How pace24 Works
The Coach Offered a Lighter Week It Could Not Build
After a break, a normal plan asks for more than your fitness now carries. Our coach says so and offers an easier week. Testing found it never read its own notes, and could not build the week it offered.
The pace24 team7 min read
Key takeaways
- After a week off, a plan that carries on as normal asks for more than your fitness now carries: 124 percent after 7 days away, for our made-up rider.
- The coach now says so as a fact, reads a short note on returning to training, and offers a lighter first week. You decide.
- First test: the coach read the note in 0 of 12 break turns, because the instruction sat in the tool's description. Moving it into the tool's result took that to 12 of 12.
- Second test: when the athlete said yes to the lighter week, 1 of 15 got one. The planner tool could not plan a smaller week. After the fix it was 15 of 15.
- Our checks passed all of it. Reading the replies still found false claims.
About this post: the riders, messages and scores are all made up. The coach in these tests is an AI model, Anthropic's Claude Haiku 5.5, with parts of the app stubbed out, and the checks are partly code and partly our own reading. Nothing here describes a real athlete.
Take a made-up rider with a fitness score of 60, whose plan asks for a load of 441 this week. Now take seven days off.
Fitness does not vanish in a week. It fades, and on the usual 42-day model a week off takes about 15 percent of it. That is the whole arithmetic.
| Days off | Fitness | Same week asks |
|---|---|---|
| 3 | 55.9 | 113% |
| 5 | 53.3 | 118% |
| 7 | 50.8 | 124% |
| 10 | 47.3 | 133% |
| 14 | 43.0 | 147% |
| 19 | 38.2 | 165% |
The last column is the week's load as a share of what the faded fitness carries (seven times the score, so 356 at day 7). The plan was not wrong to ask for 441 before the break. Nobody had told it a break had happened.
What we built
After four days or more without a recorded session, the coach is told how many days, and told when the week it is planning asks more than your fitness carries. It must say that as a fact, read a short note we keep on returning after missed training (or after illness, if you said you were ill), and offer a lighter first week beside the normal one. Nothing is cut, capped or moved until you choose.
The lighter week keeps your days and your key session and trims the volume, because detraining reviews say volume can come down further than intensity or frequency. We found no controlled study of catching up on missed sessions against letting them go, so the coach is never told to stack them.
Test one: it never read the note
We wrote four made-up cases, each run three times: a cold, a twelve-day holiday, nine silent days, and six days of exams in Dutch. We added a no-break control, then scored each reply on seven checks.
Before any fix, every one of the 12 break replies failed at least one.
| Check | Break turns passing |
|---|---|
| Plans the week | 12 of 12 |
| Flags that the week asks too much | 12 of 12 |
| States the break | 12 of 12 |
| Compares the load | 12 of 12 |
| Leaves the choice to the athlete | 12 of 12 |
| Offers a lighter week | 9 of 12 |
| Reads the note | 0 of 12 |
The note was never read. The planner tool was the only tool called in all 15 turns. We had put the instruction, "read this note before you reply", inside a long description of the tool. The model finished the tool call and replied straight away.
The fix was to move the instruction from the tool's description into the tool's result, where the model looks next. Reading the note went from 0 of 12 to 12 of 12. Offering a lighter week went from 9 to 12, and the Dutch case from 1 of 3 to 3 of 3.
Test two: the offer was hollow
We then made the test meaner. The athlete answers the offer: "Yes, give me the lighter first week." The test passes only if a lighter card appears.
It happened once in 15 tries.
The planner tool had no way to plan less. The coach asked it again and got the same week, load 376, again. It tried capping the athlete's available hours. It wrote days one at a time. It told one athlete "the planner returned the same card". The coach had made an offer it could not keep.
The tool now takes a lighter first-week load. Every training day stays where it is, the key session stays, and the volume comes down. For the club rider the week went from 376 to between 312 and 319 after 10 to 12 days off. The follow-up passed in 15 of 15 tries, and all 18 turns were clean.
A bug the scores did not show
The note had a defect that only showed in the text. It compared the season week's target, not the week on the card. Plan on a Thursday with nothing running and the card holds only the days left: load 292. The note still said "This week asks load 376, 114 percent". It warned about a load nobody had asked for, and in one broken run the coach explained the gap with a reason it had made up. The note now uses the load on the card.
What reading still caught
The final run passed 21 of 21 automated checks. We read all 39 replies anyway, and found things no check had seen:
- Three replies counted five or six rides in weeks that held four.
- Two replies described both weeks as about 480 minutes. The second week was 388.
- One called the athlete's stated eight hours a hard limit. Our rule is that stated hours are a plan, not a limit.
- One promised a heavy interval session once the athlete was back in rhythm. No tool had made that decision.
A clean score here means the checks passed. It does not mean every sentence is true.
The same rule on the Today screen
When sessions go unrecorded, the Today card is headed "Pick up from here". The first button reads "Continue with current plan". Moving a session sits behind a collapsed "Review an optional move" and needs its own tap.
Limits
Three runs per case, two made-up athletes. The database tools were stubs and the calendar was held in memory, so no real production turn was tested. The note reaches you through the coach's reply, not as a card.
Want a coach that tells you what a break did, then lets you choose? Get early access.
Sources
- Our own evaluation notes for the after-break coaching work, 2026-11-05 and 2026-11-06: four made-up cases plus a control, Claude Haiku 5.5 as the coach, three runs each, every reply read by hand.
- Mujika and Padilla (2000). Detraining: loss of training-induced physiological and performance adaptations. Part II. Sports Medicine 30(3), 145-154: doi.org/10.2165/00007256-200030030-00001
- The fitness figures in the first table are arithmetic on a made-up rider with the 42-day model, not measurements.
- No real athlete data, and no data of Strava origin, reached any model in these tests.