Training

Say your plan says "4 × 8-12" for bench press and this week you finished the last set with 40 kg for 12 reps. On the next set 42.5 kg shows up.

That 2.5 kg is the load step written in that movement’s prescription — applied once, the moment you touch the top of the band.

The step is not random but a field carried in the movement’s prescription: every planned movement has its own `loadStepKg`, defaulting to 2.5 kg (the smallest valid value is 0.1 kg). The decision passes through a single core that looks at three things at once: (1) did you reach the top of the rep band on the last set, (2) was the effort too high, (3) is this movement open to load changes. If you hit the top and the effort was reasonable, the load rises one step. If the effort was too high (RPE 9.5 and above) there is no increase — it drops one step instead; it also drops if you stayed two reps below the band. If none of these applies, the weight stays the same. For bodyweight movements the step is zero, so the weight never changes; but the moment you clip weight to a belt and enter it, the system sees that and opens load progression.

So "+2.5 kg" is a threshold, not a timer: it arrives the week you touch the top of the rep band and does not arrive the week you do not. If you want the weight to go up, your target is completing the top of the band — not doing more sets.

The system calls this the "load step", and the decision that applies it the "next-set load decision".

Say two users open the same plan; one is on an old version of the app, the other on a new one.

Both see the same load step for the same movement, because the step arrives from the server with the response.

The load step is a prescription: it derives from your level, your goal and the movement’s class. For a while this number did not exist in the weekly plan body at all and was produced entirely on the device — meaning a user’s weight progression was driven by a number their own phone wrote. When the decision sits on the phone, the owner of the phone can change it; that is a training-safety drift. The step is now carried as an explicit field in the day’s response.

It is impossible to be running on an old progression table because you did not update the app; when the rule changes, everyone moves to the new rule at the same moment.

The system calls this a "server-authoritative input".

Say you finished your session on Tuesday evening and tapped "Finish", then logged your weight on Wednesday morning. When you open the app on Thursday you see a coach suggestion.

The coach does not speak on its own; it wakes when one of a fixed number of defined events occurs.

The right to speak is bounded by a closed list: twenty-two events are defined across five areas, and nothing outside that list can trigger the coach. The six on the training side are: training day completed, training week completed, missed session detected, progression stall detected, overload risk detected, deload candidate detected. If the list were not closed, every small data change would spawn a notification and the coach would turn into noise — a coach that becomes noise gets muted, and a muted coach is useless.

A silent coach does not mean "broken": it means no event occurred to trigger it. The quickest way is to finish a session and close the week.

The system calls this "event-driven coaching".

Say you missed your session three days in a row and the system tries to produce the same suggestion each time.

You see that suggestion once; the rest are filtered out before they reach you.

A suggestion passes four gates before it reaches you. First, the confidence threshold: a suggestion whose own confidence score is below 0.4 is never even considered. Second, freshness: an expired suggestion is not shown, because old advice does not describe today. Third, the repeat filter: the same suggestion produced twice within 30 minutes is dropped; if you rejected a suggestion, the same topic does not return for 7 days; if the validator rejected one, it is not retried for 7 days either. Fourth, the quota: a suggestion of the same kind is given once a week. The only exception is severity: a new, more important suggestion deactivates a pending low-importance one and takes its place.

When you dismiss a suggestion with "I do not want this", you are not bothered about the same topic for a week; that is not a bug but a deliberate silence.

The system calls this the "conflict–dedupe–cooldown policy".

Say you see the "Your training coach — keep transitions short" line in the day detail every day, but the long coach note above the weekly plan sometimes does not appear at all.

There are two different coaches: a deterministic one that always speaks, and an AI one that speaks only when its conditions are met.

The short line in the day detail is a fixed cue derived from your goal and level — no internet needed, identical on every open. The long note on the weekly plan is produced by the AI, and if it cannot be produced the area is not drawn at all; no filler sentence is put in its place. Being unable to reach the AI is not treated as a failure: the base plan is already valid, only the commentary is missing.

If you do not see the long coach note, there is nothing wrong with your plan; no commentary was produced that week.

Say you log the third set of your squat as "40 kg × 10 reps", rest as 95 seconds and effort as RPE 8.

That single row is stored as a record and turns into four separate numbers at the end of the week.

The record carries eight fields: movement, set number, reps, weight, timestamp, effort (RPE), reps in reserve (RIR), actual rest duration and — for time-based movements — seconds. When the week ends these rows are aggregated per movement into: (1) completion rate — how many sets were planned and how many you actually did; (2) rep accuracy — how many of your sets stayed inside the rep band; (3) load retention — was your average weight at least 95% of the planned target; (4) rest overrun — did the average of your actual rests exceed 1.2× the planned rest. That last measure requires at least two rest records, because a single long break (the phone rang) is not a habit. A separate calculation keeps the movement’s best set of the week: the heaviest weight wins, and on equal weight the higher rep count. Rows with zero reps are not counted at all — a set that was not done does not count as done.

Entering weight and reps is mandatory; RPE and rest duration are optional, but when you enter them the engine goes from two measures to four, so next week’s decision is more accurate.

The system calls this the "performance signal" — raw records turned into a decision.

Say you finish a set and mark it "RPE 9.5"; where the screen would normally suggest 90 seconds, this time it shows 120.

The rest period lengthens according to how hard the last set was.

Starting again after the same rest following a very hard set lowers the next set’s reps and drags down the week’s completion rate. The rule is simple: if the effort was "very hard" 30 seconds are added to the base rest, if "hard" 15 seconds; otherwise the duration is unchanged.

Marking a hard set honestly does not slow you down; it raises the chance that you complete the next one.

Say this week you did all 12 of the 12 planned squat sets, your reps stayed inside the band and your weight did not drop. On bench press, meanwhile, you completed only 7 of 12 sets.

In next week’s plan the squat progresses while the bench press is held where it is — the decision is made per movement, not for the plan as a whole.

At the end of the week a performance signal is built for each movement and one of three outcomes is chosen: progress, hold, pull back. The order is fixed and safety comes first. If your average effort is above 9.5, it pulls back. If your completion rate is below 70%, it pulls back. If rests were overrun and completion is below 80%, it pulls back again. Progression comes only when all four conditions hold: completion at least 90%, rep accuracy at least 85%, load retained and rest not overrun. If none applies, the decision is "hold". On top of that, if movement-form analysis saw a risk, it overrides this order and closes progression — adding load while technique is broken produces injury. For female users the progression thresholds are slightly higher (93% and 90%); that is a deliberately more cautious setting.

If a movement sits at the same weight for weeks, the reason is not "the system forgot": that movement’s completion or rep accuracy is not clearing the threshold. Completing the sets you left unfinished gets you further faster than adding weight.

The system calls this the "adaptation decision".

Say loads and sets rose for three weeks, and in the fourth week the plan suddenly got lighter.

That is not a regression but the fourth step of the plan’s four-week rhythm.

Week types are read from a fixed cycle: week one is establishment, week two accumulation, week three peak, week four pull-back. The cycle is anchored to a fixed starting point in the calendar, so which week is which type can be computed independently of when you started. Without the pull-back week, week four’s load would land on an unrecovered body; fatigue would accumulate and week four would produce regression rather than progress. At advanced level the model differentiates further (block schemes such as accumulation → transmutation → realisation → taper come into play).

Adding extra sessions in the light week because "I did not train enough" defeats the purpose of the rhythm; that week the gain is in recovery, not in training.

The system calls this "periodization", and the light week a "deload".

Say the AI produced a suggestion to "increase shoulder volume" and you approved it. When you start today’s session you see the list has not changed.

An approved suggestion cannot change the week you are in; it only enters next week’s plan as a candidate.

This is not a limitation but a safety oath the system carries: the rules "the current week cannot be modified" and "movements cannot be swapped directly" sit in the code as hard `false` values and are not configurable. The reason is simple: changing the plan halfway through a week also breaks that week’s measurement — what you measured becomes unclear and the next decision loses its ground. On top of that, a suggestion cannot even become a candidate without a "passed" stamp from the validator.

When you approve a suggestion you see its effect on Monday; this week’s plan is deliberately left untouched.

The system calls this "next-week candidate only".

Say Tuesday was a training day, you never opened it, and today is Thursday.

Tuesday is marked "missed"; today and the days after it stay "planned".

A day counts as missed only if it is in the past and carries no record at all. A day you started but did not finish stays "in progress"; a day you finished becomes "completed". If there is a record on a rest day, that counts as completed too — stepping outside the plan is not punished, it is recorded. The distinction matters because the missed-day count is both one of the events that trigger the coach and a determinant of the week’s completion rate.

To make up a missed day, just tap it on the plan screen and open it; the system knows which day you are making up.