Question: do the forward-matrix output scales discriminate well enough to ever support an "orthogonal-tomorrow" declaration?
Substrate: BTCUSDT@250 ร 10 crypto regimes (16 features) + EURUSD@5 ร 10 forex slices (141 features). 20 cells, real + shuffled-null per cell. Gates: G1 spread ยท G2 anchor-separation ยท G3 threshold-ability ยท G4 cross-regime signal ยท G5 null-collapse (blocking).
| Metric | Crypto | Forex | Reading |
|---|---|---|---|
| adv_auc (residual drift) | USEFUL [10111] | USEFUL [10111] | Best performer โ spread 0.96/0.88, regime-range spread 0.77/0.60, KS vs null 0.55/0.32 |
| wold_R (fresh vs echo) | 4/5 (fails anchor only) | USEFUL [10111] | Massive spread (0.999), strongest null separation (KS 0.90/0.55) |
| granger_echo_p (setโf echo) | USEFUL [10111] | DEGENERATE | Split verdict โ works on crypto; forex set-factor degenerates, needs rework |
| redundancy (max|ฯ| vs full set) | input-only | input-only | Anchor fix VERIFIED (duplicates read 1.000 both sides). Fails G5 by construction (rank-correlation is permutation-invariant) โ stays the INPUT series, not a forward gate |
| stab_freq | MISCALIBRATED | MISCALIBRATED | Weak null separation (KS 0.09/0.01) |
| bootci_w | MISCALIBRATED | MISCALIBRATED | Rides on redundancy; tiny scale |
turnover_imbalanceโofi, forex duration_usโintra_duration_us) now read exactly 1.0 (v1 read 0.589, backwards).adv_auc + wold_R โ real spread, features measurably differ in cross-regime stability, distributions collapse under shuffling.