How the coach decides
The honest version, for people who want to know what is actually computed before they connect their data.
The numbers are math. The judgment is the coach's. Every number Milewise cites comes from a deterministic decision core: plain, testable arithmetic over your own data, computed fresh each time you talk to the coach. The AI reads those results as facts and decides, as a coach would, what today and the plan should be. Code enforces what must be true and what keeps you safe; the AI does not invent your readiness, your load, or your limits.
Your baselines, not population norms
Overnight HRV and resting heart rate are judged against your own rolling baselines, not against tables of what is normal for other people; sleep is judged by what last night actually delivered. A resting heart rate of 52 is an alarm for one athlete and a Tuesday for another; the only meaningful question is how today compares with your recent trend, and trends beat single readings: one noisy HRV night does not flip a verdict on its own.
What gets computed
- Readiness: an HRV-trend ladder over your baseline, combined with resting heart rate and sleep, resolving to green, amber or red.
- Training load: acute vs. chronic workload plus a sanctioned weekly ramp, so volume growth is judged against what your history supports, and a jump beyond it is put in front of the coach every time the plan changes.
- Plan compliance: what you actually ran against what was planned, by sport, so a stroll never silently completes a planned run.
- Goal feasibility: your current performance level against the goal and its date, including gates for debut distances.
These results reach the model as a fixed block of labels and numbers. When the coach cites a figure, that is where it comes from.
Missing data is not caution
If there is not enough data for a call, the verdict is no verdict, stated as such. It is deliberately distinct from amber: an athlete without wellness data would otherwise live in permanent artificial caution, which teaches you to ignore the coach. When Milewise does not know, it says so.
Hard rules live where plans are written, not in the prompt
Two things cannot be written, however the conversation went: a session on a day you said you cannot train, and a hard session while pain you reported is fresh. Everything else is measured on every write and handed to the coach as facts: the weekly volume jump, with 30 percent as the line the research draws; how many running days in a row your history supports; the finer ramp targets. The coach then changes the plan, or keeps it and tells you why in one sentence. This split is a deliberate design choice: safety rules that live only in an AI model's instructions can be argued with; rules in the write path cannot.
The research behind the rules
- Your baselines and trends: Plews et al., Eur J Appl Physiol 2012; Plews et al., Int J Sports Physiol Perform 2014; Buchheit, Front Physiol 2014.
- Training steered by HRV: Kiviniemi et al., Eur J Appl Physiol 2007; Nuuttila et al., Med Sci Sports Exerc 2022.
- The 30 percent limit: Nielsen et al., J Orthop Sports Phys Ther 2014.
- Acute vs. chronic load: Gabbett, Br J Sports Med 2016, read with the critique in Impellizzeri et al., Int J Sports Physiol Perform 2020.
What the AI model actually does
Language, context and coaching judgment inside those bounds: which session fits the week, how to adjust around your race, what your splits say about pacing. Four companies answer those requests over five routes, and where each one processes is set out route by route in section 8 of the privacy policy; who they are and what our contracts with them require is section 7. We train no models and pass no data on for training.
Limits, stated plainly
Milewise is an AI coach and can make mistakes. It gives training and fitness guidance only, never medical advice. If something feels wrong in your body, that signal outranks every number on this page.
More on the data itself: what the coach is given, and what it can change without asking, and what we process at all.