AGENTS.md

Agent instructions — diffmedia-loop

Work only in this repository. The cardiac and neural scores are cartoons on a fixed box (CHIR99021, IWP2, SB431542, LDN-193189). Expected improvement beating random on that cartoon is the result you can prove with no hood.

Do not

• Move the synthetic peaks and then claim the search “found biology.”
• Propose factor values outside the encoded published windows.
• Turn a proposal into a dose for a person or animal.
• Delete the repeat-EI policy that is supposed to show batch-uniqueness violations.
• Edit the candidate CSV from inside a reservation transaction.

First commands

python -m pip install -e .
python -m unittest discover -s tests -v
PYTHONPATH=src python3 -m medialoop.cli --objective cardiac --seeds 4
PYTHONPATH=src python3 -m medialoop.bakeoff --objective cardiac --seeds 2

Use fewer seeds than the README for a smoke run. Do not replace the README's 12-seed numbers unless you re-ran 12 seeds.

Improve, in this order

1. If you change the GP or acquisition, re-run the bake-off and report simple regret against the best grid point, plus uniqueness violations.
2. Allowed: one new policy behind the same columns (regret, violations). Do not remove EI, UCB, Thompson, or random.
3. A real assay hook must keep medialoop-plan from calling the synthetic observe. The planner prints a point; it does not plate a well.
4. Cross-check any suggested micromolar value against cell-protocol-compiler windows before writing it into a fixture. If they disagree, stop and report the disagreement.

Done when

Tests pass and the README still says the surface is synthetic.