# results_rank_gauss

**Claim:** `rank-gauss-hp30` — causal **past-only rank→Gaussianize** (empirical CDF → `norm.ppf`) on **4-seed** classifier logits, then mean and/or causal `hp30` (τ=30). Also **hp30→RG→mean**. Rolling windows W∈{15,30,60,120}s + full past. **NOT** multi-τ / seed-reliability / seed-disagree / online-moment / CORAL / fuse↔hp30.

**Honesty:** Label-free. Select mean LODO **Acc then H** @ HMM **0.994/0.8**. Never AASD/Carta labels.

**Compute:** CPU-only (~485 s; numba RG). No `gpu.lock`. Artifacts: `rank_gauss_hp30.py`, `results/research/rank_gauss_hp30.json`, `models/rank_gauss_hp30/selection.json`, `experiments/rank_gauss_hp30.log`.

## Selected rule

- **Selected:** `equal_mean` (raw equal mean of 4 seeds — all RG orders lose LODO Acc to this baseline)
- **LODO mean:** H 0.7932 Acc 72.68
- **AVGC:** H 0.7173 Acc 68.51
- **AASD:** H 0.6417 Acc **52.91** (ΔAcc vs linear_hp30 53.12: **−0.21**)
- **Carta2026:** H 0.6454 Acc 59.28

Best RG-by-LODO (not selected): `rg_w120_mean` LODO 0.7888/68.57; AASD 0.6412/52.92.

## LODO top (selection Acc then H)

| Variant | mean H / Acc |
|---------|--------------|
| **equal_mean** (selected) | **0.7932 / 72.68** |
| rg_w120_mean | 0.7888 / 68.57 |
| rg_full_mean | 0.7834 / 68.46 |
| linear_hp30 | 0.7797 / 68.15 |
| rg_w60_mean | 0.7737 / 67.08 |
| rg_full_hp30 | 0.7410 / 63.53 |
| rg_full_mean_hp30 | 0.7407 / 63.49 |
| hp30_rg_full_mean | 0.7445 / 63.37 |

Short rolling windows (15–30 s) collapse LODO Acc (~55–62). RG→hp30 and hp30→RG both hurt vs RG→mean alone.

## External (headline = selected)

| Variant | AASD H / Acc | ΔAcc vs 53.12 | Carta H / Acc |
|---------|--------------|---------------|---------------|
| **equal_mean** (selected) | **0.6417 / 52.91** | **−0.21** | **0.6454 / 59.28** |
| linear_hp30 | 0.6422 / 53.12 | 0.00 | 0.6520 / 60.54 |
| rg_full_mean_hp30 (illegal best Acc) | 0.6425 / 53.14 | +0.02 | 0.6428 / 59.76 |
| rg_w120_mean_hp30 | 0.6419 / 53.13 | +0.01 | 0.6432 / 59.63 |
| rg_full_hp30 | 0.6424 / 53.12 | 0.00 | 0.6428 / 59.76 |
| rg_w120_mean | 0.6412 / 52.92 | −0.20 | 0.6500 / 58.45 |

Illegal AASD Acc peek stays within **+0.02** of `linear_hp30`. Rank-gauss does **not** close AASD Acc→60.

## Verdict

- Honest LODO selection → `equal_mean` (LODO **0.7932/72.68**). Every RG / RG↔hp30 variant loses LODO Acc.
- AASD Acc **52.91** vs portfolio external best **53.12** (Δ **−0.21**). Carta regresses to 59.28.
- Open targets unchanged: AASD Acc>60 still −7.09 under selected (−6.88 under `linear_hp30`); AVGC still short of 0.85/78.
- Portfolio unchanged: fuse LODO / `linear_hp30` external.

## Do not repeat

Do not rerun this causal past-only rank→Gauss (full|W∈{15,30,60,120}) × {mean, RG→hp30, mean→hp30, hp30→RG} grid on 4-seed logits at HMM 0.994/0.8. Distinct from exhausted multi-τ / seed-reliability / seed-disagree / online-moment / CORAL — still do not repeat those either.

## NEXT IDEA

**Causal per-seed temperature / emission-scale** fit on LODO train only (one T_i or σ_i per seed from train emissions or past-only running moments; no AASD/Carta labels), then equal mean → HMM. Orthogonal to rank-CDF reshape and exhausted 4-seed weight/disagree/multi-τ. Avoid fuse↔hp30.
