Audits that verify the numbers are real: database reconciliation, leakage checks, saturation detection, and honest documentation of what failed.
| Version | Date | Finding | Entry point |
|---|---|---|---|
| v29 | 2026-06-08 | falsified prediction documented with post-mortem — "One falsified claim documented" (prediction accuracy 85.7%, 6/7) |
cycle synthesis |
| v31.4 | 2026-06-11 | measurement integrity report — metric saturation audit; corrected walk-forward σ=0.1184 across 9 out-of-sample folds |
report |
| v32.4 | 2026-06-12 | audit — overall PASS; verbatim caveat: "PASS WITH CAVEAT — The 0.9257 AUC is real and computed from actual training… there is feature leakage: entity embeddings were generated using domain information"; Historical Convergence = 0 bug documented | report |
Sibling themes: Oracle Grounding & Calibration · GNN & Machine Learning · Knowledge Graph & Convergence
Bridge pathways: Brier provenance chain unbroken since v27 (0.0413) → v27 cycle status · honest AUC reporting — out-of-fold 0.7716, "the only defensible metric" → GNN V3 performance report
Version lineage (measurement integrity): v29 falsified claim documented → synthesis · v31.4 saturation audit → report · v32.4 current full audit (overall PASS) → report