콘텐츠로 이동

GWN Teacher Phase Status (2026-04-09 Final)

GWN Teacher N=50 Conditional PASS (2026-04-08). Phase 2 KD Ablation completed and REJECTED (2026-04-09).

Key Numbers

  • GWN N=50 5HH Avg MSE (kW): 0.5049
  • DLinear M0 baseline: 0.5207
  • 5% gate threshold: 0.4947
  • Household breakdown: GWN wins Apt6(-6%), Apt30(-7.9%), Apt51(-9.5%); DLinear wins Apt15(+5.8%), Apt88(+3.3%)

Phase 2 KD Ablation Final Result (2026-04-09)

2x2 Factorial (A1/A2/A7/A6) + Null Test (AN) executed.

Condition Mean MSE vs A1
A1: DLinear, No-KD 0.5223
A2: DLinear + GWN MSE KD 0.5186 -0.7%
A7: SC-DLinear + No-KD (A_adp only)
A6: SC-DLinear + GWN MSE KD + A_adp 0.5091 -2.5%
AN: A6 + Random matrix (null test) ~0.5048 similar to A6

Statistical tests — ALL NON-SIGNIFICANT: - KD main effect: +0.0059, p=0.2403 - A_adp main effect: +0.0073, p=0.3339 - KD × A_adp interaction: -0.0044, p=0.4446 - Null test (A6 vs AN): diff=0.0043, p=0.3711

Apt51 accounts for 98.6% of A6-AN difference — single-household dependence.

Root Cause of Failure

A_adp diagonal dominance ratio = 23.84 (threshold < 5.0, 4.8x over limit). Off-diagonal entries carry negligible spatial information. FiLM conditioning on a diagonal-dominated matrix is functionally equivalent to random initialization.

Storyline REJECTED

"GWN → A_adp → SC-DLinear → performance gain" is NOT supported by current data.

Next Step (P0 Immediately)

docs/reference/chronos_hidden_kd_design.md — priority elevated to P0. Chronos Hidden State KD (temporal feature path, independent of A_adp).

Why: Phase 2 complete failure drives direction change. A_adp spatial path abandoned. How to apply: Do not propose A_adp-based experiments. Focus on Chronos T5 encoder hidden state extraction as the next KD signal.