Scientific ML T&E · governed synthetic benchmark

What survives when the gamma spectrum shifts?

Choose a model and a labeled-spectra budget, then compare in-envelope validation, a frozen compound shift, background false alarms, and unknown-isotope rejection. The point is not the prettiest score. It is the trade-off exposed when one threshold has to survive all four lanes.

Independent prototype · synthetic classifier benchmark · no hardware or field deployment

Training and evaluation spectra are synthesized with PyRIID from public GADRAS-derived seed signatures in Sandia's BALDR study. N42 files are used only for parsing and replay compatibility.

01Public seed signatures
02Synthesis + disjoint splits
03Five limited-data models
04Frozen shift + rejection tests
05N42 / API replay prototype

FIELDGAMMA / FORMAL-V1-20260825

Inspect the frozen result

Model settings and abstention thresholds were fixed before protected-test evaluation. These controls read saved results; they do not retrain anything.

HASH-BOUND

Model

Labeled spectra per isotope

In-envelope validation macro-F1—

Frozen shifted-test macro-F1—

Background false-alarm rate—

OOD unknown-rejection rate—

Validation—

Compound shift—

Background false alarm—

Unknown-isotope rejection—

T&E reading

Random Forest · 25-shot

Loading the hash-bound result…

RUN LEDGER

What this run left behind

60pre-registered fits

3random seeds

96,000prediction rows recomputed in R

1.11e−16maximum Python / R difference

10 / 10Python tests passing

76.1 sformal-run time

ANSI N42.42 / PUBLIC FIXTURE

N42 interoperability lane

My parser and SandiaSpecUtils agree on the same public Mn-56 fixture: channel counts, total counts, and live time match exactly, while calibrated energy edges agree within 0.0003 keV. That validates file interoperability on one fixture—not measured classifier performance.

16,384 channels
698,514 total counts
1,800 s live time
≤ 0.0003 keV edge difference