When you ask Betterance for a plan, no AI writes your numbers. Every distance, weight, and calorie comes from published sports science, turned into code, checked by machines before every release — and then it adapts to what you actually do, explaining every change it makes.
Most AI fitness apps ask a language model to “write a plan,” and the model improvises numbers that look right. Betterance never does that. The plan is assembled by deterministic builders — ordinary code that gives the same answer every time — and every constant inside them traces to a named, published source: the NHS Couch-to-5K program, Hal Higdon’s marathon plans, ACSM and WHO position stands, the Owen equations for metabolism, peer-reviewed meta-analyses for protein and training volume.
All of it lives in one rulebook (SCIENCE.md),
where every rule carries an honesty tag: established (real evidence),
convention (what good coaches do, said plainly), or
contested (deliberately not built — things like “the 10% rule
prevents injury,” which the research doesn’t actually support). The rulebook was
built by researchers whose numbers were then attacked by independent fact-checkers;
they killed, among other things, a 16-week marathon plan that traced back to a
competitor’s marketing PDF rather than to any coaching canon.
Your goal in plain words decides which parts the plan needs — “Run a marathon” is a running plan, “Lose weight” is meals plus training, “Sleep better” is habits. The About You screen then asks only the questions your goal’s science needs: a runner states how far they can comfortably run; a gym goal asks about equipment; everyone states bodyweight, sex, and health conditions.
Your comfortable distance seeds the whole week: easy runs at 70% of it, one long run at 110%, spread across non-consecutive days with recovery walks between. Every session carries a ceiling set by the goal itself — a marathon plan trains toward a 32 km peak (the race is run on race day, not in training), a distance-less goal caps at 25 km. No paces, ever: with no race times or heart-rate data, honest intensity is “you should be able to speak a full sentence.”
Sessions are composed from a 111-movement catalog, filtered by your equipment, experience, and injuries. Weekly training volume is counted per muscle and kept inside the evidence range (roughly 10–20 hard sets, never more than 20), sessions cap at four days a week, and every session opens with a warm-up and closes with clear stop-if-something-feels-wrong rules.
Calories start from a published metabolism equation (Owen) times a deliberately modest activity level, plus the exact cost of the training this same plan prescribes — so a marathon plan feeds the running it demands. A weight-loss goal subtracts the guideline 500 kcal, with floors that never break. Meals are then chosen from an 86-recipe catalog with known macros, portion-scaled to hit the day’s targets, filtered for your diet and allergies.
Before any builder runs, your conditions are screened the way a clinician would triage them. A heart condition or diabetes keeps all training at an easy, conversational effort and feeds you at maintenance instead of a deficit — cutting calories in that situation is a decision for you and a doctor, not an app. Pregnancy freezes running targets in place. Medications matter too: appetite-reducing drugs like Ozempic bar the deficit and raise the protein priority; certain antibiotics freeze training progression while you take them (they carry a tendon warning). And a condition the app doesn’t recognize is never silently ignored — the plan says so, in writing, on the sessions.
The same honesty applies to ambition. A marathon from a 10 km base honestly takes about eighteen weeks — the app tells you so, but it never compresses the plan to flatter a deadline. Refusal is an output: the ramp is sacred, and squeezing the plan would mean squeezing you.
After the build, exactly one thing changes your plan: your own recorded data. Three loops run continuously, and they all follow the same consent rule — the plan moves up when you’ve proven it, and moves down only when you tap yes.
A finish-line goal becomes a path of stones, each one a real capability your recorded data can bank. A marathon path climbs through 12, 15, 18 km, the half, 25, 28 — summiting at the same 32 km training peak the engine caps promotion at, then an “ease off” stone (the taper, demonstrated by two genuinely shorter runs, not scheduled by a calendar), then: Ready for race day. There are no dates anywhere. Race-readiness is climbed, not declared.
Every stone shows its receipts — tap one and see exactly which run or weigh-in earned it, or exactly what it still needs. Earned stones are permanent: on the weight ladder you’re judged by your lowest recorded weight, so an ordinary water-weight bounce can never take a stone back.
Every number that moved writes a line saying what it did and why. Once a week, when there’s something real to say, a card presents it all: calories recalibrated and the trend arithmetic behind them, run targets grown, weights stepped up, suggestions waiting, stones banked. A plan that changes silently reads as a bug. One that explains itself reads as a coach.
Every rule in the rulebook has an executable twin — a machine check that runs the entire engine before any version ships: 4,320 gym configurations, 2,016 meal plans, 4,312 running weeks, plus twenty-week forward simulations of both the running and lifting loops proving that targets stay capped, earned, and never walk down on their own. Around 1.5 million assertions in all; a single red check blocks the release. These checks catch real mistakes — during development they stopped, among others, a plank that would have progressed one second per week and an energy floor that quietly neutered heavy users’ deficits.
And beyond the machines: eleven full test plans — marathons with injuries, pregnancy, heart conditions, medications; weight loss clean and gated — were built end-to-end and independently audited against their configuration. Eleven of eleven matched, down to the calorie.