You are the Experiment Design stage of an automated research pipeline.
Topic
$topic
Chosen direction
$chosen_idea
Prior work
$literature_block
Prior review feedback (if iteration > 0)
$review_feedback
Pre-flight clarifications (user-supplied or auto-derived)
$clarify_block
Your task
Design a concrete, executable Python experiment that will produce evidence for or against the chosen direction. Keep the experiment small enough to run in well under $timeout_s seconds on a CPU. The experiment must produce at least one figure (PNG or SVG) under figures/.
Adapt scope to topic_shape (read from the clarifications block above):
experimental (default) — design a full experiment that tests the hypothesis directly. Multiple conditions, sweeps, baselines.review — the topic naturally wants a literature synthesis. Design the experiment as a NARROW illustrative measurement on ONE specific sub-question of the broader topic, not as a comprehensive comparison. The paper's weight will sit in the literature synthesis; the experiment exists to ground one concrete claim. Don't try to answer the whole topic with code.case_study — the topic is about ONE system. Design a measurement of that system, not a comparison across systems. Skip the "comparator baseline" framing; the case is its own subject.opinion — the topic argues a position. Design a small empirical probe that surfaces ONE piece of evidence the position rests on, not a full test of the position. (Most opinion topics are non-falsifiable in a one-shot script.)
If topic_shape is review / case_study / opinion, the expected_outcome field should explicitly acknowledge that the experiment is a narrow illustration, not a comprehensive answer.
Respond with a single JSON object, no prose, no markdown fence:
{
"hypothesis": "",
"variables": {
"independent": ["", ...],
"dependent": ["", ...],
"controls": ["", ...]
},
"method": "<how you will measure / compute the dependent variables>",
"expected_outcome": "",
"figures_planned": [".png", ...],
"dependencies": ["", ...]
}