Benchmark data: Pixelbolt with and without

The files behind Does a pixel-art MCP server help Claude? (run 23 October 2026). Published copies: local paths and internal names are removed, replaced in [square brackets], or renamed to BENCH_* and bench-* in the scripts.

• PLAN.md: the plan, written before the runs, with its one dated amendment
• tasks.json: the 12 requests (and the pilot task, excluded from results)
• results.json: the 72 main jobs: every check, file facts, cost, time, tools called (pixel_compile_formats, pixel_render_gif_args and delivered_gif_is_pixelbolt_render were added afterwards from the job logs and folders; Bash commands with paths are shown as [compound command])
• results-told.json: the 36 exploratory "told" jobs
• judgements.json and judgements-told.json: every blind verdict, both orders, with the judge's reason
• summary.json: the per-model totals and the claim-rule outcome
• searches-2026-10-23.txt: the five web searches and what ranked
• pixelbolt-0.9.0-sizes.txt: Pixelbolt 0.9.0's tool list and help sizes
• frozen-hashes.txt: SHA-256 of the plan and scripts at the freeze, before the runs
• outputs/: every requested file each job delivered, by model and job (outputs/<model>/<task>-<arm>/)
• Scripts: run_bench.py, score.py, score_told.py, sheets.py, judge.py, summary.py, figures.py, measure_sizes.py
• Figures: opus-T11.png, opus-T02.png

MIT-licensed, like Pixelbolt.