README.md

metadata

title: Terminal-Bench 2.1, annotated
emoji: 🔎
colorFrom: yellow
colorTo: gray
sdk: static
pinned: false
short_description: The TB2.1 leaderboard, trial by trial, with atif-scan review

Terminal-Bench 2.1, annotated

The public Terminal-Bench 2.1 main leaderboard, re-read trial by trial with atif-scan. For every row:

• Highlights — masked excerpts of where findings and benchmark-awareness stages happened;
• Trajectories — atif-scan's static viewer of every finding-related step;
• a task × attempt strip, review leads beside the leaderboard's own reward-hack rate, and a checks × rows matrix linking straight to the excerpts;
• data holes & accounting: how much reasoning each row's traces hold (full, summary, hidden, unrecorded), compacted history, web results missing from the trace, unpaired tool results, and per-trial costs against the leaderboard's total.

Findings are review leads, not verdicts. Leaderboard scores, reward-hack rates and costs are the leaderboard's own. Trace text is atif-scan's credential-masked export of public leaderboard trials.