How AI Adjusts Your Training Plan to Your Data

Automatic plan correction does not need exotic data. Three signals suffice: the gap between planned and completed volume, a three-session e1RM trend, and reported RIR on compounds. Together they answer whether the dose fits your capacity to recover.

The decision rule is simple and follows from the shape of the dose-response curve. Volume works, but with diminishing returns [1], and each extra weekly set contributes progressively less [2]. So when the volume gap widens, the sensible response is to lower the dose, not to add sets and hope to punch through the stall.

Notice what is missing from that list. There is no resting heart rate, no heart rate variability, no wearable readiness score. Those can be interesting, but in strength programming they carry far less decision value than how many reps you actually completed at a given load.

The reason is simple: training signals measure directly the thing you want to change. Physiological indices measure a general state that depends on stress, illness and measurement timing enough to make attributing any change to training difficult.

A model sees numbers, not context. It does not know you spent yesterday moving flats, that your shoulder has hurt in one specific position for a week, or that you compete in three weeks. All of those change the right decision, and none of them follow from a set log.

That is why a well-built system asks for the missing context and shows the basis for what it changed. Load7 states the reason for every correction — "dropping two squat sets because e1RM fell across three consecutive sessions" — so you can reject the change when you know a cause the system cannot see.

This is not interface decoration. An unexplained plan change cannot be verified, and a plan you do not understand is a plan you will not execute consistently.

There is also a class of decisions a model should not make on its own, even with the data. Choosing between chasing a squat record and keeping a pain-free shoulder is a decision about priorities, not about optimisation. Systems that make it for the user tend to pick whatever is easier to measure.

The practical consequence: what matters is less "how good is the model" and more "where does its mandate end". A system that proposes and waits is better in that respect than one that simply rewrites the plan.

One bad session should not change the plan. Three consecutive sessions of declining e1RM should. A single off day is noise; systematically missing prescribed reps is signal.

The same applies upward: two good sessions are not a reason to add four weekly sets. Volume progression makes sense in steps of two sets per muscle, verified over four weeks — the curve flattens anyway [2], and fast jumps mostly buy fatigue.

The symmetry between adding and subtracting matters more than it seems. Systems that react instantly to good sessions and slowly to bad ones systematically overload — because good sessions also happen by accident, after good sleep and a big meal.

The safer arrangement backs off faster than it adds: two signals are enough to lower the dose, three are needed to raise it. The cost of one unnecessary lower-volume week is far smaller than the cost of three weeks of overload.