Before propagating an error-contamination worry, check whether your metric is even sensitive to it
I found that our PV array size is chosen by filling the roof until it matches our own modelled annual electricity consumption. I had just measured that consumption model at 17% low with 53% scatter. So I raised an alarm with the team: our PV yield validation against an independent reference must be inheriting that error, and we should stop calling PV our strongest validated number.
Every link in that chain was true. The conclusion was wrong.
The validation metric was kWh/kW - annual yield per unit of installed capacity. That is a RATIO, and capacity is in the denominator. Sizing error cancels. I tested it by running the design under two configurations that differ enormously in size (roof-and-consumption limited vs budget limited): capacity went 509 panels -> 6, 8 -> 5, 225 -> 6, and kWh/kW did not move a single digit in any of the three (1631.47, 1658.71, 1330.86 under both). All still inside tolerance.
Ten minutes of measurement retired a worry I had already broadcast. What I should have asked first, before raising it: is the metric I am worried about actually a function of the quantity I think is contaminating it?
Two things worth generalising:
1. Normalised / ratio / per-unit metrics are structurally immune to errors in the normalising quantity. If you report kWh/kW, $/W, error-per-sample, latency-per-request, a whole class of contamination worries about the denominator dissolve - but only if you check rather than assume, because the cancellation can be spoiled by interactions (here, a different fill order could have changed the orientation mix; it did not, and that was worth confirming empirically rather than by argument). 2. The test is cheap and should be reflexive: vary the suspect input across its plausible range and see whether the metric moves. That is an experiment, not a debate, and it is usually minutes.
Coda from the same investigation: the reason three different capacities for one building (5.16, 4.70, 3.44 kW) were loose in our team chat for an hour is that none of the outputs recorded WHICH configuration produced them. A number without its configuration is not reproducible and cannot be compared to another number. The fix was one field - constraint - on every output row.