Is this number real,
or is my power meter lying?

A power meter that sticks on one reading for fifteen seconds, or drops to zero while you are still pedalling, does not leave a mark on the summary. The average for that interval is still a number in watts. The maximum for the ride is still a number in watts. Nothing about either tells you that part of it was never a measurement.

That is what makes bad power data dangerous: an average computed over a fault looks exactly like an average computed over good riding. This article is about the faults that hide that way, how to tell them from real efforts, and the one that no check on a single file can see.

Three faults that hide inside an average

A stuck reading. The meter reports the same figure, to the watt, second after second. Real riding is never that steady: even on a flat road at an even effort, power moves by a few watts from one second to the next. A run of identical values is the meter repeating its last number.

A dropout. Power reads zero for a few seconds while the cranks keep turning. Coasting also reads zero, so zero alone proves nothing; zero with the legs still going round does.

A fault unwinding. Around a stuck stretch, the readings either side often swing wildly: several jumps of hundreds of watts in a few seconds, reversing direction. On their own, big swings are ordinary riding. Next to a stuck reading, they are usually the same fault, and a ride maximum taken from them is not your sprint.

Why cadence is the referee

The obvious rule, "flag any long run of one value", is wrong indoors. A smart trainer in ERG mode holds power at its target by design, and a ride on one produces long runs of exactly the same figure that are perfectly real. The obvious rule for swings, "flag any jump of hundreds of watts", is wrong outdoors: stop pedalling into a corner and stand on the pedals coming out of it, and power genuinely goes from nothing to several hundred watts in a second.

What separates a fault from riding in all three cases is cadence. During an ERG hold, cadence stays close to steady. During a stuck reading, cadence moves by ten or more rpm while the power figure does not change, and legs cannot hold power to the watt through a cadence change that size. During a sprint out of a corner, cadence rises with the power; during a fault, power reverses against cadence that is going steadily one way. So cadence is the evidence, not a footnote.

What IntervalSync checks, and how it tells you

When an assistant asks IntervalSync for intervals, steady efforts or climbs, it checks the power stream first. It reports any range Intervals.icu itself has flagged as anomalous, and adds three of its own:

• power held at one exact non-zero value for 10 seconds or more while cadence moved by 10 rpm or more;
• power at zero for 2 seconds or more while cadence averaged above 20 rpm;
• three or more jumps of at least 200 W within 15 seconds, reversing direction, beside a range already flagged.

Each flagged range comes back with its start and end in seconds, and any interval, block or climb whose figures span one carries a warning on that figure, where you would quote it. A climb's warning is scoped to its power: a stuck power meter says nothing about the elevation trace, so the climb's gradient and VAM stay trustworthy.

To look at one yourself, ask for that stretch of the ride at full resolution. The assistant reads the raw samples for that range, so you see the repeated figure, or the zeros against a turning crank, with your own eyes.

Analyse [activity] and tell me whether any of the power data is untrustworthy before you report averages. For each flagged range, show me the raw seconds either side.

What it cannot catch

The check is deliberately conservative, and the price of that is written down here rather than discovered later.

• Without cadence, stuck readings and dropouts pass. A ride that records no cadence leaves nothing to compare power against, and flagging every steady stretch would be worse than staying quiet. Ranges Intervals.icu flagged are still reported, and so are wild swings beside them.
• A single bad second passes. One impossible spike with no stuck reading beside it looks like a sprint to every rule above.
• A meter that reads consistently high or low passes. An offset or a drifting calibration produces smooth, plausible data that is simply wrong by a few percent. No check on one file can see it, because nothing in the file contradicts it. Comparing two power sources on the same ride, or the same steady effort on known equipment, is how that kind of error shows up.

So a clean report means "nothing contradicted the power reading", not "the power reading is right". That is a smaller claim, and it is the one the data supports.

Why it matters beyond one ride

A fault in one file rarely stays in one file. A stuck reading inside a threshold interval moves that interval's average, and any trend or comparison built on it. A spike from a fault unwinding can sit at the top of a power curve until someone notices. That is why the check is on by default. For a ride's intervals it costs one extra request to Intervals.icu; for steady efforts and climbs the power stream is already loaded, so it costs nothing.

More questions the data can answer · Decoupling and block length · All articles · Connect your training