Verified against three real sources on hardware, plus a 52-check regression
suite that runs on the device from synthetic spectra.
Identification
- Model the spectral features that are not gamma lines: backscatter, Compton
edge, single and double escape peaks, sum peaks, and 511 keV annihilation.
The matcher assumed every peak was a line and handed these to whichever
nuclide sat nearest, which was the single cause behind phantom Na-22,
Ba-133 and Am-241 rather than three separate faults.
- A nuclide identified entirely below 110 keV must clear 9 sigma: that region
is unmodelled X-ray fluorescence and backscatter, and a 4.9 sigma bump at
63 keV (really Th-234) was being reported as Am-241 at 100%.
- Decay chains: two members with one line each corroborate one another, and a
single member matched at 8 sigma can carry its chain. Requiring one nuclide
to show two lines lost a thorium mantle whose Pb-212 was unmistakable.
- Name the dominant chain rather than reporting a mixture. At this resolution
both natural chains fit the same peaks - their line ratios differ by 0.4%,
far below what 8.4% FWHM resolves - so real ore read 97% uranium against
70% thorium and was being called NATURAL MIX.
- Chain anchors and verdicts now have hysteresis, so evidence sitting near a
threshold no longer empties the candidate list or retracts a correct answer.
Honesty of the display
- An elevated sample is never called background. Identifying which nuclide
takes counting time; noticing that something is there does not.
- Verdicts are graded: URANIUM? below 80% confidence, with the number shown.
- Quality grades the identification, not the strongest peak, so an
unidentified bump no longer reports EXCELLENT.
- Natural chains and K-40 must be above ambient to earn a verdict; every wall
contains both.
Regression suite (ANALYSIS_SELFTEST=1)
Synthetic spectra built from the library at the detector's real resolution,
with a Compton continuum and reproducible noise: uranium, thorium, mixed,
Cs-137, Co-60, Am-241, K-40, background, indoor background with annihilation,
ambient thorium in building materials, elevated-but-unidentified, insufficient
data, and centroid accuracy. It fails on real defects - it caught a background
false positive never seen on hardware, and rejected two fixes that broke other
cases.