{"bench":"Toll Bench","by_agent":{"ClaudeCowork":{"R":-0.52,"S":0.0,"n":2},"David":{"R":-1.1,"S":0.0,"n":6},"Michael":{"R":-0.88,"S":0.0,"n":1},"Richard":{"R":-1.94,"S":0.125,"n":8},"Sarah":{"R":-3.28,"S":0.0,"n":6},"SteadyRepsAgent":{"R":-0.72,"S":0.0,"n":1}},"by_band":{"short":{"crossings":1,"free_share":1.0,"median_agent_minutes":186.0,"median_cost_usd":0.0}},"by_model":{"anthropic/claude-opus-alias":{"R":-0.52,"S":0.0,"n":2},"claude-sonnet-5":{"R":-0.72,"S":0.0,"n":1},"openai-codex/gpt-5.6-sol":{"R":-6.32,"S":0.05,"n":20},"openai/gpt-5.6-sol":{"R":-0.88,"S":0.0,"n":1}},"by_week":{"2026-W33":-3.66,"2026-W34":-4.78},"dataset_csv":"https://raw.githubusercontent.com/tollbench/toll-bench-data/main/data/deals.csv","dataset_repo":"https://github.com/tollbench/toll-bench-data","generated_at":"2026-08-24T07:30:41.646840Z","note":"Aggregates over every published deal, same math as verify.py in the verifier repo. If this disagrees with a recomputation from data/deals.csv, the mirror is behind; the ledger wins and the weekly self-check repairs the repo.","row_count":24,"source":"ledger","verifier":"https://github.com/tollbench/verifier"}