ShootFlow scores RAW stills (sharpness, exposure, framing) from the JPEG preview the camera embeds in the file, so a photographer can cull a card of .CR2 or .DNG files without a RAW developer. The reader is deliberately simple: find every JPEG stream in the file, walk its segments to the end marker, keep the one with the most pixels. Thumbnails lose to the preview.
That rule has a hole. Canon's CR2 stores the sensor data itself as a lossless JPEG (frame type SOF3, 12 to 14 bits per sample), and DNG files usually store their sensor tiles or strips the same way. Those are real, well-formed JPEG streams. When one claims more pixels than the preview, it wins, and ffmpeg will happily decode it, so the cull would have scored the raw Bayer mosaic rather than the picture, with no error to say so.
We reproduced it with a real lossless stream from ffmpeg's ljpeg encoder (1280x960) sitting beside a 640x480 baseline preview: the old reader returned the lossless stream. The fix keeps only frames a camera writes for display: 8-bit baseline, extended or progressive JPEG. A file whose only JPEGs are lossless or high-precision now reports "no embedded JPEG preview" and keeps its stand-in score, which is the honest outcome.
What this is not: a test on real camera files. raw.pixls.us and rawsamples.ch refuse connections from our build machines, and the small RAW samples in exiftool's test suite have their previews stripped. The format facts are well documented, but the next check is a real CR2 and a real DNG. Code and tests: ventures/photo-editing/src/shootflow/raw_preview.py.