Publish five-trial Daytona article-suite leaderboard

[bingran-you]

Summary

• run all four canonical models on all nine article-derived tasks in Daytona
• increase every task's agent timeout to 3x its original limit
• publish five independent trials per model-task pair (180 accepted cells)
• compute the final score as RLiable-style IQM over the pooled 5 x 9 normalized score matrix
• publish mean and sample standard deviation for native task scores and per-trial IQMs
• add 180 trial-level score files and 180 matching provenance metadata files
• make Daytona the authoritative default sandbox and record exact provider-specific inference settings
• update the README, detailed leaderboard report, scoring docs, website JSON, and native HTML/CSS plots
• add responsive uncertainty whiskers and fix clipped leaderboard values on desktop and mobile

Final pooled IQM

Rank	Model	IQM	Trial-IQM SD	Positive display index
1	Claude Opus 4.8	62.03	10.38	162.03
2	GPT-5.6 Sol	36.65	5.52	136.65
3	GPT-5.5	10.66	18.68	110.66
4	GPT-5.4 Mini	-17.41	5.90	82.59

Validation

• uv run python scripts/validate_tasks.py
• uv run bench tasks check tasks/simulation_heuristics_ant_v1 --level publication-grade
• uv run python scripts/audit_simulation_heuristics_ant_v1_leaderboard.py
• uv run ruff check .
• uv run pytest -q (106 passed, 18 skipped)
• uv build
• CI-equivalent Docker oracle smoke (1/1, 100%)
• localhost HTML audit at desktop and 390px mobile width: zero clipped values or uncertainty ranges

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
genesis-bench	Ready	Preview, Comment	Jul 15, 2026 1:52pm