The RC-110's top channel is an overflow bin: every event above the last
calibrated energy is counted there, so it is not spectral content at that
energy. It shows up as a single-channel spike around a hundred times its
neighbours - 186 counts of 91224 in the indoor background capture, and 618 of
81225 (0.76% of all counts, in one channel) in a new 2.5 h forest background.
No template can reproduce a one-channel spike, so leaving it in the Poisson
likelihood does not invent a nuclide - it hides them. The only way the fit can
cover the spike is by lifting the topmost continuum bump, and that raised
continuum then absorbs genuine counts across the whole high-energy region.
Measured, toggling FIT_OVERFLOW_CHANNELS 0/1 over the real captures:
capture residual reduced chi2 overflow bin
background 6.8 -> 6.4% 2.18 -> 1.69 0.20% of counts
forest 9.7 -> 7.2% 3.72 -> 1.28 0.76% of counts
thorium 11.7 -> 11.5% 1.94 -> 1.84 -
uranium 8.7 -> 8.5% 2.01 -> 1.84 -
Goodness-of-fit improves on all four and by far the most on the capture with
the largest artefact, which is what ties the effect to this cause. Verdicts
and time-to-identification are unchanged: background 900 s, thorium 120 s,
uranium 300 s, forest 3600 s.
Significance moves modestly in both directions - K-40 rises 3.7 -> 4.6 sigma
on the forest capture but falls 6.8 -> 6.2 on the indoor one. The change is
justified by the artefact being an artefact and by the fit becoming
well-specified (chi2 1.3-1.8 where it was 1.9-3.7), not by larger numbers.
The post-fix figures are the trustworthy ones precisely because the model now
fits.
The peak search already skipped this channel; the template fit, which is what
actually decides identification now, did not.
Also adds:
- host/spectra/forest.csv - a real 9002 s / 81225 count outdoor background
from a forest, 9.0 cps against the indoor capture's 21.7, now a permanent
regression case for the low-rate end.
- tools/xml_to_replay_csv.py - converts a RadiaCode XML export into the
replay harness format. tools/radiacode_to_background.py writes
counts/second references for the device's background library, which replay
cannot consume, so there was no way to get an XML export under test.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>