ClawForge - State Repair: leaderboard

ClawForge-Bench (2026) gives a command-line agent 362 generated workflow tasks over seeded calendars, inboxes, task boards, files and configuration, where pre-existing state may be partial, stale or duplicated, and grades the final state with executable validators. This board is the strict full-pass accuracy on the 92 tasks whose primary ability is state repair, repairing or replacing stale and incorrect state; percent, higher is better.

Metric: Strict full-pass accuracy (%) on the 92 tasks whose primary ability is state repair (repairing or replacing stale and incorrect state), on the ClawForge-Bench snapshot (17 generator-backed scenarios, 362 executable state-conflict workflow tasks over tasks, calendar, email, messaging, files and runtime configuration), command-line agent in multi mode with full interaction history and a 25-step budget, scored by result-first validators over the final normalized state; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 7 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1O333.7#237
2Claude Opus 4.629.3#113
3GPT-5.226.1#185
4Kimi K2.518.5#266
5Claude Sonnet 4.617.4#173
6GPT-5.3 Codex13#124
7GPT-5.412#135

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=clawforge-state-repair · How It Works · Data refreshed daily, snapshot 2026-11-21.