Assisted progression, not automatic

Load progression is the part of the job that slips first and shows last. A client repeating the same weight for five straight weeks complains about nothing: they train, they log, and the plateau surfaces two months later as a cancellation that looks like it came from somewhere else.

The work itself is not hard. Look at what the client hit, estimate what they can handle, write the number for the next session. What is hard is doing it twenty times a week, every week. That is where assisted progression comes from: let the system do the arithmetic and let the coach keep the decision.

Assisted progression is load progression where the software writes the weights and reps of upcoming sessions from what the client logged, within limits the coach set. In automatic progression the load also moves on its own. The difference is who wins when the calculation and the coach disagree.

What gets automated is arithmetic: estimating what someone can lift from what they already lifted, rounding that number to an increment that exists on the gym floor, and carrying it forward into the following weeks. None of that asks for judgment. It asks for time, and the time always comes out of the same place.

What does not get automated is the other half. A log says the client did eight reps at 60 kilos with two left in the tank. It does not say they slept four hours, that they are coming back from a lower-back episode, or that form fell apart on the last set. None of those three has a column in any database, and all three change what the right number is on Monday.

It is also worth separating two things that travel together in marketing copy and do not in the product. The progression math is a coefficient table and a calculation, not a language model. It is deterministic, auditable, and gives the same answer twice.

The per-client task is small and the sum is not. Opening someone's session, reading what they hit set by set and rewriting the following week takes a few minutes. With twenty clients it takes hours, always on the same day, and always the day that was already full.

When the calendar tightens, the week gets copied from the last one. The client trains and logs, but the program has stopped answering them. It still works as a calendar and has stopped working as a program, and nobody names the moment it happened.

The saving Kaizer publishes on this work is 2 hours per week, per client. It grows with the roster; it is not a flat total that appears because you installed something.

And the cost is not only the coach's. A client lifting the same weight for six straight weeks draws a reasonable conclusion — that the program is not taking them anywhere — and that conclusion lands well before any conversation about adherence.

Three numbers per set, and all three come from the client: the weight they moved, the reps they completed and the RIR they reported. RIR is reps in reserve, the ones they felt were left when they stopped. Helms and colleagues (2016) formalized that scale for strength work, and it is still the cheapest way to measure effort because it needs no equipment at all. The scale and its uses are worked through in RIR and autoregulation.

Those three numbers produce an estimated 1RM: the load the client could move exactly once. There is no need to test it, and that is much of the point. The relationship between how many reps you complete and what percentage of your maximum they represent has been tabulated for decades — the NSCA publishes its own training load chart — and Kaizer estimates the maximum with coefficients derived from those tables, indexed by reps completed plus RIR reported. Eight reps at 2 RIR and ten at 0 RIR land in the same cell: both sets describe the same ten-rep maximum.

If you want to see the mechanism from the inside, the RIR to percentage of 1RM conversion runs exactly that step by hand.

The number that comes out almost never exists in a gym: 63.4 kilos is not a weight you can assemble. So the target gets rounded to that exercise's minimum increment, which depends on the equipment — 2.5 kilos on barbells and machines, 2 on dumbbells, 4 on kettlebells, one repetition when the movement is bodyweight — rather than to some arbitrary round figure. The engine also separates primary exercises from accessories, because they neither progress at the same rate nor start at the same target RIR.

The full list has two halves: what enters the calculation and what stays out because it is not measured. That is where the math stops.

This is where the implementation meets the evidence. Plotkin and colleagues (2022) published an eight-week experiment in PeerJ with 43 trained subjects split into two groups: one progressed by adding load, the other by adding reps at the same weight. They measured muscle thickness at the end and found no significant difference between them.

Progressing in reps counts as progressing. That is the argument in progressive overload is not just adding weight, and the full definition of the principle sits in the progressive overload glossary entry.

The practical consequence is double progression, and it is exactly what the engine does when the number does not work out: if the target weight does not cover a whole increment, it neither forces the jump nor leaves the week identical. It adds a repetition and holds the load. That is a design decision, not a side effect.

The case where it shows most is the intermediate client. Someone a year into training can go months without earning another 2.5 kilos on a press. In a progression that only understands kilos, that client is stuck for six straight weeks. In one that also understands reps, they progressed six times.

The load does move on its own, in the weeks the coach has not touched, and it stops the moment they do. That happens through three concrete mechanisms, and all three sit on the coach's side.

The first is that the routine does not reach the client until you send it. A generated routine is born as the coach's draft and moves to the client through a human action. It is a switch: the setting that turns it off is called, literally, send AI routines without reviewing them, and it ships off.

The second is the one that changes the day to day most. The engine only rewrites weeks that are still marked pending and still authored by it. The moment you edit a week's weight, that week leaves its reach and does not come back. Your edit displaces the calculation permanently rather than competing with it every Monday. Automation fills in what you left untouched, and nothing else.

The third is the reach of the switch. Automatic progression turns on and off globally, and gets overridden client by client: the one coming back from an injury can sit outside it while everyone else carries on. That is how AI routine adaptation works in practice.

There is a fourth, smaller case you see constantly: when you add an exercise to a session, the system proposes a weight from the client's history in that movement. It is a proposal you confirm or overwrite, and nothing is saved until you decide.

The first is that reported RIR is an estimate, and estimates carry error. Steele and colleagues (2017) showed that the ability to predict how many reps remain before failure improves with training experience: a beginner who reports 2 RIR often had five reps left. Every calculation that starts at RIR inherits that error, and it shrinks by teaching the scale, not by improving the algorithm. It helps to know the margin is generous: Refalo and colleagues (2023) found comparable hypertrophy training with reps in reserve and training to failure, so being slightly off costs less than people fear.

The second is missing history. For a new exercise, or a client who has never done that movement, there is nothing to estimate from and the field stays blank. That is better than an invented number, and it means the first value for every new movement is written by the coach.

The third is that there are no automatically scheduled deloads. A cycle's target RIR undulates — it starts conservative and moves toward failure as the weeks advance — and that is not the same thing as a deload week. Coleman and colleagues (2024) studied exactly this: 39 participants, nine weeks, and a deload week in the middle cost no muscle growth compared with training straight through. It is a reasonable tool, but deciding when to use it stays with the coach, and the criteria are in the deload week.

The fourth is the broadest and contains the other three: the system knows nothing about what it does not record. That gap is exactly what the coach occupies.

The arithmetic gets automated: estimating the maximum from logged sets, rounding to an increment that exists on the floor, and writing the following week. So does the case where the weight cannot go up, by adding a repetition instead of forcing the jump.

What does not get automated is when a rise is warranted. The engine only writes where the coach did not, switches off client by client, and hands over drafts rather than decisions.

A system that does the arithmetic well frees up exactly the time the rest of it needs.