What is your config actually worth?
Run bench/run_bench.py on your repo — same tasks, same model, config stripped vs present — then load the results here. The judge is your own test command, never the agent's self-report.
Accepts results-summary.json or raw results.jsonl
How to produce results
python3 bench/run_bench.py \
--tasks bench/tasks.example.json \
--repo ~/code/your-project \
--agent-cmd 'claude -p "{prompt}" --dangerously-skip-permissions' \
--repeats 3 --max-turns 5 --out bench/results.jsonl
Turns to pass, per task
Agent invocations until the task's check command exits 0 (median of passing runs). Fewer turns = less re-discovery and rework.
Estimated tokens, per task
Median output size (chars ÷ 4) across runs — a cost proxy, not billing data.
Full numbers
The table is the source of truth — charts summarize it.