roundlisted

# /tml:round — design the next experiment Turn a question into a study whose answer will actually mean something. The whole value is in two things: **one goal**, and an honest **scientific / nuisance / fixed** split. Everything else is bookkeeping. ## First action, always ```bash ls -la ./.tml/rounds/ 2>/dev/null | tail -5; cat ./.tml/frontier.md 2>/dev/null | head -40 ``` Number this round after the highest existing one. If earlier rounds exist, read the most recent spec and its verdict — a round that repeats a settled question is wasted budget, and a round that ignores the previous round's caveats inherits them silently. ## Hard rules 1. **One goal per round.** If the goal needs an "and", it is two rounds. Two simultaneous questions cannot be disentangled afterwards. 2. **Write the role assignment down.** A round whose scientific / nuisance / fixed split was never recorded cannot be checked for fairness later — and it will be, by `/tml:analyze`, possibly in a different session. 3. **Every fixed hyperparameter is a caveat on the conclusion.** Record it as one, in those words. 4. **Never put `max_train_steps` in the search space.** Fixed per study. 5. **Read-first.** Writes only under `./.tml/rounds/NNN/`. Never runs training. ## Procedure ### 1. Establish the regime — before designing anything `references/regime.md`. Three questions: how many trials can run concurrently, where do they run, and does this plugin get to see the results directly. The answers