{
  "label": "thermal sensation in held-out countries: PMV vs learned trees vs physics-as-starting-belief with a learned person correction",
  "registered": "2026-10-10",
  "code": "experiment/buildings/comfort.py at commit 8688579f36d41823525fc03059ef05aa5f96540d, mode run",
  "data": {
    "source": "ASHRAE Global Thermal Comfort Database II v2.1.0, db_measurements_v2.1.0.csv.gz sha256 9722be588ec3e779de21679513711e2d980138359a1b44bb9775e9499bc68380, db_metadata.csv sha256 af6770a2b469f079cd172bd6632d1ae726a3a663515126816ff6d2d8f51bfc10",
    "development": "dev.csv.gz sha256 bef92d664c36ebad61c28328ceb3e29810b9a0973d8a2c32b78817caee92edd5 (80,815 rows, 14 countries)",
    "holdout": "sealed/holdout.csv.gz sha256 cdbce5f6bd66afbd8c77a56172e62ecd48d6907e669987c07bc8438b346229a6 (28,218 rows; australia, brazil, denmark, france, germany, greece, indonesia, italy, japan, mexico, nigeria, philippines, singapore, south korea, tunisia), unread",
    "scored_rows": "held-out rows with a thermal sensation vote and a database PMV"
  },
  "arms": {
    "physics": "database PMV",
    "learned_reg": "LightGBM regression (L1) on measured room, person and place inputs, no physics",
    "learned_cls": "LightGBM most-likely-vote classifier on the same inputs",
    "lived": "LightGBM most-likely-vote classifier started from PMV class scores (width 2), same inputs"
  },
  "primary": "mean absolute error of the predicted vote against the voted sensation, pooled over scored rows",
  "bar": "MET if, for each of physics, learned_reg and learned_cls, the 95 % interval of MAE(lived) minus MAE(baseline), resampling whole buildings 2,000 times with seed 20261010, lies entirely below 0. Otherwise NOT MET.",
  "reported_only": ["exact-vote accuracy", "per-arm MAE"],
  "environment": "[host name redacted] CPU, Python 3.12.3, lightgbm 4.7.0, numpy 2.5.3, pandas 3.0.6, 2 threads",
  "runs": 1,
  "seed": 20261010
}