diff --git a/examples/decode_capture.py b/examples/decode_capture.py
new file mode 100644
index 0000000..c5d16df
--- /dev/null
+++ b/examples/decode_capture.py
@@ -0,0 +1,210 @@
+#!/usr/bin/env python3
+"""Decode RYLR998 LoRa frames from wideband SDR captures.
+
+This script processes real BladeRF captures at 2 MHz sample rate,
+channelizes to 125 kHz, and decodes LoRa frames.
+
+Usage:
+    python decode_capture.py <capture_file.raw>
+    python decode_capture.py --list  # List available captures
+"""
+
+import argparse
+import sys
+from pathlib import Path
+
+import numpy as np
+
+# Add parent to path for local development
+sys.path.insert(0, str(Path(__file__).parent.parent / "python"))
+
+from rylr998 import (
+    Channelizer,
+    FrameSync,
+    PHYDecode,
+    sync_word_to_networkid,
+)
+
+
+# Capture parameters (BladeRF defaults)
+INPUT_SAMPLE_RATE = 2e6      # 2 MHz capture rate
+CENTER_FREQ = 915e6          # Center frequency
+CHANNEL_FREQ = 915e6         # LoRa channel frequency
+CHANNEL_BW = 125e3           # LoRa bandwidth
+
+# LoRa parameters (RYLR998 defaults)
+SF = 9
+CR = 1
+
+
+def decode_capture(
+    capture_path: Path,
+    sf: int = SF,
+    verbose: bool = True,
+) -> list:
+    """Decode LoRa frames from a wideband capture file.
+
+    Args:
+        capture_path: Path to .raw capture file (complex64)
+        sf: Spreading factor
+        verbose: Print progress
+
+    Returns:
+        List of decoded frames
+    """
+    # Load capture
+    if verbose:
+        print(f"Loading {capture_path.name}...")
+    iq_raw = np.fromfile(capture_path, dtype=np.complex64)
+
+    if len(iq_raw) == 0:
+        print("  ERROR: Empty file")
+        return []
+
+    # Check for NaN/Inf
+    if np.any(~np.isfinite(iq_raw)):
+        nan_count = np.sum(~np.isfinite(iq_raw))
+        print(f"  WARNING: {nan_count} NaN/Inf values ({100*nan_count/len(iq_raw):.1f}%)")
+        # Replace NaN with zeros
+        iq_raw = np.nan_to_num(iq_raw, nan=0.0, posinf=0.0, neginf=0.0)
+
+    duration_s = len(iq_raw) / INPUT_SAMPLE_RATE
+    if verbose:
+        print(f"  Loaded {len(iq_raw):,} samples ({duration_s:.1f}s at {INPUT_SAMPLE_RATE/1e6:.1f} MHz)")
+
+    # Channelize to LoRa bandwidth
+    if verbose:
+        print(f"Channelizing to {CHANNEL_BW/1e3:.0f} kHz...")
+    channelizer = Channelizer(
+        input_sample_rate=INPUT_SAMPLE_RATE,
+        channel_bw=CHANNEL_BW,
+        center_freq=CENTER_FREQ,
+        channel_freq=CHANNEL_FREQ,
+    )
+    iq_ch = channelizer.channelize(iq_raw)
+    if verbose:
+        print(f"  Channelized: {len(iq_ch):,} samples ({channelizer})")
+
+    # Frame synchronization
+    if verbose:
+        print(f"Searching for LoRa frames (SF{sf})...")
+
+    sync = FrameSync(sf=sf)
+    frames = []
+
+    # Process in chunks (1 second at a time)
+    chunk_size = int(CHANNEL_BW)  # 1 second of channelized data
+    n_chunks = (len(iq_ch) + chunk_size - 1) // chunk_size
+
+    decoder = PHYDecode(sf=sf)
+
+    for i in range(n_chunks):
+        start = i * chunk_size
+        end = min((i + 1) * chunk_size, len(iq_ch))
+        chunk = iq_ch[start:end]
+
+        result = sync.sync_from_samples(chunk)
+        if result.found:
+            # Extract NETWORKID
+            nid = result.networkid
+
+            if verbose:
+                print(f"\n  Frame detected at chunk {i} ({start/CHANNEL_BW:.1f}s):")
+                print(f"    CFO: {result.cfo_bin:.2f} bins")
+                print(f"    Sync word raw: {result.sync_word_raw}")
+                print(f"    NETWORKID: {nid}")
+                print(f"    Data symbols: {len(result.data_symbols)}")
+
+            # Decode
+            cfo_int = int(round(result.cfo_bin))
+            frame = decoder.decode(
+                result.data_symbols,
+                cfo_bin=cfo_int,
+                use_grlora_gray=True,   # Real captures use gr-lora_sdr convention
+                soft_decoding=False,    # Real captures need -1 offset
+            )
+
+            if verbose:
+                print(f"    Header OK: {frame.header_ok}")
+                print(f"    Payload len: {frame.payload_length}")
+                print(f"    CRC OK: {frame.crc_ok}")
+                if frame.payload:
+                    try:
+                        text = frame.payload.decode('utf-8', errors='replace')
+                        print(f"    Payload: {repr(text)}")
+                    except Exception:
+                        print(f"    Payload (hex): {frame.payload.hex()}")
+
+            frames.append({
+                'time': start / CHANNEL_BW,
+                'networkid': nid,
+                'cfo_bin': result.cfo_bin,
+                'frame': frame,
+            })
+
+            # Reset sync for next frame
+            sync.reset()
+
+    if verbose:
+        print(f"\n{'='*60}")
+        print(f"Total frames found: {len(frames)}")
+
+    return frames
+
+
+def list_captures():
+    """List available capture files."""
+    logs_dir = Path(__file__).parent.parent.parent / "gnuradio" / "logs"
+    if not logs_dir.exists():
+        print(f"Logs directory not found: {logs_dir}")
+        return
+
+    print(f"Capture files in {logs_dir}:")
+    print("-" * 60)
+
+    for raw_file in sorted(logs_dir.glob("*.raw")):
+        size_mb = raw_file.stat().st_size / 1e6
+        n_samples = raw_file.stat().st_size // 8  # complex64 = 8 bytes
+        duration_s = n_samples / INPUT_SAMPLE_RATE
+        print(f"  {raw_file.name:<45} {size_mb:>7.1f} MB  ({duration_s:.1f}s)")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Decode LoRa frames from SDR captures")
+    parser.add_argument("capture", nargs="?", help="Capture file path")
+    parser.add_argument("--list", action="store_true", help="List available captures")
+    parser.add_argument("--sf", type=int, default=SF, help=f"Spreading factor (default: {SF})")
+    parser.add_argument("--quiet", "-q", action="store_true", help="Less verbose output")
+    args = parser.parse_args()
+
+    if args.list:
+        list_captures()
+        return
+
+    if not args.capture:
+        parser.print_help()
+        print("\n\nTip: Use --list to see available capture files")
+        return
+
+    capture_path = Path(args.capture)
+    if not capture_path.exists():
+        # Try looking in the logs directory
+        logs_dir = Path(__file__).parent.parent.parent / "gnuradio" / "logs"
+        alt_path = logs_dir / args.capture
+        if alt_path.exists():
+            capture_path = alt_path
+        else:
+            print(f"ERROR: File not found: {capture_path}")
+            sys.exit(1)
+
+    frames = decode_capture(capture_path, sf=args.sf, verbose=not args.quiet)
+
+    if not frames:
+        print("\nNo frames decoded. Try:")
+        print("  - Different SF (--sf 7 through --sf 12)")
+        print("  - Check capture frequency matches RYLR998 setting")
+        print("  - Ensure RYLR998 is transmitting")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/python/rylr998/__init__.py b/python/rylr998/__init__.py
index f32d02b..1c96b7b 100644
--- a/python/rylr998/__init__.py
+++ b/python/rylr998/__init__.py
@@ -26,6 +26,7 @@ from .phy_decode import PHYDecode, LoRaFrame
 from .phy_encode import PHYEncode
 from .frame_sync import FrameSync
 from .frame_gen import FrameGen
+from .channelizer import Channelizer, channelize
 
 __version__ = "0.1.0"
 __all__ = [
@@ -39,4 +40,6 @@ __all__ = [
     "FrameSync",
     "FrameGen",
     "LoRaFrame",
+    "Channelizer",
+    "channelize",
 ]
diff --git a/python/rylr998/channelizer.py b/python/rylr998/channelizer.py
new file mode 100644
index 0000000..90a5390
--- /dev/null
+++ b/python/rylr998/channelizer.py
@@ -0,0 +1,179 @@
+"""Channelizer: frequency translation and decimation for wideband captures.
+
+When captures are taken at a higher sample rate than the LoRa bandwidth (e.g.,
+2 MHz capture of a 125 kHz LoRa signal), the signal must be:
+    1. Frequency-shifted to baseband
+    2. Low-pass filtered (anti-aliasing)
+    3. Decimated to match the LoRa bandwidth
+
+This module provides the channelization step needed before FrameSync.
+"""
+
+import numpy as np
+from dataclasses import dataclass
+
+
+@dataclass
+class ChannelizerConfig:
+    """Configuration for channelizer."""
+    input_sample_rate: float    # Hz - rate of incoming samples
+    channel_bw: float           # Hz - LoRa bandwidth (typically 125e3)
+    center_freq: float = 0.0    # Hz - center frequency of capture
+    channel_freq: float = 0.0   # Hz - frequency of LoRa channel
+    n_taps: int = 0             # FIR filter taps (0 = auto-calculate)
+
+
+class Channelizer:
+    """Frequency translation and decimation for wideband captures.
+
+    Shifts a LoRa channel to baseband with proper anti-alias filtering.
+    Uses a windowed-sinc FIR filter instead of a moving average.
+
+    Example:
+        >>> ch = Channelizer(input_sample_rate=2e6, channel_bw=125e3)
+        >>> baseband = ch.channelize(wideband_iq)
+        >>> # baseband is now at 125 kHz sample rate
+    """
+
+    def __init__(
+        self,
+        input_sample_rate: float,
+        channel_bw: float,
+        center_freq: float = 0.0,
+        channel_freq: float = 0.0,
+        n_taps: int = 0,
+    ):
+        """Initialize channelizer.
+
+        Args:
+            input_sample_rate: Sample rate of input data (Hz)
+            channel_bw: LoRa bandwidth / output sample rate (Hz)
+            center_freq: Center frequency of capture (Hz)
+            channel_freq: Frequency of LoRa channel (Hz)
+            n_taps: FIR filter taps (0 = auto: 4 * decimation + 1)
+        """
+        self.input_sample_rate = input_sample_rate
+        self.channel_bw = channel_bw
+        self.center_freq = center_freq
+        self.channel_freq = channel_freq
+
+        # Calculate decimation factor
+        self.decim = max(1, int(input_sample_rate / channel_bw))
+
+        # FIR filter design
+        if n_taps <= 0:
+            n_taps = self.decim * 4 + 1  # odd length for type-I linear phase
+        self.n_taps = n_taps
+
+        # Normalized cutoff (Nyquist = 1.0)
+        cutoff = channel_bw / input_sample_rate
+
+        # Design the filter
+        self._fir = self._design_lowpass(n_taps, cutoff)
+
+        # Precompute frequency offset
+        self._freq_offset = channel_freq - center_freq
+
+    def _design_lowpass(self, n_taps: int, cutoff: float) -> np.ndarray:
+        """Design windowed-sinc lowpass filter.
+
+        Args:
+            n_taps: Number of filter taps (should be odd)
+            cutoff: Normalized cutoff frequency (0 to 1, where 1 = Nyquist)
+
+        Returns:
+            FIR filter coefficients
+        """
+        try:
+            from scipy.signal import firwin
+            return firwin(n_taps, cutoff).astype(np.float32)
+        except ImportError:
+            # Fallback: simple windowed sinc if scipy not available
+            n = np.arange(n_taps) - (n_taps - 1) / 2
+            # Avoid division by zero
+            with np.errstate(divide='ignore', invalid='ignore'):
+                h = np.where(n == 0, 2 * cutoff,
+                            np.sin(2 * np.pi * cutoff * n) / (np.pi * n))
+            # Hamming window
+            window = 0.54 - 0.46 * np.cos(2 * np.pi * np.arange(n_taps) / (n_taps - 1))
+            fir = (h * window).astype(np.float32)
+            return fir / np.sum(fir)  # Normalize
+
+    def channelize(
+        self,
+        iq_data: np.ndarray,
+        freq_offset: float | None = None,
+    ) -> np.ndarray:
+        """Channelize wideband IQ data to baseband at channel bandwidth.
+
+        Args:
+            iq_data: Complex IQ samples at input_sample_rate
+            freq_offset: Override frequency offset (Hz). If None, uses
+                        channel_freq - center_freq from init.
+
+        Returns:
+            Complex IQ samples at channel_bw sample rate
+        """
+        if freq_offset is None:
+            freq_offset = self._freq_offset
+
+        n = len(iq_data)
+
+        # Step 1: Frequency shift to baseband
+        if abs(freq_offset) > 0.01:  # Only shift if significant offset
+            t = np.arange(n, dtype=np.float64) / self.input_sample_rate
+            shifted = iq_data * np.exp(-1j * 2 * np.pi * freq_offset * t).astype(np.complex64)
+        else:
+            shifted = iq_data.astype(np.complex64)
+
+        # Step 2 & 3: Filter and decimate
+        if self.decim <= 1:
+            return shifted
+
+        # Convolve with FIR filter
+        filtered = np.convolve(shifted, self._fir, mode="same")
+
+        # Decimate
+        return filtered[::self.decim]
+
+    @property
+    def output_sample_rate(self) -> float:
+        """Output sample rate after channelization."""
+        return self.input_sample_rate / self.decim
+
+    def __repr__(self) -> str:
+        return (f"Channelizer(in={self.input_sample_rate/1e3:.0f}kHz, "
+                f"out={self.output_sample_rate/1e3:.0f}kHz, "
+                f"decim={self.decim}, taps={self.n_taps})")
+
+
+def channelize(
+    iq_data: np.ndarray,
+    input_sample_rate: float,
+    center_freq: float,
+    channel_freq: float,
+    channel_bw: float,
+) -> np.ndarray:
+    """Convenience function to channelize wideband data.
+
+    Args:
+        iq_data: Complex IQ samples at input_sample_rate
+        input_sample_rate: Sample rate of input data (Hz)
+        center_freq: Center frequency of capture (Hz)
+        channel_freq: Frequency of LoRa channel (Hz)
+        channel_bw: LoRa bandwidth (Hz)
+
+    Returns:
+        Complex IQ samples at channel_bw sample rate
+
+    Example:
+        >>> # 2 MHz capture, LoRa at 915 MHz, 125 kHz bandwidth
+        >>> baseband = channelize(iq, 2e6, 915e6, 915e6, 125e3)
+    """
+    ch = Channelizer(
+        input_sample_rate=input_sample_rate,
+        channel_bw=channel_bw,
+        center_freq=center_freq,
+        channel_freq=channel_freq,
+    )
+    return ch.channelize(iq_data)
diff --git a/python/rylr998/frame_sync.py b/python/rylr998/frame_sync.py
index 680df6e..8002743 100644
--- a/python/rylr998/frame_sync.py
+++ b/python/rylr998/frame_sync.py
@@ -125,10 +125,12 @@ class FrameSync:
 
         Preamble chirps should have:
         - Strong FFT peak (high SNR)
-        - Bin very close to 0 (or consistent with CFO)
+        - Bin consistent with previous preamble chirps (if any)
 
-        The tolerance must be tight enough to distinguish preamble (bin ~0)
-        from sync word symbols which can be as low as 8 for RYLR998.
+        Real SDR captures can have significant CFO (carrier frequency offset),
+        so the preamble bin can appear anywhere in 0..N-1, not just near 0.
+        The key insight is that preamble chirps have the SAME bin value
+        (modulo small noise) for many consecutive symbols.
         """
         if peak_mag < 3.0:  # Minimum SNR threshold
             return False
@@ -137,14 +139,16 @@ class FrameSync:
         if self._preamble_count > 0:
             expected_bin = int(round(self._cfo_estimate)) % self.N
             # Tight tolerance: must be within 3 bins of expected
-            # This distinguishes preamble (bin ~0) from sync word (bin >= 8)
+            # This distinguishes preamble (bin ~0+CFO) from sync word (bin >= 8+CFO)
             tolerance = 3
             distance = min(abs(peak_bin - expected_bin),
                            self.N - abs(peak_bin - expected_bin))
             return distance <= tolerance
         else:
-            # First preamble chirp - accept bins close to 0 or N-1 (CFO)
-            return peak_bin < 4 or peak_bin > self.N - 4
+            # First preamble chirp - accept ANY strong signal
+            # Real captures have arbitrary CFO, so preamble can appear at any bin
+            # We'll validate by checking if subsequent chirps have the same bin
+            return True
 
     def _is_downchirp(self, samples: NDArray[np.complex64]) -> tuple[bool, float]:
         """Detect if samples contain a downchirp (SFD).