Knobs: simpler rules, stop early, more variety

1. Simpler rules — fewer branches, lower “tree depth,” fewer fancy features. Less room to memorize weird exceptions.
2. Stop early — pause training before practice score becomes unrealistically perfect while the surprise quiz stalls.
3. More varied practice examples — different writers, lighting, gyms, spam phrasings so quirks of one batch matter less.
4. Keep-it-simple pressure (regularization) — gently punish overly wiggly explanations.
5. Honest splits — practice / validation / final test stay separate so you do not “study the answer key” by accident.

Each knob trades something: simpler rules can miss real patterns (underfitting). Stopping early can leave useful learning unfinished. More data helps only if it looks like the real quiz conditions.

Flexible run: practice 99%, validation 84% (gap 15). After simpler rules / earlier stop: 93% / 91% (gap 2). You gave up some practice bragging rights to keep surprise-quiz strength.

Draw a panel with three sliders labeled Simpler, Stop early, More variety. For a spam-filter club project, write which slider you would move first if practice is 100% and a friend’s examples score 70%. Explain why you would not move all sliders randomly at once — change one, measure, then decide.

• Prefer a measured baseline over a flashy demo.
• Keep versions of your notebook so you can undo a bad knob twist.
• Separate “code ran” from “new-quiz evidence looks honest.”

Simpler rules: You cap how detailed the explanation may be. Practice may fall a little; surprise quizzes often rise if you were memorizing noise.

Stop early: You watch both curves. When practice keeps climbing and validation stalls, you pause — like ending a study session when you start memorizing page numbers instead of ideas.

More variety: You add examples from different friends, lighting, or phish phrasings. Variety is not “more of the exact same worksheet copied ten times.”

Write a warning on your control panel: knobs change tradeoffs; they do not grant unlimited trust. After each twist, recompute train, val, and gap, and keep the old row so you can roll back. If you cannot explain a knob to a 10-year-old, you do not understand it well enough to turn it on demo day.