FinancePy adjusted binomial: invalid support points produce negative probabilities and shift the mean

For four independent equal-loss credits with default probabilities [0.01, 0.01, 0.5, 0.99], the actual FinancePy adjusted-binomial implementation returns probability −0.1127312875. A two-credit case changes the required mean from 1.60 to 1.65. The public credit-tranche function also returns an expected surviving fraction above one.

This is the archival release of the existing 15 September 2026 finding, replayed on 17 September against pristine current master 4c7cd50bdadd6efc9ac74fa81e93374397d18e3e and official PyPI 1.1.2. Across 2,844 frozen equal-loss portfolios, an independent rational moment oracle finds 295 / 0 / 295 / 295 failing vectors for original / candidate / restored / released code. Twelve separate integration scenarios and the existing upstream test are retained. The portable runner was executed; the full source archive and wheel hashes were verified.

The candidate selects adjacent support points bracketing the mean. It fixes the tested probability and moment invariants while preserving the method’s approximation error for general portfolios. Unequal loss sizes, real bank deployment, customer losses, full upstream suite and performance were not measured. Existing issue 265 is open; maintainer acknowledgment or a merged fix is not claimed.

GitHub report and evidence · Original developer report · Existing 34-second English video

Independent GERO research by Xamit Kadirbekov. AI-assisted research and editorial preparation. The video uses original graphics and disclosed synthetic edge-tts narration; no media binary is included. Report/evidence descriptions: CC BY 4.0. FinancePy source, copied tests and derived code retain GPL 3.0; see LICENSES.md. Archive SHA-256: 45637af8cf3c1399639c98c8748e8e5ad2df1e8c4c10b139628d27d0ad42edab.