Skip to content

P00: The Harness Mystery

What You Do

A colleague wired up a new model and ran the discover-prove-fix benchmark. It failed: no PoC, no patch, nothing saved. But the model's response contains a correct solution. Read the trace and figure out what went wrong.

Harness Mechanism

Native tool calling, reasoning retention, and explicit sampling/output-token parameter forwarding. The harness, not the model, determines whether the model can act.

Open First

Why This Comes First

Every later project teaches you to read a trace and tune a lever. This one motivates the whole course as a debugging exercise: a bad harness can make a strong model look broken, and you find it by reading the trace, not by guessing.

Two real-world cases hit the same class of bug:

  • A CyberGym-E2E run where fixing three harness settings raised S1 from 40.7% to 70.0%: same model, same tasks.
  • An OpenAI experiment where enabling two API settings tripled their ARC-AGI-3 score and cut output tokens 6×.

The lesson: a benchmark score measures the whole stack, not just the model.

Built as a friendly front door for the runnable OpenHands harness lab.