One published result, with the command that produces it and the full log. If a task fails because our machine fell over, we rerun it. If it fails because Kernl got it wrong, it counts as a failure.
In plain terms: handed a public set of real computer jobs, Kernl finished about half of them correctly, unattended, with AI that runs on one desktop machine and costs nothing per task. Everything below is so you can check that claim yourself.
All 89 tasks in the public Terminal-Bench 2.1 set, each one a real job at a command line. Kernl got one attempt at each, and the benchmark's own checker decided whether the result was correct. We did not write anything task-specific or change the test rig.
harbor run -d terminal-bench/terminal-bench-2-1 -a kernl_harbor_agent:KernlAgent -k 1The run above used an open model on one desktop graphics card, so it costs nothing per task beyond the electricity. A model that size does not match the large commercial models from Anthropic, OpenAI, or Google on hard problems. What the number tracks is our own software: we hold the model and the test suite still, change the engine, and re-run. Movement is the engine.
Terminal-Bench and the test rig it runs on are public projects. The only part we wrote is the small piece that lets the rig drive Kernl, and it is in a public repository with the complete log of the run above, exactly as it came out. Results from a different rig or a different set of tasks are not comparable to this one. If you run the command and get a different answer, open an issue and we will look at it.