Benchmark ResidualMoE on real REDD data

[nipunbatra]

What changed

• register ResidualMoE as a residual mixture-of-experts family
• advance the single model-agnostic CPU/CUDA image family to merged nilmtk-contrib commit 0769d00893a178277d06e83ccd2e92c03a5ba9ef
• publish three immutable real-REDD result bundles and approve their exact runtime tuple
• regenerate the checked-in JSON/CSV leaderboard from 72 source bundles

T0 result

Protocol: corrected T1 REDD, building 1 fridge, 60 s sampling, sequence length 299, 3 epochs, 3960-sample limit, seeds 10/20/42, NVIDIA A100-SXM4-80GB.

Rank	Model	MAE	F1	Parameters	Time	Peak GPU memory	Status
5/24	ResidualMoE	50.66 ± 0.95	0.755 ± 0.019	440,870	18.25 ± 0.94 s	6424.7 MiB	smoke-verified

ResidualMoE improves over its TimesNet anchor on this slice (53.54 MAE, 0.728 F1) and over the earlier unconstrained NILMMoE (56.90 MAE, 0.707 F1). It is not the overall winner: NILMFormer remains rank 1 at 48.09 MAE. The result is retained without post-hoc tuning.

Provenance and reproducibility

• all three publication runs passed on the first attempt
• an independent second three-seed campaign reproduced every per-seed MAE and F1 value exactly
• runner: e33b94d44d07b098f774e8569298c9853dd71422
• nilmtk-contrib: 0769d00893a178277d06e83ccd2e92c03a5ba9ef
• image: nilmbench:t0-e33b94d-cuda
• image digest: sha256:3eb89c6d2bcdc8942426e359ed8b1b6ed74108060d86a293101a186844347c86
• REDD SHA-256: 757e694a420157aa035c8fd345915080c2396cb7a37bfeb5f925bcb84f9532b8
• final leaderboard: 72 bundles, 24 models, one ranking cohort, contiguous ranks 1–24, all entries provenance-verified

Verification

• 130 tests passed; 1 dataset-dependent skip
• Ruff passed for src and tests
• uv lock --check passed
• direct CUDA model/expert/router/zero-anchor contract passed on an A100
• independently regenerated leaderboard.json and leaderboard.csv are byte-identical to the committed artifacts
• git diff --check passed