ShootFlow scores a frame for sharpness, grain, exposure and tilt, but its thresholds were set on synthetic patches. A new script, tools/real_photo_calibration.py, takes real photographs, adds a known blur, noise, rotation or brightness change, and prints what the analyser reads back.
On two OpenCV sample photos (a building and a butterfly): sharpness falls steadily as blur grows (butterfly 0.98 to 0.04 over 0 to 8 px) and does not mistake grain for detail. The grain estimate rises with injected noise but under-reads, 8.7 for a true 32 on the building, probably because of JPEG compression. Exposure peaks at the original and falls both ways.
Tilt failed. Rotating the photos 0 to 8 degrees did not move the estimate in any usable way, and its confidence stayed at 0.04 to 0.10. The estimator needs long straight edges; a butterfly has none and a building has perspective lines at many angles. Because the story penalty is scaled by confidence, nothing is wrongly penalised, but crooked frames go uncaught.
Limits: two photos, no human picks, no photo with a real horizon. The next step is to repeat this on a sea or table-edge shot and try a stronger line detector.