CreditVintage v0.2.1

Adds an offline benchmark on 30,000 attributed UCI credit-card records, with five frozen train/calibration/test splits, four fixed methods and three equal review budgets. Every score, split, model result, label-free queue and source notice can be downloaded and refitted.

Numerical dependency versions are pinned for replay. CI also transfers the same Linux Python 3.12 evidence and wheel to Python 3.13 and Windows for independent refitting. A mismatched numerical runtime reports the recorded versions before fitting; score and metric tolerances remain unchanged.

The v2 card protocol uses Newton-Cholesky logistic fitting with a fixed 1e-10 convergence tolerance and a six-decimal input grid for isotonic calibration to prevent amplification of platform-dependent logistic tail differences; raw logistic scores are retained. Older v1 evidence keeps its original calibration protocol.

These are random partitions of one historical cohort, not future-period or point-in-time validation. The native subsequent-default target remains distinct from the synthetic 180-day DPD target.

Assets are the exact wheel and source archive that passed clean installed-package replay in the linked successful main CI run. release-manifest.json records the commit, version and digests; CHECKSUMS.sha256 covers all other assets. Dataset attribution and CC BY 4.0 terms are bundled with the data.