Output falls about 0.9% for each °F of panel heat on 26 clear days (likely −1.1% to −0.7%), about 5 times the 0.2% silicon loses, so most of it isn't heat and the evidence for heat alone is still weak.
2 readings far outside the rest aren't drawn.
Silicon panels lose output as they heat up, typically about 0.4% per °C of cell temperature (0.2% per °F). This page checks whether that shows in this system's data. The dashed grey line is that typical loss drawn through the middle of the readings; the gold one is what the readings show.
Readings: the 15-minute readings between 10:00 and 15:00 on clear days since Jul 3 (when the array reached its current size), with at least 600 W/m² of sunlight at the station. A clear day is one whose 10:00–15:00 sunlight averaged at least 700 W/m². Output per unit of sunlight is solar_w ÷ solar_radiation_w_m2(effective m²).
Panel temperature isn't measured, so it's estimated with the NOCT model: T_panel = T_air + (NOCT − 20 °C) ÷ 800 × sunlight, NOCT 45 °C, about 1.8 °F hotter than the air for every 100 W/m². The slope is a least-squares line, stated as a share of the average output. Its likely range is a 95% interval from 1,000 resamples of whole days (readings on the same day share clouds and sun path, so they aren't independent), with a fixed seed so the page doesn't change between visits.