WattPair U-Shaped (Saddle) Window Air Conditioner Dataset

The first path. Certified capacity over certified efficiency ratio, both read off the same certification record.

The second path, printed per row so the arithmetic is auditable. 750 hours is the cooling-season figure the federal procedure assumes.

10 CFR 430 subpart B appendix F §5.2.1–5.2.2 (Uniform Test Method for Measuring the Energy Consumption of Room Air Conditioners)

The two paths are the same identity, not two measurements. Because appendix F DEFINES the efficiency ratio in terms of the annual energy figure, capacity ÷ CEER and annual kWh × 1000 ÷ 750 are algebraically equal. Their agreement here confirms that the three published columns are mutually consistent to within their own rounding — it catches a transcription or rounding error, and it does not independently confirm any machine’s power draw. Anyone presenting it as corroboration, including us, would be overstating it.

The derived figure is an upper bound on cooling-mode input power, not that power itself. Appendix F folds inactive and off-mode energy into the same annual total, so capacity ÷ CEER carries a machine’s standby draw amortised across 750 cooling hours in addition to the compressor and fan. The excess is small, and it is in the direction that does not flatter the number.

Certified efficiency ratios are published to one decimal place, so a derived wattage inherits about ±0.4% of implied precision at the low-CEER end of this file. Figures are rounded to whole watts and no more precision than that is claimed. Cooling capacity is the certified rating, not a measurement of the machine you buy.

Stable id for this row, built from the ENERGY STAR product id. Use it rather than the model number: nine model numbers appear on more than one row.

Brand name as filed with ENERGY STAR, verbatim.

Manufacturer model number as filed. NOT unique across this file — six TCL models were filed twice a day apart, and three model numbers are shared by different brands.

Sibling or private-label model numbers listed on the same certification. Null on all but a couple of rows.

The upstream primary key. Unique across the file.

The programme’s own long-form certification identifier.

The official mounting column, which is what makes this dataset reproducible: `straddles_window` is the U-shaped body that surrounds the raised sash, `straddles_windowsill` sits astride the sill. Nothing here was classified by eye.

Certified cooling capacity under the federal test method.

Combined Energy Efficiency Ratio, the certified rating. Published to one decimal place, which bounds the precision of everything derived from it.

running_watts = cooling_capacity_btu_h ÷ ceer_btu_per_wh. Input power under the federal cooling test, plus standby amortised over 750 hours — an upper bound, rounded to whole watts. Not a manufacturer figure.

Certified annual energy, on the federal 750-cooling-hour assumption. It is a comparison metric between machines, not a forecast of your electricity bill and not a basis for battery runtime.

running_watts_from_annual_energy = annual_energy_kwh × 1000 ÷ 750. Printed so the arithmetic is auditable per row — but see the caveats: this is the same identity rearranged, not a second measurement.

Absolute difference between the two derived wattages, as a percentage of the capacity ÷ CEER value, computed before rounding. Non-zero values are the published columns’ own rounding, and nothing else.

Supply voltage as filed. A 230 V row cannot be run from an ordinary US outlet or from any portable power station on this site.

True where the unit is a reverse-cycle machine — a saddle-mounted heat pump.

How far the model’s energy use sits below the US federal minimum for its class.

Whether the model meets the programme’s Most Efficient criteria for its year.

The DOE product class the model is certified in — reverse cycle or not, louvered sides, and the capacity band. It is what the standard it is measured against depends on.

When the certification was recorded.

Manufacturer-declared market availability. It does not mean the model is still on sale — this is a certification list, not an inventory.

The markets the certification covers. Two rows are Canada-only and are kept, with this column saying so, rather than dropped to round the count.

Cabinet width as filed. It is the machine, not the opening it needs — the sash clearance a saddle unit requires is in the installation manual, not in this filing.

Cabinet height as filed, measured over the whole body rather than over the part that sits inside the room.

Cabinet depth as filed, front to back across both halves of the saddle.

Unit weight as filed — the number that decides whether one person can hang it.

Whether a support bracket is supplied with the unit.

Refrigerant charge type as filed.

The same fact with the programme’s global-warming-potential annotation.

Whether the compressor is inverter-driven. Every certified room air conditioner in the filing reads Yes, saddle or not — an observation, with no claim here about why.

Product type as filed.

Whether the model is filed as network-connected capable.

None, anywhere in the file. This is a reference table, and a reference table with shopping hooks in it is an advert wearing a lab coat.

Not published. That library is what this site sells against, and opening it would be giving away the commercial core rather than the research.

A retail barcode. It identifies a package on a shelf rather than a certified machine, it is the field that turns a reference table into a shopping index, and it is absent from about a third of the rows upstream anyway. The ENERGY STAR product id exported here identifies the certification record itself, and the upstream file linked in `source` still carries the UPC for anyone who needs one.

Reads "None" on all 204 rows, so it separates nothing. It is withheld rather than shipped as a column of identical values because a reader would reasonably take it to mean "no saddle unit fits a casement window", and the field does not establish that — it was checked and found to carry no information at all (2026-08-03).