program.md

Figure Autoresearch Program

Your job is to improve one figure: a single, wide visual for a Substack post that explains the worker-critic pattern in one glance.

Objective

Raise the score reported by prepare.py.

The current evaluation averages four 0-10 criteria:

1. semantic_fidelity
2. one_glance_clarity
3. readability_layout
4. visual_coherence

The score is computed by the script, not by you. A run is only considered accepted if:

• the average score is at least 8.5, and
• every individual criterion is at least 8.0.

Files

• plot.py is the only file you should modify.
• prepare.py is the fixed evaluation harness. Do not edit it.
• artifacts/autoresearch/current/figure.svg and figure.png are the outputs of plot.py.
• artifacts/autoresearch/current/review.json and review.md are the outputs of prepare.py.

Figure brief

The figure should make these ideas visible:

• a worker produces drafts;
• a critic reviews those drafts;
• feedback flows back into revision;
• the worker and critic are persistent roles, not one-shot calls;
• the loop stops only when the result is approved.

The ideal figure is:

• understandable in about five seconds;
• legible at publication scale;
• visually coherent, not cluttered;
• explicit about the direction of information flow.

Workflow

1. Edit plot.py.
2. Run uv run python plot.py.
3. Run uv run python prepare.py.
4. Read review.md and review.json.
5. Keep changes only if the score improves, or if the average ties but the weakest criterion improves.

Constraints

• Do not change the task.
• Do not change the scoring rubric.
• Do not add extra files unless they are outputs inside artifacts/autoresearch/current/.
• Prefer clearer structure over more decoration.
• Prefer fewer words over denser explanations.