I dug out my old homebrew 1GHz diode detector and tested it this evening. To minimise uncertainty, it's best to use it in a levelling system as in the diagram below. This system uses ALC (feedback) to level the signal generator to the accuracy of the power meter in the ALC feedback path. The source VSWR is defined by the quality of the 50R resistors in the splitter.
This is the classic method used by calibration houses to test the frequency response of various devices as it minimises uncertainty.
In the test I've just done. I put the diode detector in the ALC path (in place of the power meter) so the system now levels the sig gen to the frequency response of the diode detector. A sig gen with low harmonic content is needed here or uncertainty can become significant. However, the OP only needs to prove +/- 1dB accuracy.
I've used a thermocouple power meter as the DUT in the setup below so the frequency response plot below is my Anristsu thermocouple power meter measuring the levelled output. It therefore measures the frequency response of the diode detector.
I know that this detector is only good to about 1GHz. Above this frequency, the internal parasitic package inductance and capacitance cause the diode detector to over read the power level. The input return loss of the diode sensor is better than 30dB below 1GHz. Below 500MHz it's much better than 30dB return loss. You can see that the levelling performance degrades markedly above 1GHz. By 1.5GHz there is a 0.5dB error for example.
Whatever ripple is present on the plot below is partly due to the frequency response of the diode detector and from the influence of harmonics from the test generator (on the diode detector) and from the fact that the thermocouple sensor efficiency isn't perfectly flat vs frequency. The software knows the cal factor vs frequency for the sensor (as stamped in a table on the thermocouple sensor body) but in my experience, a typical thermocouple sensor will still have subtle efficiency variations between the cal points stamped on the sensor. So there will be various things contributing to the fact that the plot below isn't perfectly flat.
The performance is good when you consider the diode in the detector costs less than about £0.50.
Hardly anyone needs performance better than this. In a real system, the wiggles in my plot will be lost in the inevitable mismatch uncertainty between the stages of real world circuit designs. So there's not much point trying for better performance than this unless you are in the business of designing very exotic and expensive RF power detectors.
I'll have a look at extending the frequency range of the diode detector but I doubt I'll be able to get it to work to 3GHz. My recommendation would be to buy an old 26GHz HP 8473C diode detector for higher frequencies like this. The diode detector is good for measuring flatness but it will need to be referenced in some way if you want to make absolute power measurements.
I usually do this at a low frequency and then rely on the frequency response of the diode detector to verify the power at other frequencies. This diode detector method would be a good way to cross check the performance of any DIY thermal power sensor.