Pass A's background-free centroid divided the disk's first moment by
net = I_sum - n*bkg whenever net > 0. An integer sum against a rational
background reaches zero to rounding (net ~ 1e-14), and any small net gives a
position off the disk: the first moment's noise is radius x sigma, so the
centroid's is radius x sigma/net. Below sigma the raw centroid is taken, as it
already was for net <= 0. Both engines; Metal shares the CPU code.
XDET/YDET up to 1e16 in the unmerged MTZ was the visible part: 67 of 68
rotation sets scanned had rows off the detector. Their I is right (profile fit
and box sum are centred on the prediction; z against symmetry mates has rms
0.92 vs 1.04 for the control).
The real harm was in post-refinement. Such a centroid can reach the strongest
20000 positions, where the held-out cost is a plain mean with a noise from the
fourth moment: one outlier inflates the noise and the cross-validation refuses
a fit that the paired change supports at -17 sd. Measured, --no-gpu, this commit
against its parent, 14 sets: 5 change their verdicts and so their geometry;
the 9 others have merged MTZs bitwise identical. On the half-image in-house
lysozyme: joint noise 6.8e-8 -> 2.4e-8, reject -> COMMIT, cell a 78.44 -> 78.35
(XDS 77.74), R_meas 56.0 -> 46.8 %, ISa 4.3 -> 4.7. Open-arm 8c3e: 1.74 -> 1.67 A.
Small-molecule L-alanine single sweep commits too, ISa 57.3 -> 55.6.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HVeDQs733XDnbQQyUbV1vs