No rule survives contact
with individual physiology.

It tells being weaker apart from being fatigued

Your true strength is never observed directly — what you see in a set is capacity minus whatever fatigue you brought to it. Platebook estimates the two separately, running a Kalman filter over a hierarchical state: a shared per-muscle factor plus a near-static per-exercise offset, updated by every set you log and read through the fatigue you were carrying at the time.

Because the state is shared, a hard session of rows sharpens the estimate for pulldowns too. Because it drifts between sessions, a lift you haven't trained in months widens its own error bars instead of pretending the last data point still holds.

In plain terms A bad session and getting weaker look identical in a training log. Platebook can tell them apart — and tracking that gap over time is what lets it see a deload coming before you dig yourself into a hole.

Fatigue is two pools per muscle, and it feeds back into your strength

An acute pool saturates within a session and clears fast. A slow damage pool accrues through the session, commits at the end with super-linear compounding — residual fatigue amplifies fresh damage — then decays on a time constant set by that muscle's tissue class.

This isn't a soreness score in a corner of the app. It's the same state that suppresses your strength estimate, gates when a muscle may be trained again, and prices every candidate session the solver considers.

In plain terms Some fatigue is gone by tomorrow; some takes days and stacks up if you keep training through it. Training hard on a muscle that hasn't recovered costs you more than the identical session would have cost you fresh — so the plan doesn't schedule it that way.

Stimulus is the currency — in its own units, not "hard sets"

Each set earns a growth stimulus that saturates across the working rep range and rises sharply as the set approaches failure, normalized against a reference set. Weekly totals accumulate in those native units. Muscles that aren't the target earn graded credit, scaled by how close that tissue actually works to failure inside the movement.

In plain terms Three sets taken to the edge aren't the same as three comfortable ones, and counting "sets per week" pretends they are. Platebook counts what each set is actually worth — and gives your triceps proper credit for the pressing they already did, instead of a round number somebody guessed at.

Your volume ceiling is computed, not looked up in a table

Published landmarks give population ranges. Platebook computes a conditional ceiling per muscle: the largest weekly dose whose best possible spacing still clears the recovery gate, given your training days, your intensity, and how much of that muscle's work arrives indirectly from compounds. A boundary of feasibility, never a target.

In plain terms There's a point where extra sets stop buying growth and just cost you recovery. That point isn't the same for you on four days as it is for someone training six — so it gets worked out for your schedule instead of read off a chart.

The week is solved in stages, each with its own constraints

A target vector is scaled at the sub-muscle level until it meets those ceilings. Muscle rhythms are enumerated against a running spillover ledger, so what compounds already delivered is subtracted before any direct work is assigned. Assembly then enforces a rule most programs violate blind: a downstream muscle may not sit in the recovery shadow of an upstream compound day that recruits it.

The daily solve enumerates feasible integer-RIR paths across a rest grid per movement class and works backward from the terminal set, so sets, reps, RIR and rest come out of one joint decision. Packing places it all under a strict-priority objective in which your session-length window is a tolerance — and inside it, physiological coherence outranks making every day the same length.

In plain terms It doesn't fill slots in a template. It works out what each muscle needs, subtracts what your big lifts already gave it, refuses to schedule anything that's still wrecked from two days ago — including your triceps before a bench day — then decides sets, reps and rest together rather than one at a time.

Heavy work carries a systemic cost no muscle pool can see

Heavy triples aren't just local fatigue — there's a central and bracing cost no per-muscle pool captures. Platebook carries a separate systemic pool, so the layout question (which day each heavy lift lands on) is scored on realized systemic load across the whole week.

In plain terms Three heavy days stacked together will bury you even when every individual muscle says it's ready. The plan knows that and spreads them out — because it loses on the objective, not because someone had an opinion.

Exercise choice is scored on three axes; your gym is a filter

Candidates compete on lengthened-tension (stretch and range, gated by whether the movement genuinely holds tension at length), loadability resolved against your real equipment — including whether a bodyweight movement can be loaded at all and how finely you can add weight — and practicality carrying a stability tax. Emphasis re-weights the axes; implement preference composes on top as a separate dimension.

In plain terms It picks the movement that grows the target muscle best with the equipment you actually have. Gear you don't own never scores badly — it simply isn't considered, because an exercise you can't do isn't a worse option, it's not an option.

Variety means movement identity, not a different implement

Platebook tests genuine movement identity, so rotation buys new stress patterns rather than the appearance of change — within the week, and slowly across months at rates matched to each lift's class.

In plain terms Swapping a barbell for dumbbells isn't variety, it's the same movement in different clothes. Big lifts hold for months; smaller work refreshes sooner.

One generator, parameterized by your state

Adaptive mode isn't a second algorithm bolted on. There is one generator, and a fresh program is the zero-state case. A deviation changes the state; the same solver re-phases what's ahead from there. Deviate from nothing and it reproduces your cycle exactly — behavior continuous in your data, not a jump between "the plan" and "the improvised plan."

In plain terms Miss a week and it re-plans around what actually happened instead of pretending you didn't. Follow it exactly and nothing moves. Short on time today, it recomposes the session two ways and lets you choose; want a different exercise mid-workout, it ranks the alternatives your gym allows and shows what the swap does to your week before you accept it.

It can see the models, not just the output

The coach can query the engine directly — your program's full analysis, per-muscle stimulus, what's fed by compounds, how the week is laid out, the trade-offs the solver made, your strength estimates, your training history. The stimulus model in particular is deliberately exposed to it: functional form, constants and every computed intermediate, because an explanation grounded in the actual math beats a plausible-sounding one.

In plain terms Ask "why is my chest only twice a week" or "why this exercise and not that one" and you get the real reason the program was built that way — not a generic answer about hypertrophy.

It can change things, not just discuss them

Read and write, across the app: log sets mid-session, swap an exercise for a ranked alternative, add or remove work, restructure a program, build a one-off session, import training you did elsewhere, and edit your own settings — priority muscles, injuries, equipment and training environments, warm-up protocols, units, dashboard, theme.

In plain terms "Swap the leg press for hack squats, my knee's cranky" or "I did this workout at a hotel gym yesterday, log it" — and it does it, rather than telling you which screen to go find.

Every change is logged and reversible

Write access is only acceptable with an audit trail. Each mutation snapshots what it touched before it writes, so the full change log is inspectable and any edit can be reverted — by you, or by telling the coach to undo it.

In plain terms If it changes something you didn't want, say "undo that." Nothing it does is a one-way door.

It's a production frontier model, not a hobby LLM

No home-grown model, no thin wrapper around a fine-tune. Platebook runs Anthropic's Claude, with the model selectable in Settings, given a purpose-built tool surface and the engine's own state as its context.

In plain terms The reasoning quality is the same as the assistants you already use — except this one can see your training data and the model that built your program.

Entirely optional

The coach is a switch in Settings. Off, every AI surface disappears and the engine still builds, adapts and explains your programs — the physiology and the solver are local, and none of the programming depends on the model.

In plain terms Don't want AI anywhere near your training? Leave it off. You lose the conversation, not the program.