SniperGold_ML/repaired/pre_purge/pre_purge_run.log

=== PRE-PURGE BASELINE (SB-01/SB-02 PRE) ===
Tanggal      : 2026-08-21
Sumber       : run sinkron SEBELUM edit (kode lama), command:
               python train_model.py XAUUSDc --out repaired\pre_purge\SniperGold_ML_prepurge.mqh --tag v20260821_prepurge
Hash kode lama (train_model.py, sebelum edit) : 96BC258FA62630A67060DBD03655834798DA77799A72228BF8529BE6169B360D
Dataset (features_XAUUSDc.npz)                : E60285F55AB5AF7B397653373F0EE5B8B814708DC178DD58691DAE0BE122946A
Artefak model PRE sejati = FREEZE Include\SniperGold_ML.mqh (hash C0D5CAE6..., tidak tersentuh).

Hasil (kode lama; metrik validasi RUSAK = SB-01; split lama tanpa purge = SB-02):
  total bar: 60000 | sampel berlabel: 47537 (bull 26125, bear 21412)
  AUC validasi (long head): 0.5005 (epoch terbaik 0)   <-- metrik rusak (ravel 2-head vs 1 head)
  LONG : TRAIN AUC 0.6523 | VAL 0.6443 | TEST 0.6270
  SHORT: TEST AUC 0.6207
  CALIBRASI LONG A=1.2115 B=0.2612 | SHORT A=1.6690 B=-0.1410
  n_tr=30304 n_te=11885 tag=v20260821_prepurge
  TEST LONG 0.6270 / SHORT 0.6207 == freeze header (reproduksi persis).

CATATAN INTEGRITAS:
- File SniperGold_ML_prepurge.mqh yang sempat dihasilkan TIDAK valid sbg artefak PRE:
  run background tidak sengaja memakai kode yang SUDAH diedit (0.6582/0.6582).
  File tsb di-rename -> SniperGold_ML_prepurge_INVALID_overwritten_by_postedit_run.mqh.
- Evidence PRE yang sah: freeze mqh (immutable) + output di atas (transkrip sesi).