Skip to content

Fairshift Lab v0.3.0

Choose a tag to compare

View all tags
@lindgreendavid lindgreendavid released this 12 Aug 11:54
· 14 commits to main since this release
v0.3.0
8f3952f
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Fairshift Lab v0.3.0 makes probability calibration and decision-threshold dependence visible under controlled distribution shift.

Highlights

  • Independent training, source-calibration, source-evaluation, and target-evaluation populations
  • Source-only temperature scaling selected by held-out negative log likelihood
  • Raw and calibrated Brier score, expected calibration error, and reliability bins
  • Formal 19-point source and target threshold sensitivity analysis
  • Interactive reliability diagram and linked fairness/performance decision curves
  • 39 deterministic Python tests with 100% statement and branch coverage
  • Patched React, Vite, Vinext, and Cloudflare build chain with a production dependency audit

Calibration can fail after distribution shift, ECE depends on binning, and threshold sweeps do not choose a real-world decision policy. This synthetic laboratory is research and educational infrastructure—not a fairness certificate or compliance assessment.

Assets 2
Loading