Evidence / How we know what we know
Coaching built on what your data shows.
Not on what sounds good.
Your heart rate swings more from sleep, heat and stress than from a week of training. A coach that reads every swing as fitness chases noise. Below: what held up when we tested it, what didn’t, and what that changes in your plan.
- public rides tested
- 127,730
- riders in that dataset
- 552
- claims we make
- 6
- claims we refuse
- 8
By Jim Camut · Former pro and Johan Bruyneel Cycling Academy racer · Updated October 8, 2026
Figure 1. Why we show you a range, not a number
Same rider, same power, twelve weeks.
The dots are what a heart-rate monitor reports ride to ride. The line is what fitness actually did. The change you care about is smaller than the noise around it.
Drawn to the measured numbers: a 6.8 bpm ride-to-ride swing around a real effect of one to two beats. The swing was measured across 127,730 rides from 552 riders in the public Golden Cheetah OpenData set, then repeated on a second group with the same result.
What we’re willing to tell you
Four things that held up when we checked them.
Some were written down as predictions before we looked. Others come from measuring our own riders. Each one says what it rests on.
Figure 2
What a break actually costs
Extra beats per minute at the same power when you come back, by how long you were away. The steps are small, and they are real.
449 riders, 1,461 separate breaks. As a check, ordinary weeks with no break went through the same test and showed no rise. That is how we know the step comes from the break, not from our method. The 10–14 day step is small enough that it could be zero.
Figure 3
At most one rider in eight
About a quarter rise more after a break, but most of them are within their own normal swings. At most 12 in 100 clearly drop further.
380 riders with three or more breaks each. We can’t tell which group you’re in from a form, so the coach measures you instead of assuming.
We’ll tell you
Take a break and your aerobic fitness dips. It comes back.
After a break, your heart works a little harder at the same power, and longer breaks cost more. Ten to fourteen days off barely registers. Two to four weeks adds about 1.3 beats a minute, a month or more about 2.2. It fades back to normal after about three weeks of riding.
Based on: 449 riders and 1,461 separate breaks, in a public dataset anyone can download.
We’ll tell you
Some riders lose more after a break. It takes measuring to know who.
About a quarter of riders show a bigger rise after a break. But for more than half of them, that rise sits inside their own normal ride-to-ride swings. At most one rider in eight clearly stands out. We can't tell which one you are from a form, so the coach measures you instead of assuming.
Based on: 380 riders with at least three breaks each.
We’ll tell you
Your heart rate bounces around a lot more than your fitness does.
Ride to ride, at the same power, heart rate swings by several beats for reasons that have nothing to do with training: sleep, heat, caffeine, stress, how the day went. Real fitness change is smaller than that swing. It's the single most important fact about reading your own data, and it's why we show you a range instead of a number that jumps around.
Based on: 127,730 rides from 552 riders. Measured twice, on two separate groups, with the same answer.
We’ll tell you
The same break costs different riders very different amounts.
Across 17 breaks over seven seasons, one of our riders came back with heart rate about two beats higher at the same power. Another rider barely moved after the same kind of break. An average hides that spread, so the coach reads your own comebacks instead of assuming.
Based on: Our own riders, measured across seven seasons. One of them is the honest exception: they don't detrain over their usual breaks.
What your power shows
Your watts, tested the same way. Including where our model loses.
The findings above read heart rate. These read power: your curve, your FTP, and the point where a model stops being useful.
Figure 6
Three guesses at your next 5-minute best
How far each guess missed the best 5-minute power a rider actually put on file over the next 90 days. Shorter bars are better.
222 riders in a public dataset. The curve and the recent best both read only the 90 days before; neither saw the days they were scored on.
We’ll tell you
Your power curve beats an average rider's numbers, for about two riders in three.
From your best efforts between 3 and 20 minutes, the coach fits a curve of how long you can hold each power. We tested that curve on the next 90 days of riding, which it had never seen. At 5 minutes it missed a typical rider's real best by 7.5%. An average rider's numbers missed by 13.7%. The coach uses the curve to check that a whole session is one you can finish, and eases it if it isn't.
Based on: 222 riders in a public dataset, each scored on 90 days the curve never saw.
We’ll tell you
Your starting FTP comes from your rides, not your birthday.
Age matters. Threshold power per kilo falls about half a percent a year, and riders 65 and up sit about 18% below riders 35 to 44. But inside any one age group, the middle half of riders spans about a full watt per kilo. So when your rides carry power, your starting FTP comes from them, or from the FTP you already set in your app. Age only shapes the first guess when there is nothing to read.
Based on: 1,249 riders with 20 or more powered rides each, in a public dataset. Only 23 are women, too few to measure separately.
We won’t say
“Our model knows your two-minute power.”
It doesn't. At most durations we tested, a rider's own best from the last 90 days predicted the next 90 better than our curve did. Outside the efforts the curve is fit on, it got worse: at 2 minutes it missed by 19.6%, more than an average rider's numbers at 14.9%. Part of that is the data, since nobody in it was asked for an all-out test. So short, hard targets start from your own best at that duration, not from the curve.
What we won’t tell you
Seven things a training app could say to you. We tested each one. None of them held.
Several of these are on other apps’ home screens today. Here’s what happened when we tried to back them up with data.
We won’t say
“Here's what your fitness will be next month.”
We built the forecast. Then we tested it against two simple guesses: a straight line through your recent numbers, and “next week looks like this week.” It beat both for 49 riders out of 100. That's a coin flip, so we deleted it. Plenty of apps still show one.
Figure 4
Our forecast against a coin flip
Our forecast won49A coin flip50riders out of 100 · full bar = 100We won’t say
“We've learned your personal recovery timeline.”
It sounds great. It isn't measurable. We tried 20 different settings for how fast a rider absorbs training. For the typical rider, 14 of the 20 fit the data equally well. When that many answers look equally right, you don't have an answer. Published studies run into the same wall.
Figure 5
14 of 20 settings fit equally well
We won’t say
“This block will add X% to your five-minute power.”
We can't give you a number here, in either direction. When we measured it, the answer changed depending on which weeks we compared, and riders who did no intervals at all showed similar moves. We'll tell you a block worked. We won't attach a percentage we can't stand behind.
We won’t say
“Polarized training beats the way you're training now.”
We ran this one as a check on ourselves: if our methods were finding patterns that aren't there, this is where it would show up. It came back empty every time, at every size of dataset. Whether polarized training is better is a real debate. Our data doesn't settle it.
We won’t say
“Your long rides built your durability.”
True as general physiology, and we'll happily explain it that way with a citation. But as a claim about your specific data, it didn't hold up: right direction for some riders, wrong direction for others. “Science says” and “your data shows” are different sentences.
We won’t say
“A hard week hurt your recovery.”
We found this in an early look at a few months of data and got briefly excited. When we re-ran it across the rider's full history, it vanished completely. It was an artifact of where we happened to start looking.
We won’t say
“It was the heat.”
Heat is real and we track it. But it explains only a small slice of why a heart rate reads high, and for some riders almost none. We'll mention it as one possibility. We won't hand it to you as the explanation for a bad day.
How we test our own claims
None of this is exotic. All of it is unusual in training apps.
- Predict first. Then look.
- We write down what we expect to find before we run a test, lock the file, and log every change we make after that. It's easy to find a pattern in ride data after the fact. It's hard to find one you called in advance.
- Beat doing nothing.
- Every result has to outperform the dumbest alternative, like drawing a straight line through your recent numbers. A surprising number of training claims don't.
- Audit the wins, too.
- We re-check the results we liked, not just the ones we didn't. That review overturned three conclusions we had already reached, one of them days from publishing. Nothing on this page skipped it.
Everything above is about riders in general. This part is about you.
What the coach learns about you
A small set of numbers, all read from your rides.
Here’s how they get set, what stops them from over-reading thin data, and which changes need your OK before they touch the plan.
- 01
It keeps 16 numbers about you, and every one comes from your rides.
Not a questionnaire. Your repeatable power, how much training you actually absorb in a week, how many hard days suit you, how well you hold power deep into a long ride. They update once a week from the riding you did that week.
- 02
Statistics set the numbers. The AI only explains them.
Every learned value is computed by ordinary statistics, the same kind of math used to track a signal through noisy measurements. The AI never sets one and never adjusts one. It reads the finished numbers and tells you what they mean in plain words.
- 03
Thin data makes it say less, not more.
Each learned value needs a minimum amount of evidence before it gets used. Below that line it quietly falls back to a sensible starting point and tells you nothing personal, instead of dressing up a guess as your number.
- 04
Easing off is automatic. Asking more of you needs your OK.
If your rides say the plan is too much, it backs off on its own. If they say you could take more, that waits for you to approve it. Same rule for your FTP: it can rise by itself, but a drop is your call, and a number you typed in always beats one we estimated.
- 05
When something we built doesn't beat a plain number, we switch it off.
We once learned how fast each rider's weekly training could safely rise. Tested against one sensible fixed number, the personal value won for only 35 riders in 100. So the fixed number is what everyone gets now.
Your numbers
Updated weekly
Demonstrated weekly load
590TSS
Well-established21 build weeks at 590–675 TSS, none with a failed quality session.
Hard days per week
3days
Calibrated8 weeks where three quality sessions landed and a fourth did not.
Fatigue red line
Using the starting point
Still learning2 weeks of evidence so far. Using the starting point until there is more.
Eased on its own
Hard days per week 4 → 3 days
Backing off never waits.
Proposed · waiting for your OK
Raise critical power 265 → 272 W
A mock-up, built from the same parts as the settings page every athlete gets.
What this means for your training
Three things you’ll notice
in the first week.
Too small to read? It says so.
Missed the set? It checks the plan first.
Formulas do the math. The AI explains it.
See how that compares to the apps you already know, or read why a bad week is so hard to read from your own data.
Common questions
The questions this page usually raises.
Why would a training app publish what it can't do?
Because a claim you can check is worth more than a claim you can't. Every coaching app tells you what it knows. Almost none will tell you where the evidence runs out, which leaves you no way to judge the rest. We'd rather be the one you can verify.
Does AdaptCycling predict my future fitness?
No. We built a fitness forecast and tested it against two simple guesses: a straight line through a rider's recent numbers, and "next week looks like this week." It beat both for 49 riders out of 100. A coin flip isn't a forecast, so we removed it. We do show the training load planned ahead of you, which is arithmetic, not prediction.
Where do these numbers come from?
Two places. Our own athletes' ride histories, several seasons each, and the Golden Cheetah OpenData project, a public set of ride files anyone can download and re-analyze. For the public dataset, we wrote down what we expected to find before we looked, and we logged every change we made after that.
What happens when a result stops holding up?
We remove the claim and update this page. It has already happened: reviewing our own results overturned three conclusions we had reached earlier, including one we were close to publishing.
Get started
Your next workout
shouldn’t be a guess.
Free 14-day trial, no card.