The per-cell fit in ApplyCellSurface dropped every observation with Is <= 0.
Both of its sums are linear in Is - there is no logarithm and no division by
it - so the filter protected nothing arithmetic; the only hazard of keeping
the negatives is a cell whose `cross` sums to <= 0, which now keeps its
factor for that round instead of dividing by it.
What the filter did was select on the noise: it kept the up-fluctuated half
of a weak cell, inflating `cross` by the truncated mean, so each cell was
pulled down by an amount set by its own signal-to-noise. Per round, on a
Wilson-distributed cell with six equivalents: 0.1% at <I>/sigma = 5, 1.0% at
2, 4.9% at 1, 15% at 0.5. That push is almost entirely along the surface's
resolution null direction, which is why the surfaces walked.
Simulated with a flat truth and a 12% cos(2 phi) flat field:
- no gauge: shell at I/sigma = 1 suppressed 40% after 30 rounds with the
filter, flat to 1% without it;
- 16x16 detector grid, gauge on, outer I/sigma = 0.5: with the filter the
fitted field carries 1.9x its true amplitude at CC 0.48 to the truth and
the gauge removes a 5x ramp; without it 1.04x, CC 0.97, 1.09x.
On 12 open rotation sets against the gauge build (rc172-2):
- the null walk the gauge removes collapses where it was large (45.9x ->
2.0x, 5.0x -> 2.4x, 3.1x -> 1.7x; a 56x and a 4.9e5x walk are gone because
those surfaces no longer pass cross-validation);
- held-out gains lose the part the filter inflated (23.2 -> 20.1%,
14.1 -> 12.9%);
- sets at I/sigma >= 10: per-shell CC(Fo, Fc) against the deposited model
moves by -0.006..+0.003 (fine half), ISa unchanged;
- the two weakest sets (I/sigma 6.8 and 4.6) lose a surface the gate no
longer admits, and with it 0.019 / 0.031 of fine-half CC(Fo, Fc).
Forcing the unfiltered surface on the first recovers all of it, so there
it is the gate that refuses a real correction; on the second the
unfiltered surface is worth nothing and the filtered one's benefit is
not a correction the fit can reproduce.
"NOT settled in 30 rounds" does not go away (10 of 17 applied surfaces ->
8 of 15): it has a second cause the filter was not driving.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013nW6FNRP1bBJJ8pfHiByAT
leonarski_fandClaude Opus 5
1d1a40d908