Model card

AlphaTrack AT-L1

Detects track-like features in labelled synthetic / public CR-39 imagery. It makes no particle species, energy or yield claim without documented calibration.

Synthetic benchmark only. The scores below do not establish performance on real CR-39 imagery. Real detector validation, protocol-specific calibration and partner-lab review remain pending.

Intended use

Public software demonstration and a permanent explainable regression oracle for ML models.

Prohibited use

Species / energy / yield inference from uncalibrated images; any operational radiation-safety use.

Architecture

Deterministic OpenCV baseline (AT-L0) + a logistic-regression classifier over candidate features (AT-L1 stand-in for a detector/segmenter).

Training / evaluation dataset

alphatrack-synthetic-v1.0 (synthetic). Split by physical plate — never by tile.

Performance (synthetic, plate-separated holdout)

Precision1.0 95% CI 1.0–1.0
Recall0.975 95% CI 0.926–1.0
F10.987 95% CI 0.962–1.0
Counting MAE1.667
Calibration error (ECE)0.025

Calibration

Confidence calibration reported (ECE); abstention on failed QC / out-of-domain input.

Failure modes

  • Crowded fields reduce recall (merged tracks)
  • Extreme blur / low contrast
  • Novel artifacts not seen in training

Supported / unsupported domain

Supported: the declared synthetic generator domain. Unsupported: real CR-39 plates, other detectors, or acquisition outside the model card.

Status

  • External validation: Pending — no real CR-39 imagery; partner-lab review required
  • Model version: AT-L1-v0.1
  • Weights sha256: 5567f043f1d4ed3980b99215…

Try it: AlphaTrack synthetic demo → · Full card in source: docs/model-cards/alphatrack-l0.md.