Our approach to readiness, evidence and judgment
This document describes how RAMTT turns what your body did into what it tells you, and why it is built the way it is. It is the reasoning behind every number in the platform: which things we measure, how we grade the evidence behind them, when we let a rule decide and when we do not, and what we have chosen not to do at all.
It is written for the athletes and coaches who use RAMTT, and it is deliberately long. A platform that asks you to trust its judgment owes you the reasoning, and the reasoning does not fit in a tooltip. We revise it when the evidence moves, and we mark what moved.
If you want the shape without the detail, the summary covers the four systems we model, the five evidence grades, and the handful of rules we keep, in about four minutes.
Most of what a training platform tells you is a count. Kilojoules, hours, a stress score, a form number. Counts are useful, and we compute all of them, but a count answers the question of what you did, not the question you actually have at seven in the morning, which is what your body can do today and what it will cost you to find out.
RAMTT's answer is a model of the athlete, not a scoreboard. It keeps a running estimate of the things that actually change from hour to hour: how much fuel you have, where your thresholds sit today rather than on the day you last tested, how much reserve you have above them, and how recovered the systems are that you will lean on tomorrow. Each estimate is a range, not a point, and the width of that range is part of the answer. A narrow range after a lab test means one thing; a wide range after three weeks of unlogged rides means another, and the platform says so.
This is also why RAMTT grades its evidence. A number that comes from a measurement is not the same kind of number as one that comes from a population average, and neither is the same as a number the model has inferred from yesterday's ride. We show you which is which. Not because measurement is always better, but because you cannot decide how much to trust a recommendation unless you know what it stands on. Most platforms hide this, and the result is a score that looks equally confident on its best day and its worst.
There are two broad ways to build the thing that tells an athlete what to do. One is to write rules: below this form number, rest; above this ramp, warn. The other is to model what is going on inside the person, carry the uncertainty along with the estimate, and let the advice follow from the evidence. Rules are transparent, easy to audit, and hard to fool, and we keep a small number of them where the cost of being wrong is high. Everywhere else we prefer the model, because the athlete in front of us is never quite the athlete the rule was written for.
The list is long and it grows, but it has a shape. Work is what you did: power, duration, the kilojoules, and how the intensity was distributed within them. Supply is what you had available to do it with: the fuel you took on before and during, the fuel you had stored, and what the two together imply about what is left. State is how the body responded: the drift in heart rate against power over the course of a ride, the way a threshold settles as fuel runs down, and the signals you give us each morning about how the night went. Everything else, from a race-day plan to a warning about the coming week, is built from those three.
None of these are read from a single sensor. A threshold is estimated from many rides and moves slowly; a fuel level is estimated from a few hours and moves fast; a morning signal is one number from one person on one day. The model weighs them differently because they deserve different weights, and it says how much each one moved the answer. When two sources disagree, we do not average them into a comfortable middle. We show the disagreement, because the disagreement is usually the most informative thing on the screen.
It follows that RAMTT is slower to be sure than the platforms you may be used to. It is quicker to notice, and slower to conclude, and we think that is the right way round for a tool that will be trusted with a season.
Every claim in RAMTT carries a grade. The grade describes the state of the research behind the claim, not its importance and not our confidence in this particular athlete. A claim can be early and still be the best available guide, in which case we say both things. We would rather show you an honest early than a dressed-up established, and we would rather you argue with us than take the number on faith. The grades are:
The grade sits next to the number, not in a footnote. You will see it in the app, in the weekly review, and in the science notes where each claim is written out with the studies behind it. When a grade changes, because a replication landed or failed, the note changes with it, and the app tells you what moved.
This section walks through how a grade is assigned to each engine, what happens when the evidence behind it moves, and how the platform behaves while a claim is still early. We expect it to be of less interest to most readers, so it is collapsed by default.
An engine is a piece of the model that turns inputs into one kind of estimate: a threshold from a set of rides, a fuel level from a ride and what you ate, a recovery state from a night and a morning. Every engine has a grade of its own, and the grade is attached when the engine is written, not when it is shipped. It is written down with the studies it rests on, the populations those studies used, and the assumptions we had to make to get from the paper to the person. If we could not write that page, the engine does not ship.
We grade at the level of the claim, not the level of the feature. A single screen can show three numbers with three grades, and it usually does. The threshold on a ride might be established; the estimate of how far it drifted in the last hour might be supported; the amount of reserve you still had at the top of the climb might be our own. Showing one grade for the whole screen would mean choosing the best or the worst, and both would mislead.
The grade is set by two people who did not write the engine, against a written rubric, and the disagreement between them is recorded. A grade that the two do not agree on is set to the lower of the two until it is resolved.
An early claim is shown, not hidden. It carries its grade, it carries a wider range than it otherwise would, and it is never the only reason for a recommendation. If the only evidence for a piece of advice is early, the advice is phrased as a question rather than an instruction, and the app tells you what would settle it. A coach can turn early claims off for an athlete entirely, and some do. We think that is a reasonable choice and we make it easy.
The grades are a floor, not a ceiling. An established claim can still be wrong for you, and the range around the number is where that shows up. The grade tells you how much the field knows; the range tells you how much we know about you.
We generally favour judgment over rules, and where a rule exists in the platform we try to explain why it is there and what it costs. A hard stop before a race, for instance, is a rule, and we keep it because the cost of being wrong on race morning is high and the benefit of cleverness is low. Most other decisions are not like that. Whether to push a session or hold it depends on things a rule cannot see, and the right answer for you is not the right answer for the athlete next to you with the same numbers. A model can hold that; a threshold cannot.
This has a consequence for how RAMTT talks to you. It does not issue verdicts about who you are. It describes what it sees, how sure it is, and what would change its mind. If that sounds less decisive than a red badge that says rest, it is, and deliberately so. The badge is easy to build and easy to ignore. A description you can check against your own body is harder to build and harder to dismiss, and it is the only kind of advice that survives contact with a real season.
There will be days when the model is wrong about you. It will be wrong less often than a rule would be, and, more importantly, it will be wrong in ways you can see coming, because the range will widen before the answer breaks. That is the whole point of carrying uncertainty instead of hiding it: the platform gets to say I am not sure before it has to say I was wrong.
It will not invent precision. If a number rests on a population average, it says so, and it does not show you two decimals of a guess. It will not hide a disagreement between its own sources to look tidy. It will not use colour alone to tell you something matters, and it will not tell you the same thing in two different colours on two different screens. And it will not describe you with a label you did not choose, because a platform that names you has stopped measuring you.
Everything above is a description of intent, and intent is cheap. What makes it real is that each of these commitments is a thing you can check: the grade on a claim, the width of a range, the source behind a number. We have tried to make RAMTT a platform that shows its work, in the belief that an athlete who can see the reasoning will train better than one who is handed a score. If you find a place where it does not, tell us, and we will show you what we changed.
Last updated September 2026
Most of RAMTT is judgment. These are the places where we decided a rule was safer than a model, and why.
In the last seventy-two hours before a race the platform stops suggesting additional load, whatever the model says about your readiness. The cost of a bad call on race morning is high and the upside of a clever one is small, so a rule wins. You can still log what you want; we just stop asking for more.
The stop is visible on the calendar and in the plan, and it lifts the moment the race is logged.
If you mark yourself ill, the plan drops to nothing until you mark yourself well, and the first two days back are capped regardless of how good the morning numbers look. Illness is the one place where feeling fine is not evidence, and we do not let a good check-in override it.
The model will not estimate your body composition, your menstrual cycle, or any medical condition from training data, even where a correlation exists in the literature. These are things you tell us, or we do not know them. The privacy note lists what is inferred and what is only ever stated.
A coach can override any recommendation and any grade for an athlete, and the override is shown to the athlete with the coach's name on it. A coach cannot hide the model's own answer; the two sit side by side. We think the disagreement between a coach and a model is information the athlete should have, not a thing to be tidied away.
Deletion is a rule, not a request. When you delete an account, the raw files, the estimates built from them, and the validation copies are removed within thirty days, and we tell you when it is done. Nothing you delete is kept for the model's benefit.