A practical way to structure the hunt is to separate device-level instability from board-level parasitic feedback. K-factor assumes the measured two-port is complete and that both ports see 50 ohms; in a real PA, the gate may see a driver impedance, the drain may see a bias-network resonance, and the common source/ground may couple both stages. Treat any spurious tone as a loop-closure problem: identify the gain path, the feedback path, and the impedance condition that makes the loop phase reach 0 or 360 degrees with gain above unity.
Check sequence: 1) Map every spurious tone versus drain voltage, quiescent current, drive level, input termination, and output load. Note whether it starts below compression, only under drive, or only after a load-pull condition. 2) Repeat with the RF input terminated in matched, short, open, and a few reactive loads if available; conditional oscillation usually points to an out-of-band impedance issue rather than in-band matching. 3) Use a near-field probe or small capacitive probe to localize the tone, but remember it maps field strength, not the root cause. If the maximum is over a bias choke or decoupling loop, measure the bias-node impedance rather than immediately changing the RF match. 4) Probe gate and drain bias rails with a high-impedance FET probe or bias tee/spectrum setup, keeping the probe ground very short. A tone on both rails suggests common feedback; a tone mainly on one rail suggests a local decoupling or bias resonance. 5) Inspect return current paths, source/emitter vias, gate-loop area, drain-loop area, and shared bias branches. Even a small shared ground inductance can convert output current into gate-source feedback.
For fixes, prefer adding loss at the oscillation frequency while preserving in-band performance. Options include a gate or input shunt resistor, a series RC or lossy bead on bias, improved local decoupling with different self-resonant frequencies, damping of high-Q bias chokes, and tighter source grounding. Avoid simply stacking more capacitors: parallel capacitors and vias can create antiresonances that move the oscillation rather than remove it. After any change, re-run the same start-up, load, temperature, and drive matrix; a tone that shifts frequency when you alter a bias network is strong evidence that network controls the unstable mode. AI-generated technical discussion; verify against authoritative documentation.