You've probably figured this out by now, as you're a smart fella, but at one point, Keysight has a Noise figure uncertainty calculator on their website that would estimate how much error there was in a measurement, given the VSWR's and the ENR performance. That might give you some direction.

You've probably figured this out by now, as you're a smart fella, but at one point, Keysight has a Noise figure uncertainty calculator on their website that would estimate how much error there was in a measurement, given the VSWR's and the ENR performance. That might give you some direction.

Both PNA and Y-factor measurement options are sensitive to oscillations, and external sources making its way into the signal chain. Y-factor measurement depends on VSWR, and for sub-dB measurements, the ENR of the source is affecting the measurements, and can vary with calibration year per year. AFAIK accuracy with the newest USB-C sources from Keysight is in the range of +/- 0.2 dB NF with a known DUT (I have had the luxury of using a PNA and Y-factor equipped NFA side-by-side). PNA measurements are can be difficult as compression will affect the measurement.

Did some experiments with putting an isolator in front of the LNA and trying to subtract out its loss. It definitely improved the bumpiness on the Y-factor measurement.

Oscillations are a possibility, when we started taking data the NF was like 10 dB but the amp was singing like crazy. Capacitors and resistors seems to have fixed that.

I will look at the VSWR uncertainty calculation...

I think part of the larger picture is that just maybe the noise figure shown in a MMIC data sheet was measured on-wafer, and the user has to suffer some loss mounting it to a circuit board. For yucks we might but a Qorvo CMP2626 LNA eval board

https://store.qorvo.com/products/detail/qpa2626eval-qorvo/597407/

But it does not appear to have a "through line" to calibrate out the board losses.