Harness² (google-research/harness2) — vendor claims and pricing, checked_on 2026-10-28 (UTC+8)

Sources:
- README (repo HEAD 1cfe254e8d25b15f37048c3730134686abe60335, 2026-10-13): https://github.com/google-research/harness2
- Project site: https://harness2.github.io
- Paper: https://arxiv.org/abs/2610.14224
- GitHub API 2026-10-28: 20,547 stars, 2,013 forks, created 2026-10-06, last push 2026-10-14, Apache-2.0, 137 open issues.
- Package: harness2 0.1.0 (pyproject); NOT on PyPI (pypi.org/pypi/harness2 -> 404); install from source.

Pricing: Free, Apache-2.0 open source. No account for the local demo. Model-backed benchmark runs need your own model access
("Model-backed runs incur API costs") — documented path is Vertex AI (gcloud ADC); JobBench judge can use any
OpenAI-compatible endpoint via HARNESS2_JUDGE_BASE_URL/HARNESS2_JUDGE_API_KEY. "This is not an officially supported Google product."

Claims (verbatim):
C1 README Install: "uv venv --python 3.12 .venv / source .venv/bin/activate / uv pip install -e '.[render]' / harness2-demo --output runs/demo --k 2"
   "The local demo runs both recursion modes with deterministic editors and real Python executions. It requires no model credentials."
C2 demo.py docstring: "A deterministic local adapter for exercising the real adaptation workflow."
   README: "Task-level recursion refines the execution harness for the current task through a contrastive Propose–Probe–Compose step ...
   without ground-truth feedback." "Harnessing-level recursion updates a persistent improvement harness with adopted harnesses and
   evidence-supported lessons from each task to guide future refinement."
C3 SETUP.md: "harness2-check checks installed dependencies, credentials, and configuration without calling a model."
   README: "Add --dry-run to harness2 to preview task selection without model calls."
C4 README: "Across three professional benchmarks and four model–harness configurations, a single Harness² step improves GLM-5 and
   Gemini 3.5 Flash over their base harnesses, with gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and
   11.8 percentage points on LAB's partial pass rate."
   Site: "Gemini 3.5 Flash with OpenCode improves from 72.4 to 77.5. Parallel recursion at k = 5 reaches 80.8." (WorkBuddy average, means of 3 runs)

Safety notes from vendor: "Executor network access is unrestricted, and Codex runs with its sandbox disabled."