Cap or interpolate HR, power, and cadence spikes.

Cap if the underlying effort was real and only the magnitude is wrong - rare in practice. Interpolate for every other case: HR strap dropouts, power-meter zero-resets, dual-record conflicts, cadence-sensor noise. Interpolation is what cleans your normalised power and average HR, because it removes the spike entirely rather than just clamping it. Most people should default to interpolate.

Yes, and that is the point. After cleaning, average HR and average power drop to the values they should have been. The total work (kJ) computed from power drops accordingly. Normalised Power drops, because NP is sensitive to high-end variability and the spike was inflating it. The TSS computed from the cleaned NP is therefore also lower, which is what your training log should reflect.

Yes - we clean whichever streams the format carries. GPX preserves HR and cadence in the TrackPointExtension and (with the Garmin extension) power as well. TCX natively carries HR, cadence, and power. The peak-remove logic is format-agnostic; only the file encoding and which streams are present differ.

For HR, set the cap at your true max plus 5-10 bpm - typically 200-220 depending on age. For power, 2000W is comfortable for most riders; track sprinters might want 2500W. For cadence, 250 rpm catches all real spikes since no human pedals past about 230. The defaults work for 95% of people. Adjust only if you know your own true peaks exceed them.

Power meters compute force from torque sensors and angular velocity from cadence sensors. If the cadence reading dropouts or briefly mis-reads, the power formula divides by a near-zero or wrong value and the result is a giant spike. Same logic for HR straps - momentary electrode contact loss reads weird values. The streams are reliable on average but produce occasional one-second anomalies that destroy your downstream metrics.

Yes - in the right direction. Your peak power numbers stop being inflated by bogus single-second spikes. After cleaning, your real peak 5-second power is the highest real 5-second average, not whatever 2300W glitch happened during your 200W tempo ride. This is the whole reason to clean.

Yes. Trainer files have their own peculiarities - some smart trainers report power in 0.5-second bursts that look like spikes to a 1Hz analyser. The tool handles this correctly because it works on the raw values; if your trainer reports a brief 2100W on a 200W steady state, that is a spike worth removing.

Not after download - keep the original file. Drop the threshold a little less aggressive and re-run. The before-and-after charts in the tool show exactly what is being removed, so check them before downloading. If the chart shows your real efforts getting clipped, raise the threshold.