Inside the Tora training algorithm

Background

Programming exercise is just as much a science as it is an art. It is an art because, so often, you cannot decisively say that one option is better than another. A major part of the programming is adapting and individualizing the program for the athlete. To effectively personalize programs is a hard task, which is why most programs and apps still offer fixed programs. Some apps have tried to close that gap by letting LLMs do the heavy lifting of the thinking, and that creates a problem of its own.

A language model can write a program that reads as if a coach wrote it, but writing the plan is the easy part, and adapting it is where the programming is actually done. When exercise scientists and coaching experts have graded programs written by language models, they found plans that were safe and sensible on the page but generic, which came out much the same for a man and a woman with the same profile, leaned toward caution over effectiveness and only got better as the person typed in more detail [1, 2, 3]. In one study the recommendations covered well under half of what a complete exercise prescription should include [4], and in another the model's ability to fine-tune a program through ongoing interaction was judged not fully satisfying [2]. Language models also tend to agree with the person asking them [5, 6], so an athlete who asks to add more work is more likely to hear yes than to be asked how they have been recovering. Most importantly, a plan from a language model is judged by whether it sounds right, and nothing in the model itself is checked against how athletes actually did on its plans over the months that follow, which is the only test that matters in training.

Very few algorithms yet use evidence-based personalization for athletes, and that is the problem our team at Tora was highly motivated to tackle head-on. We want to provide the level and quality of adaptability that an effective coach brings to an athlete, while keeping the advantages that softwareistic algorithms allow for.

Practical Application

Practically, Tora works the way we built it. We built the algorithm with exercise scientists and coaches, and the algorithm builds a user model, which is Tora's working picture of you, covering your goals, schedule, equipment, training history and how your body responds to training. Calibration fills the user model first and gives the algorithm its starting point, down to a conservative first weight wherever there is enough information to estimate one, and anything you skip is recorded as an assumption rather than treated as known.

From there it runs as a loop. You train, the algorithm reads the results from your check-ins and from how you are progressing, and it adapts the sessions that follow using rules drawn from the studies and from our coaches' experience. After each session Auto-Load sets your next weight and reps and treats effort ratings with caution, since lifters misjudge how many reps they had left by about one [1, 2]. Every suggestion uses weights your equipment actually has and gives way the moment you type your own number, injuries and pain are handled before anything else, and recovery is considered before volume is ever reduced. The strategies that trainers with many clients use to keep people on a program are built into how it plans and adapts.

RelatedAuto-LoadYour numbers from your last session, filled in for you.

We lean toward having people train a bit harder rather than a bit easier, because muscle growth improves as sets end closer to failure and more weekly work tends to build more muscle [3, 4], and safety always comes first.

A coach can still watch your form and talk with you. Nova, the part of Tora you can talk to, covers much of that conversation. Where Tora has the advantage is that it keeps track of everything, every set you log and, with Apple Health connected, far more of what you do than one person could follow. Nova passes your requests to the algorithm, which computes every number itself.

Development and Validation

Proving that a training algorithm works is the hardest part of building one, and validation is where the program is actually made. Everything encoded from the literature into a user model is theory until it has been validated, with data from real people following the program and with coaches who understand what is happening in front of them. A model can agree with every study it was built from and still be wrong about the person using it, and only validation shows which parts those are.

The lesson the post-training of AI models keeps teaching is that validation is make-or-break. Language models became so strong at code largely through reinforcement learning in closed loops, in which the model writes code and the code is run against tests, with the result fed straight back as its reward. The approach is known as reinforcement learning with verifiable rewards [1, 2]. Where an answer cannot be checked that simply, the same idea is extended with rubrics, which are written criteria an expert would use to grade the answer [3]. Our simulations borrow both principles. Thousands of automated checks run against the engine every time it changes. A rubric built on how good coaches judge a program grades the programs it writes for a set of synthetic athletes, and a simulator runs synthetic athletes through sixteen weeks of training on the real engine to check that it keeps behaving the way a competent coach would.

How the algorithm is checked

Check	What it does
Automated checks	Thousands run against the engine every time it changes.
A coaching rubric	Grades the programs it writes for a set of synthetic athletes, the way good coaches judge a program.
A simulator	Takes synthetic athletes through sixteen weeks of training on the real engine, to check it keeps behaving the way a competent coach would.
Trainers	Coaches who have worked with many archetypes of athletes took the user model apart and found what did not hold up.
Athletes over time	Followed over long periods, with the results reported whichever way they fall. Still to come.

The same field shows where language models belong in Tora. Post-training is at its strongest where a clear winning answer can be checked, which is why these models have become so good at code and math, and it becomes much harder to apply when the literature conflicts and experts hold different opinions, because there is no clear winner to train toward [3, 4]. Programming exercise is an art for exactly that reason, since so often you cannot decisively say that one option is better than another. So language models in Tora do what they are excellent at, which is classifying and parsing what people say. When you talk to Nova, our architecture understands what your query is and feeds it to the right parts of the algorithm, but it is never the LLM that decides how your program adapts, and every number comes from the algorithm.

A program like this cannot be built by writing a plan and asking an AI, or anyone else, whether it is good, again and again. The reviewer can only judge the plan on the page, and the page is not where programs fail. To know whether a program works you have to run it, with real people following it for months and synthetic athletes taken through it week by week, and then look at what actually happened. Running it shows a deload that arrives too early, a muscle that quietly stopped getting enough work, or a block that unravels after one missed weekend, and it shows the places where an algorithm is technically correct by the science and still breaks down once people try to follow it. It also shows how a program feels to the person doing it, which decides whether they are still doing it three months later. That is why validation has to run through people as well as code. Trainers who have worked with many different archetypes of athletes took the user model apart with us and found the parts that had been built from the literature alone and did not hold up. Our exercise scientists are now working on how to make a program fun to stick to. It is slow work, and it is the part we are proudest of.

The reason is that an algorithm optimized for the science is optimized for the science, and not for your training. Your training is not the science. The two are related the way computer science theory is related to engineering, where the principles are fused together and neither gets far without the other. They are still separate things that are approached in separate ways, and proving that an algorithm is correct does not make it run well on a real machine for real users. As the anthropologist Marilyn Strathern put Goodhart's law, "when a measure becomes a target, it ceases to be a good measure." Agreement with the research is a good measure of a program, but once an algorithm makes that agreement its target, it can satisfy every study and still hand someone a program they cannot follow or stop progressing on, because the real target was always one person training well for months. No amount of foresight can measure that. Only validation can, by putting the program in front of people and watching what happens over time.

Bridges show the same idea. Picture a thousand identical bridges on one afternoon, each carrying its own mix of traffic and wind. The average strain across all of them may be perfectly safe, but nobody drives across the average bridge. Each bridge lives through its own single run of loads, decade after decade, and a crack that starts in one bad winter is still there the following spring. Physicists call a system ergodic when the average across many copies at one moment matches what a single copy experiences over time, and many of the systems that matter most are not ergodic [5, 6]. Training is one of them. Research reports what happened to a group on average, while each athlete lives through one path of weeks, in which a missed session, a bad night or a deload in the wrong place carries forward into everything after it.

Each athlete lives one path, not the average

What all of this validates is a decision architecture rather than a list of rules. The science gives principles, ranges and caveats, but it almost never gives a decision, and turning it into one is the coach's job, which is why a good coach is above all a good decision-maker who reads the nuance in front of them from first principles rather than from a template. The broad shape of a system like this is public knowledge, but the long run of judgment calls inside it is not. The first flowcharts for it were drawn by hand nearly three years ago, before AI could help with any of it, and it has been revisited many times since. Much of that work sat where the literature leaves gaps, such as which finding carries more weight for this person, what to do when the evidence points two ways and what to do when it says nothing at all. Making those calls show up correctly inside the app was its own engineering project, with calibration seeding first weights, substitutions inheriting history, missed sessions handled gracefully and every gym's real equipment respected, all alongside the adaptive nutrition targets, sleep and recovery check-ins and whatever an athlete tells Nova. We tried many strategies to shorten the validation loop, and none of them removes the need for time, a wide range of people and patient iteration.

None of that yet answers the question that matters most, which is whether it works. Testing whether any training algorithm produces better results than a well-designed fixed plan over the long term is genuinely hard, and as far as we can find, no one has published that test yet, us included. We intend to find out, with athletes followed over long periods and results reported whichever way they fall. We think this approach is the best chance of showing that it works, and we know we have not shown it yet.

Limitations and Scientific Oversight

As the mathematician Cathy O'Neil once wrote, "models are opinions embedded in mathematics," and although she wrote it about algorithms that make decisions about people, it is the most useful warning we know of for one that makes decisions about your training. Ours has opinions in it too, in the grades, the judgment calls and the order its rules are checked in, which is why the honest thing is to keep those opinions visible and keep people who can challenge them close to it, and why the exercise scientists sit inside the process rather than beside it. They decide what the evidence supports and for whom while the engineers make it exact and testable, and wherever the evidence runs out the call is recorded as judgment so it can be argued with later.

Some of the hardest problems are still open. The evidence is thinnest exactly where adaptive training lives, and nobody has yet shown that adjusting a program day to day builds more muscle than a well-designed fixed plan,[1, 2] so we do not claim it does. The match between you and a study is still made mostly on training age, which is a coarse stand-in, since two people with the same years under the bar can differ in almost everything else. Every input is noisy, from an effort rating that is off by a rep to a check-in skipped on a busy day. The algorithm cannot see you move. And our testing so far shows that the engine behaves the way we intend, not yet that it outperforms anything.

There is also a flaw in any program that is purely optimized by the science. The most optimal program on paper is not always one that people can follow, enjoy or keep doing, because a program that looks perfect in a study can ask for more time, more variety or more discomfort than a real life allows. A plan that is abandoned after a month helps nobody, which is why we treat optimizing for the science alone as a limit to design around rather than a goal in itself.

Sources [1, 2] Hickmott, L. M., et al. (2022), Sports Medicine Open; Helms, E. R., et al. (2018), Frontiers in Physiology. O'Neil, C. (2016), Weapons of Math Destruction, Crown.

Comparison with Existing Apps

Most training apps today promise a customized and personalized program, which is why repeating their marketing back to them would not tell you very much, since we would be claiming the same thing. What separates these products is the detail of what each one actually adapts on, how its programs are generated and whether it has been tested over the long term.

You can use any of these apps, and that is okay with us, since each of them does well what it set out to do and the people behind them are trying to help others train better. The point of this comparison is that if you want something that actually adapts using science, which learns from you and understands your goals and helps you meet them, and which gets better the more you train, then that is what we have built Tora to do. It will even teach you in the process, and it is reviewed by exercise scientists. This matters most for athletes who are chasing the small and difficult things that push them further, which is where we believe Tora excels, not because we think less of anyone else but because we have done the due diligence and the work.

Conclusion

What we have built is an algorithm that weighs a great many variables about you at once, draws on the literature through evidence grading and has been tested over a long period of work, which as far as we know very few training algorithms, if any, have been. A great deal of effort went into it, and we are proud to release it.

It was also built to make training easy and enjoyable rather than only optimized by the science. As we briefly went over in the limitations section, the most optimal program on paper is not always one that people will follow. That is why we worked with physical trainers who have many clients and with our beta testers, asking not only whether a program was optimal but whether people could follow it, would enjoy it and would keep doing it. Those answers are built into how the algorithm plans and adapts your training, and the rest of the app is designed so that returning to your plan is as easy as possible. We want the hard thing to be the training, so that when you go to the gym you can train hard without having to worry about anything else.

Reaching any goal, whether that is becoming a better athlete, getting healthier or eating better, also depends on education and understanding, which is why we are not trying to take the learning out of training. Nova can explain the reasoning behind your program and the studies it rests on, and the program guide teaches you everything we are doing for your specific program, including the sources, how we interpret them, the other ways they could be interpreted and how other coaches might program the same thing.

We are not asking you simply to use our algorithm and trust us. Our aim is to change the world of training by bringing together in one place what would otherwise take years to gather, with scientists analyzing the literature and practical experts turning it into programs. Open feedback is always welcome and the algorithm will keep improving, and if you would like to join our beta tester community for early access to new features, you can reach us through our contact page.