Suite18 — FMM representation experiment ======================================= FILES ----- suite18_fmm.py Standalone chronology-safe multiscale fuzzy-min-max encoder. train_suite18_fmm_forward_specialist_boosting.py Thin Suite17 derivative. It reuses Suite17's learner/evaluator and injects the FMM representation only at the R/M/L view boundary. run_suite18.sh Server command matching the existing Suite17 run layout. tests/ Unit and integration tests for FMM chronology and representation behavior. SUITE18_DESIGN.md Frozen experiment specification. PLACEMENT --------- Recommended: / train_suite17_robust_forward_specialist_boosting.py suite17_alignment.py suite17_data.py suite17_hgb.py suite17_local.py suite17_pipeline.py suite17_reporting.py suite17_specialists.py suite17_structured_cache.py suite17_temporal.py suite17_training.py ...Suite17 dependencies... suite18/ The Suite18 entry point automatically checks its parent directory for Suite17. If Suite17 is elsewhere: export SUITE17_CODE_DIR=/absolute/path/to/the/suite17/source RUN --- cd /suite18 ./run_suite18.sh For a clean FMM rebuild: ./run_suite18.sh # after adding --force_fmm to the Python command The normal rerun reuses per-fold/per-view FMM caches when the view shape, feature schema, temporal range and FMM configuration are identical. RESEARCH SAFETY --------------- 1. Committee is never retrained with FMM. 2. R_RELATION, R_RELATION_CURRICULUM, R_RELATION_LATENT and R_FULL use separate FMM encoders. 3. Historical rows are always queried before their labels update FMM. 4. Preprocessing is fitted on a strict historical prefix then frozen. 5. Raw Suite17 specialist features are retained unchanged. 6. CART/RF is diagnostic only and cannot enter forward selection. 7. Target labels are used only for post-hoc evaluation, never FMM/tree fitting. TEST ---- python3 -m pytest -q tests NORMAL SERVER WORKFLOW ---------------------- Suite18 resolves Suite17 relative to the Suite18 source file, so no environment variable is needed for the normal sibling-directory layout: cd /root/disk_3tb/earlywarning/30072026/suite18 ./run_suite18.sh The complete stdout/stderr analysis log is written to: /root/disk_3tb/earlywarning/30072026/log/suite18.log The Python entry point automatically searches the project root and every sibling directory whose name starts with "suite17". SUITE17_CODE_DIR is only an override for nonstandard layouts. LOG ANALYSIS CONTENT -------------------- The log includes the inherited Suite17 diagnostics plus Suite18-specific: - per-view FMM build/availability and theta hyperbox summaries; - FMM prequential chronology/leakage audit with violation count; - top selected raw features per view/fold; - hyperbox support, confidence factor, width and contraction diagnostics; - raw/FMM/augmented representation dimensions; - standalone R-ablation/M/L historical and target diagnostics; - forward candidate/stage decisions and correction diagnostics; - Suite18 vs Suite17 holdout AP/AUC/F2 comparisons when Suite17 results exist; - Suite18 vs every available stored lineage (Committee, Suite16/14/13/12, Suite11/10/9/8/7/shared, Council and V2) with mean/median/worst faculty deltas and faculty win rates; - paper-inspired FMM/CART/RF diagnostic performance, explicitly excluded from model selection.