Model lineage of the graph neural network, from v18 training through GPU-enhanced message passing — with honest AUC reporting.
| Version | Date | Finding | Entry point |
|---|---|---|---|
| v18 | — | model training | gnnmodelv18 |
| v19 | — | features | gnnmodelv19 |
| v20 | — | Granger causality | gnnmodelv20 |
| v24 | 2026-06-05 (initial recon: 2026-05-19) | Edge-Aware GNN AUC 0.971 (landing history table) |
coordinator summary · reconnaissance summary |
| v25 | 2026-06-05 | TGAT selected — AUC 0.9405 (target 0.910 exceeded), 4x faster than TGN |
cycle synthesis |
| v29 | — | GNN V3 honest AUC 0.7716 — data-limited, not architecture ("Honest Assessment") |
performance report |
| v31.5 | 2026-06-11 | enhancement reports ×3 (base / GPU / GraphSAGE) | report · GPU · GraphSAGE |
| v32.5 | 2026-06-12 | GraphSAGE-style message passing + residuals — AUC 0.9987 on CUDA (RTX 4060 Ti) |
GNN report |
Sibling themes: Knowledge Graph & Convergence · Measurement Integrity · Infrastructure & Ops
Bridge pathways: Honest AUC reporting — out-of-fold 0.7716, "the only defensible metric" → GNN V3 performance report · GPU training on RTX 4060 Ti (CUDA) → v32.5 GNN report
Version lineage (GNN): v25 TGAT selection → synthesis · v29 GNN V3 → report · v31.5 enhancement reports → GraphSAGE report · v32.5 current → report