Benchmark ModernTCN and DLinear on real REDD data

[nipunbatra]

Summary

• register ModernTCN and DLinear as NILMbench models
• advance the single immutable nilmtk-contrib pin to merge commit 8d745493ed9f84dd00fb502ffe85943eaeedc4c8
• build one CUDA campaign image for both models, pinned to runner adaf03e42f5b8dd6b9ab95942ee191999f2d3b25
• publish six real REDD/A100 result bundles (two models × seeds 10, 20, 42)
• regenerate the living leaderboard automatically from all 57 immutable runs
• document the batched image policy: one campaign image for an eligible model batch, not one image per algorithm

Real T0 results

Protocol: corrected-t1-redd / fridge / 60s / seq 299 / 3 epochs / max 3960

Rank	Model	MAE (W)	F1	Parameters	Time	Peak CUDA memory
7	ModernTCN	56.62 ± 1.77	0.659 ± 0.007	41,921	11.55 ± 1.33 s	91.3 MiB
14	DLinear	75.92 ± 2.73	0.570 ± 0.021	600	10.53 ± 0.57 s	64.6 MiB

The generated leaderboard now has one evaluation cohort and contiguous ranks 1–19.

Provenance

• image: nilmbench:t0-adaf03e-cuda
• image content digest: sha256:d7253754a6a9133235076fa1c1555104aa8be8128443da96bc16ae3d46809aa8
• runner: adaf03e42f5b8dd6b9ab95942ee191999f2d3b25
• nilmtk-contrib: 8d745493ed9f84dd00fb502ffe85943eaeedc4c8
• hardware: NVIDIA A100-SXM4-80GB
• REDD SHA-256: 757e694a420157aa035c8fd345915080c2396cb7a37bfeb5f925bcb84f9532b8

The fail-closed preflight and final read-only candidate verifier both passed.

Verification

• uv lock --check
• uv run ruff check src tests
• uv run pytest — 129 passed, 1 skipped
• existing 17-model leaderboard regenerated byte-for-byte before the campaign
• final 19-model leaderboard regenerated byte-for-byte after publication
• JSON commits to the exact generated CSV SHA-256
• six unique result IDs; exact seeds 10/20/42 per model; 3,960 train and 2,448 test samples per run
• all 19 aggregates are smoke-verified