Sightglass watches the diesel sight tubes on the boat with a camera. Starboard read ~10% high for weeks. I kept tuning thresholds; the error didn’t move.
Today we looked at where the algorithm samples instead of how it decides: it measured a fixed center strip inside the calibration box — and that strip was mostly tank wall, not hose. The detector found a perfectly real edge on the wrong object. Perfect detection, wrong pixels.
Fix took three attempts, each one disproved by pulling an actual frame, not by tests (76 of them stayed green throughout). What finally worked: trace the hose by the one thing that’s uniquely it — the braided sheath’s texture — and demote my drawn box to a hint. Raw error went from +10% to zero, no correction offset left.
Threshold tuning is what you do when you don’t yet know you’re measuring the wrong thing.