Discontinuity in MRF1513 fet.

I'm using an mrf1513 as a driver for a larger device. I noticed as I varied the gate bias the output on the drain would vary quite nicely and then suddenly step to a larger value. The step only occurs at high drain rf voltages. The rf on the drain swings to a peak of around 30v and minimum of a few volts negative. If I reverse the change to the gate bias I return to the original operating condition via another step response in rf amplitude. There is a slight hysteresis on the gate bias so I need to wind back a little more than I expect to return to the original state. I'm operating at VHF. Typical of an oscillation I thought. I have spent almost a day searching for untoward products, nothing to 8ghz!
In an attempt to isolate the cause I rebuilt the driver stage onto copper clad board and used the effective Zin of the next stage to form a load, this was made from a real resistor and a capacitor. The feed to this test amplifier has been a network analyser or a std sig gen. This isolated amplifier does the same thing regardless of source.
The input level from 50r is 17dbm, there is an inductor across the feed to tune out some of the gate capacitance and there is a 10 ohm feed into the mrf to enhance stability. I feed the gate bias from a psu, through a potential divider and through a 1k to the gate. As far as the drain is concerned there is an inductor of 21nh to +12, a 27pF series to a 11nh to gnd. The load is a 100pf in series with a 3ohm resistor connected to the 27pf, 11nh node.
It is resilient in its operation. I have changed resistors, transistors, psu's, added caps here and there. It can always be made to generate a step function with a smooth change in gate bias.
What mechanism could be causing this?

it is likely breaking into oscillatons.

I agree that oscillations are the most obvious and likely cause. However as already explained there are no oscillations to be found. I have loosely coupled the drain to a spectrum analyser and looked upto 8 ghz. Nothing.
Usually high frequency oscillations are very sensitive to small changes in component values. Not this. I can add capacitance to gate or drain, compensate for any loss in fundamental performance by adjusting bias or drive level and the step change still occurs.
I've been involved with rf design for almost 30yrs and oscillations either tend to affect the sideband signature of the fundamental or produce non harmonically related products.
I accept that it is usual that oscillations would cause this effect but there are no tell tale signs.
Am I missing something obvious here?