Hold the numbers. The numbers do not hold still.
Named slots hold numbers. You see them once. Instructions then rewrite them one at a time, and you keep up in your head. At the end you answer questions with nothing on screen.
The slots land somewhere new every round, never in a row. A row can be held as a list and the names read back off the order at the end. Scattered, B has to be remembered as B.
A wrong answer or a timeout costs one strike, however many questions you missed that round. Three strikes end the run. Failing does not send you backwards: the level runs again with new instructions.
A level takes three clears to pass. The marks beside the level number show how many you have banked. A strike does not spend them.
Level 1 is one instruction over two slots, A and B, and so is level 2. Level 3 adds a third slot and nothing else. Level 4 adds the second instruction, level 6 the third. Nothing ever gains a slot and an instruction on the same level. New vocabulary arrives one idea at a time, and a new idea appears at most once per round while it is still new.
8 12 3 becomes 4 9 3.Your score is the highest level you passed. Ties break on strikes remaining, then on speed. Answering inside par scores full speed. Past par it decays, and well past it the question closes.
No level here was authored. Each one is generated, simulated, and kept only if its predicted difficulty lands within 10% of that level's target.
You are modelled as a store plus a processor. The store holds items: slot values, half-finished arithmetic, the question you are holding. Mistakes happen only at accesses, the moments you reach into the store or compute something. Each access carries a failure probability that rises with three things:
Every instruction and question compiles to a trace of primitive events.
The trace is priced, and the round's total is one number: strain.
Your predicted chance of getting the round right is
e−strain.
A step is priced by what it asks for, not only by how big the numbers are. A multiplication costs more than an addition. Working out which slot is the largest costs more than reading one you already know.
Because this is falsifiable. The model claims a scan late in a round costs more than the same scan early, and that rotating the slots makes every value read as stale. Those are predictions, and the game records enough to check them. If the numbers disagree with players, the numbers are wrong.
A run ends on three strikes, so levels reached swings on luck. This asks a steadier question: how hard a round can you hold?
Play a few runs and this fills in. It needs about a dozen rounds to say anything useful. Practice and replayed seeds do not count.
The game predicts your odds before you play each round. Where the lines split is where it has you wrong.
The score itself, luck and all.
Nothing here changes reading time, processing time, gaps, or par. Difficulty is not configurable.
Unscored. Holds one level for as long as you want, and can slow the tempo.
Unscored. Strikes still end the session, but nothing is recorded.
Practice stays on the level you pick. Clearing a round draws another at the same level.