+
+
+def match_tokens_to_clusters(agnostic_path: str, clusters: list):
+    """Match agnostic tokens to segmented clusters."""
+    with open(agnostic_path, "r") as f:
+        tokens = f.read().strip().split("\t")
+
+    print(f"\nAgnostic tokens ({len(tokens)}):")
+    for i, tok in enumerate(tokens):
+        print(f"  {i:3d}: {tok}")
+
+    print(f"\nClusters: {len(clusters)}")
+    print(f"Tokens: {len(tokens)}")
+
+    return tokens
+
+
+if __name__ == "__main__":
+    sample_dir = r"C:\src\accidentals\dataset\package_aa\000141058-1_1_1"
+    sample_id = "000141058-1_1_1"
+
+    png_path = f"{sample_dir}/{sample_id}.png"
+    agnostic_path = f"{sample_dir}/{sample_id}.agnostic"
+
+    clusters, img, no_lines = segment_symbols(png_path)
+    tokens = match_tokens_to_clusters(agnostic_path, clusters)
+
+    # Save debug images
+    debug_dir = r"C:\src\accidentals\debug"
+    import os
+    os.makedirs(debug_dir, exist_ok=True)
+
+    cv2.imwrite(f"{debug_dir}/no_lines.png", 255 - no_lines)
+
+    # Draw bounding boxes on original
+    vis = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+    for i, c in enumerate(clusters):
+        color = (0, 0, 255)
+        cv2.rectangle(vis, (c["x1"], c["y1"]), (c["x2"], c["y2"]), color, 1)
+        cv2.putText(vis, str(i), (c["x1"], c["y1"] - 2),
+                    cv2.FONT_HERSHEY_SIMPLEX, 0.3, color, 1)
+    cv2.imwrite(f"{debug_dir}/segmented.png", vis)
+    print(f"\nDebug images saved to {debug_dir}/")
diff --git a/train.py b/train.py
new file mode 100644
index 0000000..899cad4
--- /dev/null
+++ b/train.py
@@ -0,0 +1,315 @@
+"""Train a small CNN to classify accidentals (sharp/flat/natural).
+
+Outputs an ONNX model for inference via OpenCV's dnn module.
+"""
+
+import os
+import sys
+import time
+import random
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils.data import Dataset, DataLoader, WeightedRandomSampler
+from PIL import Image
+from torchvision import transforms
+
+sys.stdout.reconfigure(encoding="utf-8")
+
+CROPS_ROOT = Path(r"C:\src\accidentals\crops")
+OUTPUT_DIR = Path(r"C:\src\accidentals\model")
+LABELS = ["flat", "natural", "sharp"]  # alphabetical → class indices 0,1,2
+IMG_SIZE = (40, 40)  # square, small enough for fast inference
+BATCH_SIZE = 256
+NUM_EPOCHS = 30
+LR = 1e-3
+SEED = 42
+VAL_SPLIT = 0.1
+TEST_SPLIT = 0.05
+MAX_PER_CLASS = 5000  # subsample: more than enough for this easy problem
+
+
+class InMemoryDataset(Dataset):
+    """Dataset that holds all images as tensors in RAM."""
+
+    def __init__(self, images: torch.Tensor, labels: torch.Tensor, augment=False):
+        self.images = images  # (N, 1, H, W) float32
+        self.labels = labels  # (N,) long
+        self.augment = augment
+        self.affine = transforms.RandomAffine(
+            degrees=5, translate=(0.1, 0.1), scale=(0.9, 1.1)
+        )
+        self.erasing = transforms.RandomErasing(p=0.2, scale=(0.02, 0.08))
+
+    def __len__(self):
+        return len(self.labels)
+
+    def __getitem__(self, idx):
+        img = self.images[idx]
+        if self.augment:
+            img = self.affine(img)
+            img = self.erasing(img)
+        return img, self.labels[idx]
+
+
+class AccidentalCNN(nn.Module):
+    """Small CNN for 3-class accidental classification.
+
+    Input: 1x40x40 grayscale image
+    Architecture: 3 conv blocks + 1 FC layer
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.features = nn.Sequential(
+            # Block 1: 1 -> 16 channels, 40x40 -> 20x20
+            nn.Conv2d(1, 16, 3, padding=1),
+            nn.BatchNorm2d(16),
+            nn.ReLU(inplace=True),
+            nn.MaxPool2d(2),
+            # Block 2: 16 -> 32 channels, 20x20 -> 10x10
+            nn.Conv2d(16, 32, 3, padding=1),
+            nn.BatchNorm2d(32),
+            nn.ReLU(inplace=True),
+            nn.MaxPool2d(2),
+            # Block 3: 32 -> 64 channels, 10x10 -> 5x5
+            nn.Conv2d(32, 64, 3, padding=1),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True),
+            nn.MaxPool2d(2),
+        )
+        self.classifier = nn.Sequential(
+            nn.Dropout(0.3),
+            nn.Linear(64 * 5 * 5, 64),
+            nn.ReLU(inplace=True),
+            nn.Dropout(0.3),
+            nn.Linear(64, len(LABELS)),
+        )
+
+    def forward(self, x):
+        x = self.features(x)
+        x = x.view(x.size(0), -1)
+        x = self.classifier(x)
+        return x
+
+
+def load_all_into_ram():
+    """Load all images into RAM as tensors. Returns (images, labels) tensors."""
+    print("Loading all images into RAM...", flush=True)
+    all_images = []
+    all_labels = []
+
+    for label_idx, label_name in enumerate(LABELS):
+        label_dir = CROPS_ROOT / label_name
+        files = [f for f in os.listdir(label_dir) if f.endswith(".png")]
+        random.shuffle(files)
+        if MAX_PER_CLASS and len(files) > MAX_PER_CLASS:
+            files = files[:MAX_PER_CLASS]
+        print(f"  {label_name}: {len(files)} files", flush=True)
+
+        for fname in files:
+            img = Image.open(label_dir / fname).convert("L")
+            img = img.resize(IMG_SIZE, Image.BILINEAR)
+            arr = np.array(img, dtype=np.float32) / 255.0
+            all_images.append(arr)
+            all_labels.append(label_idx)
+
+    # Stack into tensors: (N, 1, H, W)
+    images = torch.tensor(np.array(all_images)).unsqueeze(1)
+    labels = torch.tensor(all_labels, dtype=torch.long)
+    print(f"  Total: {len(labels)} images, tensor shape: {images.shape}", flush=True)
+    print(f"  Memory: {images.nbytes / 1024**2:.0f} MB", flush=True)
+    return images, labels
+
+
+def split_data(images, labels):
+    """Stratified train/val/test split. Returns index arrays."""
+    indices = torch.randperm(len(labels))
+
+    by_class = {i: [] for i in range(len(LABELS))}
+    for idx in indices:
+        by_class[labels[idx].item()].append(idx.item())
+
+    train_idx, val_idx, test_idx = [], [], []
+    for cls, idxs in by_class.items():
+        n = len(idxs)
+        n_test = max(1, int(n * TEST_SPLIT))
+        n_val = max(1, int(n * VAL_SPLIT))
+        test_idx.extend(idxs[:n_test])
+        val_idx.extend(idxs[n_test : n_test + n_val])
+        train_idx.extend(idxs[n_test + n_val :])
+
+    return (
+        torch.tensor(train_idx),
+        torch.tensor(val_idx),
+        torch.tensor(test_idx),
+    )
+
+
+def make_weighted_sampler(labels):
+    """Create a sampler that oversamples minority classes."""
+    class_counts = torch.bincount(labels, minlength=len(LABELS)).float()
+    weights = 1.0 / class_counts[labels]
+    return WeightedRandomSampler(weights, len(labels), replacement=True)
+
+
+def evaluate(model, loader, device):
+    """Evaluate model on a data loader."""
+    model.eval()
+    correct = 0
+    total = 0
+    class_correct = np.zeros(len(LABELS))
+    class_total = np.zeros(len(LABELS))
+    total_loss = 0.0
+    n_batches = 0
+
+    with torch.no_grad():
+        for imgs, lbls in loader:
+            imgs, lbls = imgs.to(device), lbls.to(device)
+            outputs = model(imgs)
+            loss = F.cross_entropy(outputs, lbls)
+            total_loss += loss.item()
+            n_batches += 1
+
+            _, predicted = outputs.max(1)
+            total += lbls.size(0)
+            correct += predicted.eq(lbls).sum().item()
+
+            for i in range(len(LABELS)):
+                mask = lbls == i
+                class_total[i] += mask.sum().item()
+                class_correct[i] += (predicted[mask] == i).sum().item()
+
+    acc = correct / total if total > 0 else 0
+    class_acc = np.where(class_total > 0, class_correct / class_total, 0)
+    avg_loss = total_loss / n_batches if n_batches > 0 else 0
+    return avg_loss, acc, class_acc
+
+
+def main():
+    random.seed(SEED)
+    np.random.seed(SEED)
+    torch.manual_seed(SEED)
+
+    device = torch.device("cpu")
+    print(f"Device: {device}", flush=True)
+
+    # Load everything into RAM
+    images, labels = load_all_into_ram()
+
+    # Split
+    train_idx, val_idx, test_idx = split_data(images, labels)
+    print(f"\nTrain: {len(train_idx)}  Val: {len(val_idx)}  Test: {len(test_idx)}", flush=True)
+    for i, name in enumerate(LABELS):
+        nt = (labels[train_idx] == i).sum().item()
+        nv = (labels[val_idx] == i).sum().item()
+        ne = (labels[test_idx] == i).sum().item()
+        print(f"  {name:8s}: train={nt:6d} val={nv:5d} test={ne:5d}", flush=True)
+
+    # Create datasets (in-memory, fast)
+    train_ds = InMemoryDataset(images[train_idx], labels[train_idx], augment=True)
+    val_ds = InMemoryDataset(images[val_idx], labels[val_idx], augment=False)
+    test_ds = InMemoryDataset(images[test_idx], labels[test_idx], augment=False)
+
+    # Weighted sampler for class imbalance
+    sampler = make_weighted_sampler(labels[train_idx])
+
+    train_loader = DataLoader(train_ds, batch_size=BATCH_SIZE, sampler=sampler, num_workers=0)
+    val_loader = DataLoader(val_ds, batch_size=BATCH_SIZE, shuffle=False, num_workers=0)
+    test_loader = DataLoader(test_ds, batch_size=BATCH_SIZE, shuffle=False, num_workers=0)
+
+    # Model
+    model = AccidentalCNN().to(device)
+    param_count = sum(p.numel() for p in model.parameters())
+    print(f"\nModel parameters: {param_count:,}", flush=True)
+
+    optimizer = torch.optim.Adam(model.parameters(), lr=LR)
+    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
+        optimizer, mode="max", factor=0.5, patience=3
+    )
+
+    # Training loop
+    best_val_acc = 0.0
+    OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
+    best_model_path = OUTPUT_DIR / "accidental_classifier.pt"
+
+    print(f"\n{'Epoch':>5s}  {'TrLoss':>7s}  {'VaLoss':>7s}  {'VaAcc':>6s}  "
+          f"{'Flat':>5s}  {'Nat':>5s}  {'Sharp':>5s}  {'LR':>8s}  {'Time':>5s}", flush=True)
+    print("-" * 70, flush=True)
+
+    for epoch in range(1, NUM_EPOCHS + 1):
+        t0 = time.time()
+        model.train()
+        train_loss = 0.0
+        n_batches = 0
+
+        for imgs, lbls in train_loader:
+            imgs, lbls = imgs.to(device), lbls.to(device)
+            optimizer.zero_grad()
+            outputs = model(imgs)
+            loss = F.cross_entropy(outputs, lbls)
+            loss.backward()
+            optimizer.step()
+            train_loss += loss.item()
+            n_batches += 1
+
+        avg_train_loss = train_loss / n_batches
+        val_loss, val_acc, class_acc = evaluate(model, val_loader, device)
+        scheduler.step(val_acc)
+
+        elapsed = time.time() - t0
+        lr = optimizer.param_groups[0]["lr"]
+
+        print(
+            f"{epoch:5d}  {avg_train_loss:7.4f}  {val_loss:7.4f}  {val_acc:6.1%}  "
+            f"{class_acc[0]:5.1%}  {class_acc[1]:5.1%}  {class_acc[2]:5.1%}  "
+            f"{lr:.6f}  {elapsed:5.1f}s",
+            flush=True,
+        )
+
+        if val_acc > best_val_acc:
+            best_val_acc = val_acc
+            torch.save(model.state_dict(), best_model_path)
+
+    # Final evaluation on test set
+    print(f"\n=== Test Set Evaluation ===", flush=True)
+    model.load_state_dict(torch.load(best_model_path, weights_only=True))
+    test_loss, test_acc, test_class_acc = evaluate(model, test_loader, device)
+    print(f"Test accuracy: {test_acc:.1%}", flush=True)
+    for i, name in enumerate(LABELS):
+        print(f"  {name:8s}: {test_class_acc[i]:.1%}", flush=True)
+
+    # Export to ONNX for OpenCV dnn inference
+    print(f"\nExporting to ONNX...", flush=True)
+    model.eval()
+    dummy = torch.randn(1, 1, *IMG_SIZE)
+    onnx_path = OUTPUT_DIR / "accidental_classifier.onnx"
+    torch.onnx.export(
+        model,
+        dummy,
+        str(onnx_path),
+        input_names=["image"],
+        output_names=["logits"],
+        dynamic_axes={"image": {0: "batch"}, "logits": {0: "batch"}},
+        opset_version=13,
+    )
+    print(f"Saved ONNX model: {onnx_path}", flush=True)
+    print(f"Saved PyTorch model: {best_model_path}", flush=True)
+
+    # Quick ONNX verification via OpenCV
+    import cv2
+    net = cv2.dnn.readNetFromONNX(str(onnx_path))
+    test_img = np.random.randint(0, 255, (IMG_SIZE[1], IMG_SIZE[0]), dtype=np.uint8)
+    blob = cv2.dnn.blobFromImage(test_img, 1.0 / 255.0, IMG_SIZE, 0, swapRB=False)
+    net.setInput(blob)
+    out = net.forward()
+    pred = LABELS[np.argmax(out[0])]
+    print(f"ONNX+OpenCV smoke test: input=random noise -> pred={pred} (expected: arbitrary)", flush=True)
+    print(f"\nDone. Labels: {LABELS}", flush=True)
+
+
+if __name__ == "__main__":
+    main()