plot

name	plot
description	Aggregate and generate publication-quality figures from experiment results across all BFTS stages.

Plot Aggregation

You are generating publication-quality figures for an AI research paper.

Arguments

• --exp-dir <path>: Experiment directory (required)

Parse from the user's message.

Procedure

1. Load Experiment Context

Read the experiment state and gather all stage summaries:

python3 -c "
import json, sys, os
sys.path.insert(0, '.')
from tools.state_manager import load_experiment_state, load_journal, get_best_node, get_journal_summary
state = load_experiment_state('<exp_dir>')
for stage in ['stage1_initial', 'stage2_baseline', 'stage3_creative', 'stage4_ablation']:
    journal = load_journal('<exp_dir>', stage)
    summary = get_journal_summary(journal)
    best = get_best_node(journal)
    print(f'=== {stage} ===')
    print(json.dumps(summary, indent=2))
    if best and best.get('plot_paths'):
        print(f'Plots: {best[\"plot_paths\"]}')
"

2. Inventory Existing Plots

List all plots generated across all stages:

find <exp_dir>/workspace/figures -name "*.png" -o -name "*.pdf" 2>/dev/null | sort
find <exp_dir>/figures -name "*.png" -o -name "*.pdf" 2>/dev/null | sort

View each existing plot using the Read tool to understand what's available.

3. Read Best Experiment Code

Read the best experiment code from the final stage to understand what data is available:

ls <exp_dir>/state/stage4_ablation/best_solution_*.py 2>/dev/null || \
ls <exp_dir>/state/stage3_creative/best_solution_*.py 2>/dev/null || \
ls <exp_dir>/state/stage2_baseline/best_solution_*.py 2>/dev/null

4. Generate Aggregator Script

Create a Python script that produces publication-quality figures. The script should:

1. Load raw experiment data from .npy, .json, or .csv files in the workspace

2. Generate 4-6 key figures for the paper:

   • Training curves: Loss and accuracy over epochs (all datasets overlaid)
   • Method comparison: Bar chart comparing baseline vs. proposed method
   • Ablation results: Grouped bar chart showing component contributions
   • Dataset comparison: Performance across different datasets
   • Any additional figures that illustrate key findings

3. Follow publication standards:

   • Use matplotlib with seaborn style
   • Font size ≥ 10pt for readability
   • Clear legends and axis labels
   • Color-blind friendly palettes
   • Tight layout with no clipping
   • Save as both PNG (300 DPI) and PDF

cat > <exp_dir>/auto_plot_aggregator.py << 'PYTHON_EOF'
<generated aggregator script>
PYTHON_EOF

cd <exp_dir> && python auto_plot_aggregator.py

5. Review Generated Figures

View each generated figure using the Read tool. Evaluate:

• Are axes properly labeled?
• Are legends clear and complete?
• Do the figures tell a coherent story?
• Are there any visual artifacts or issues?

6. Iterate (up to 3 rounds)

If figures need improvement:

1. Identify specific issues
2. Edit the aggregator script
3. Re-run and re-check

7. Finalize

Copy final figures to the experiment's figures directory:

mkdir -p <exp_dir>/figures
cp <exp_dir>/workspace/figures/*.png <exp_dir>/figures/ 2>/dev/null
cp <exp_dir>/workspace/figures/*.pdf <exp_dir>/figures/ 2>/dev/null

List the final figures:

ls -la <exp_dir>/figures/

Figure Guidelines

• Minimum 4 figures for a complete paper
• Each figure should have a clear caption-worthy takeaway
• Use consistent styling across all figures
• Prefer vector formats (PDF) for line plots
• Use raster (PNG 300DPI) for complex plots with many points
• Ensure all data is real (from experiment runs), never synthetic/fake