Publish verified REDD T0 leaderboard

[nipunbatra]

Outcome

Publishes the first automatically generated, real-data NILMbench T0 leaderboard.

• corrected REDD T1 refrigerator task
• 17 methods, including the NILMTK Mean baseline
• 3 seeds per method: 51 accepted runs
• 3 epochs, 3,960 training rows, 2,448 test rows
• NVIDIA A100-SXM4-80GB execution
• every aggregate is smoke-verified

The leaderboard is generated from committed result bundles; no metric was entered by hand. JSON and CSV are cross-checked, and the JSON commits to the exact CSV checksum.

Reproducibility

• runner: 83fb39e
• nilmtk-contrib: 825740b39bcd44b3f4bfaf146f4c0d944843b131
• image: nilmbench:t0-83fb39e-cuda
• image digest: sha256:dd977962c2e0d72e2d923f8f5c3e92e538f67a00495e545c2f22571001872e91
• REDD SHA-256: 757e694a420157aa035c8fd345915080c2396cb7a37bfeb5f925bcb84f9532b8

Verification

• 129 NILMbench tests passed
• 4 container-contract tests passed
• scoped Ruff, lock check, diff check, wheel and sdist passed
• immutable-image pip check passed
• full REDD/A100 preflight passed
• all 51 candidate bundles independently reverified
• finalizer confirmed 51 source runs, 17 aggregates, three exact seeds per method, and 105 expected publication paths

This is deliberately a T0 smoke leaderboard, not a converged accuracy claim. It establishes a reproducible, continuously extensible real-data baseline.