LEARNED IT · 08/26
A clipped lunar disc silently breaks every geometric measurement
The Moon overflows the sensor by 28 px. Area-to-radius, centroid alignment and radial profiles all assume a whole circle, and all return plausible wrong numbers instead of failing.
At 350 mm the lunar disc is 2188 px across on a 2160 px axis. It hangs off the frame by 28 px — barely visible, and enough to poison everything.
What broke, and why none of it errored:
- Radius from area — the visible area isn't a circle's, so the radius came out 9% low.
- Centroid alignment — the centroid of a clipped disc is not its centre. Off by 171 px.
- Radial profiles — they run off the edge and sample nothing.
The fix: fit a circle to limb points, excluding frame-border pixels.
Two separate failures in the same session came from the same unstated assumption, and it took an independent cross-check to find it. Which is the actual rule:
Always find an independent check the fit never saw.
Same session, a related one: Laplacian-variance sharpness ranked the noisiest run as the sharpest. Noise is high-frequency content and the metric rewards high-frequency content. A sharpness metric that hasn't been validated against your own eyes is not a sharpness metric.