01 / Cold snapBlocked
Comfort-floor stability and forecast value did not pass.
PENGAby GlacialRift
PENGA Research Lab
Where the controllers are trained.
PENGA trains controllers in a simulator built from Alberta buildings and Alberta weather. What clears the benchmarks moves toward a real boiler room — and the failures stay on the record.
The promotion rule
Models earn their way into the boiler room.
A model that saves gas but loses heat does not pass. Failed runs stay visible so the next experiment starts from the truth.
ComfortCold-tail losses stop the run.
Plant behaviorCycling and recovery stay visible.
Forecast valuePrediction has to beat reaction.
Internal simulation
What internal simulation currently shows.
These are internal simulation results, not field results. No candidate controller has been promoted or proven in a real building. Simulation tells us where to look next; only field evidence can make a claim sellable.
Internal simulation signal
6–35% less gas
Across three Alberta-weather replays versus neutral outdoor reset. Internal simulation only.
Average internal simulation reduction17.0%
Internal simulation signals3 / 3
Peak internal simulation reduction35.5%
Internal simulation: reference controller versus neutral outdoor reset. Not a field result, a promoted controller, or a guaranteed saving.
Thermal memory
Buildings hold heat for hours after the boiler stops.
Weather, water, concrete, and plant operation move on different clocks. PENGA learns where those delays create useful room to act.
MinutesWeather changes
HoursWater and plant respond
LongerThe building gives heat back
Public evidence surface
Failure stays visible.
Every benchmark run is logged with its baseline, comfort floor, and the reason a candidate did not advance.
Latest benchmark
0 / 3 promoted
Blocked at the promotion gate — the bar is high on purpose
02 / Warm frontBlocked
Comfort held close. Predictive value was not demonstrated.
03 / Shoulder dayBlocked
Safe coasting appeared. Solar lookahead was unavailable.