This page explains what data SisuLift reads, which rules it uses to propose weights and deload weeks, which parts are exact calculation and which parts use a language model, and what the AI never does. It is written from the code, not from marketing material, and we update it when the logic changes.
Weekly progression weights and reps are calculated in the database with deterministic rules. The same data always yields the same suggestion. A language model does not decide your weights.
A commercial language model is used for conversation, written feedback and program drafts. When it proposes numbers, they pass through guards before they reach your program.
Every suggestion is accepted or skipped. The AI does not change your program, weights or goals on its own, and it makes no medical decisions.
Every decision is based on your own data. You are never compared to other users.
Weight, reps, set type and RPE (rate of perceived exertion). Warm-ups, drop sets, cluster sets and technique sets are filtered out of progression decisions so they do not distort how working sets are read.
The rep range of an exercise is read from your program, not from a stale target in the log. If the program has been regenerated, the old target does not drive the decision.
The post-workout rating (light, moderate, heavy), completion percentage and the reason if a session was cut short. A session cut short for lack of time is not a fatigue signal.
Sleep, stress and energy are asked at the start of a workout. They produce a readiness score from 0 to 100 that the AI coach also sees.
Sleep with stages, HRV and resting heart rate from Apple Health or Health Connect. Values are compared only to your own 14-day average, never to a general norm. Without a wearable, the calculation works from logged data.
Recovery of each muscle group is calculated from the sets you did, weighted by RPE, and last week's load is compared to your four-week average. The load ratio is a descriptive metric, not an injury risk prediction.
Goal, experience level, age, training style (cautious, guided, all out), gym equipment and per-equipment weight increments. Menstrual cycle phase is considered only if you have opted into cycle tracking.
Every Monday the previous week is analysed one exercise at a time and you are shown suggestions to accept or skip. This is pure calculation in the database. The rules below are listed in the order they take effect.
SisuLift uses double progression: you collect reps at the same weight until every working set hits the top of the rep range. Only then does the weight go up and reps return to the bottom of the range. Weight increases when it is earned, not by the calendar.
If the perceived exertion of your working sets exceeded the threshold for your goal, weight is not increased even if the reps were hit. Reps are added first instead. The training style setting adjusts the threshold: cautious brakes earliest, all out allows higher exertion.
An increase is rounded to the smallest real step of the equipment (for example barbell 2.5 kg, dumbbell 2 kg, machine 5 kg), and you can define your own steps. If the step would be too large relative to the weight, reps are collected first. On light loads a small absolute step is still allowed so progress does not stall.
If the same weight has been used for several weeks and the bottom of the rep range is met, an increase is proposed as a test. If reps have been raised for several consecutive weeks at the same weight, the weight is forced up. A forced increase never exceeds a safe step.
If reps run clearly past the target, or the target is already at the ceiling for your goal, the weight goes up and reps reset. The size of the increase scales with how far past the target you went, but a single jump is capped. Isolation exercises have a higher ceiling than compound lifts.
When every set has hit the top of the program's range for several weeks in a row and the muscle group's weekly volume is still below its tolerated limit, one extra set is proposed.
Deload and intensification weights are normalised before analysis. A decrease is never proposed from deloaded weights, and a runaway increase is never proposed from intensification weights. After a break, the weight is restored close to the previous working weight, not from zero.
The outcome of every accepted suggestion is measured the following week: hit, partial or missed. The success rate, last week's readiness and RPE adjust the size of the next increase. The multiplier is bounded, so one good or bad week does not swing the step unreasonably.
Accept or skip: suggestions are shown on Monday, and you can accept all of them, some, or none. An accepted change is applied immediately to this week's and next week's workouts. A skipped suggestion affects nothing, and the reason for every suggestion is always shown as text.
Training moves through three phases: accumulation, intensification and a deload week. Transitions are mostly decided by rules, and the mandatory safety cap is always a rule.
In accumulation, volume and weights are normal. In intensification, weights go up by about five percent and the top of the rep range is lowered. In the deload week, volume drops to about half and weights by about 15 percent. The deload comes primarily from volume, because that restores you best without losing strength.
Intensification starts once accumulation has lasted enough weeks, or earlier if perceived exertion has clearly risen across several logged sessions or progression success has dropped. The transition is proposed, not done silently.
Three levels in order: 1) a mandatory safety cap on the length of intensification that depends on age (longer for younger lifters, shorter over 50), 2) a data-based recommendation when RPE rises and readiness falls, completion stays below 80 percent for three weeks, or readiness drops more than 15 percent from your own level, 3) a calendar-based precautionary recommendation. Strongly green recovery can postpone the safety cap by at most two weeks.
The language model makes a separate periodisation assessment from the same data and can recommend a deload before a rule fires. A fresh assessment can override the calendar-based recommendation, but it never overrides the mandatory safety cap.
You can start a deload week yourself with one tap at any time, and a proposed deload can be skipped. During a deload week, weekly progression produces no suggestions, because deloaded data would distort them.
Known weakness: the weekly analysis's own deload recommendation currently fires too easily when several rated sessions are marked heavy on a small sample. The fix is on the list. We say this because transparency also means saying unfinished things out loud.
Readiness and recovery are pure calculation. They never call a language model.
The score combines sleep, HRV, resting heart rate, per-muscle recovery, load and your own feeling. The weights scale with which signals are available: without a wearable, sleep and feeling weigh more. A missing signal is left out, not guessed.
HRV and resting heart rate are related only to your own 14-day average. That is why different watches measuring differently does not matter, and you are never compared to anyone else.
If today's readiness is low or a muscle group is still recovering, the app proposes a lightened session or a substitute exercise. You choose the level of lightening yourself, and a manual choice always wins over an automatic one.
SisuLift has not trained its own training model, and we do not claim to. We use a commercial language model through an API. It receives your data and the rules it must speak by, and it does not decide your weights. Here is every place it is used.
Conversation in Finnish and English. Grind fetches numbers with tools straight from the database: training history, strength progression, recovery, volume levels, sleep, body metrics and the food log. Charts are drawn from database numbers, not from the model's memory. The only writing action is logging a meal at your request. Grind does not change your program.
The written feedback comes from the language model, but the weight target for the next session comes from the same deterministic calculation as weekly progression. The model may not invent numbers, and if its text and the calculation disagree, the calculation wins.
A written summary of the week with recommendations. The model also returns a fine-tuning factor that can shrink or grow the size of the next increase. The factor works within the bounds of the progression rules and cannot set a weight directly.
The program draft (split, exercises, sets, rep ranges) comes from the language model. Starting weights are given to the model as hints from your history or a strength-ratio estimate, and the model's proposal passes three guards: it may not go below your proven level, may not go below an empty bar on barbell lifts, and a single-session increase may not exceed 15 percent of your last one. The draft is accepted or rejected.
Meal photo and nutrition label recognition, plateau analysis, periodisation assessment and generating exercise instructions. Here too the model's output is a suggestion or text, never an automatic change.
The AI provider does not use your data to train models, and we have a data processing agreement with them. Details are in the privacy policy. Privacy policy →
It makes no medical decisions, diagnoses or rehabilitation programs. If you have an injury or illness, the app points you to a professional.
It forces nothing. Every suggestion can be skipped, and a program can be built and edited entirely by hand without the AI.
It does not change your program, weights or goals without your approval. The only exception is the mandatory deload week when the safety cap is reached, and even that is shown and explained.
It does not compare you to other users. Every baseline comes from your own history.
It does not invent numbers. Every number the AI shows or refers to comes from the database or from deterministic calculation.
It does not hide its reasons. Every progression suggestion has a written justification, and every deload recommendation has a trigger.
No. Conversation and written text use a commercial language model through an API, but weights, reps, mesocycle transitions and readiness are calculated with deterministic rules in the database. The language model receives those numbers ready-made and does not decide them.
No, and we do not claim to. The progression logic is rule-based and built on evidence-based principles: double progression, RPE-based autoregulation, volume landmarks and periodised deloads. The rules are updated based on user data, and every change is run against the whole user base's history before release.
Yes. Readiness is calculated from logged sleep, stress and energy plus training load. A wearable adds HRV and measured sleep, which changes the weighting, but no feature requires a wearable.
You can. Weekly progression suggestions are accepted one by one, weights can be overridden in a workout, a deload can be skipped or started yourself, and a program can be built entirely by hand. The AI features are an addition, not a requirement.
The skeleton is the same, but the thresholds are not. Your goal (strength, hypertrophy, fat loss) sets the RPE thresholds and rep ceilings, training style adjusts the brake, age sets the safety cap, equipment sets the step size, and the outcome of accepted suggestions sets the size of the increase. Isolation exercises also have their own rep limits.
Skip it and tell us in the support chat which exercise and which suggestion. Every suggestion is stored with its reasons, so we can check why it was made. Most rule updates have started from exactly these messages.
Try it yourself
Try Premium for 14 days without a card. The first weekly progression appears after your first full training week.