Tune the machine you are the optimizer

The regression unit's sleep-and-scores data is back — on purpose. This time you are the machine: turn a knob, watch one error number, make it smaller. Then unlock a second knob and feel why training is genuinely hard — and then let the machine finish with your own strategy, done consistently.

ink dots: the same six students from the regression page · gold line: your machine · thin gray drops: each miss (largest drawn thick)

your descent trace — gray: your knob turns · gold: the machine's steps · log scale so the tail stays visible

Phase 1

slope knob — m

0 … 20 · arrows turn it (Shift = coarse)

start-level knob — b

locked until Phase 2

Advance the game

phases advance by button, never on their own

Regression continuity

the faint squares' areas are exactly what the error adds up

The machine's whole rule, in words: measure which way each knob pushes the error down, nudge it that way a little, repeat. One honest detail: it measures the level knob at the middle of the data (6.75 h) so the two knobs don't fight each other, then reports b = level − m · 6.75. Same knobs, same error, steadier walk.

Try: with one knob you can't get under 94.3 — why does unlocking b change that?