Goose Swarm 3.0.2 — SB7.1 benchmark

Goose Swarm 3.0.2 brings SB7.1 into the desktop Benchmark view with a bundled scorer, guided installation of the optional benchmark tools, and recorded browser evidence.

Benchmark and scoring

• Run the payments task with a configured single model or Swarm. The benchmark retains backend consistency and recovery checks and adds a detailed, data-driven 3D inspector and committed-update animation.
• See the earned behavioral score, the final score, and the specific quality conditions that limit it. A high score measures the published contract; it is not a claim of subjective artistic excellence.
• Inspect the individual checks, screenshots and a short recording from the graded browser, then publish the result to LeanZero from the app.
• Scoring uses the bundled benchmark files. The release includes a file-hash manifest and validation provenance so that results can be tied to their actual scorer.

Verified installed-app run

Gemini 3.8 Flash completed the model build in 10m 16s. The bundled scorer graded the saved build in 14m 13s and awarded 49.92%, with no admission ceiling reducing the earned score. The app published the report and graded animation clip successfully. Inspect the full result and recording.

Desktop fixes

• Benchmark activity now shows meaningful model actions with expandable excerpts and raw details. Long paths wrap inside the panel.
• Completed builds receive a private receipt before scoring. If scoring is interrupted, eligible sessions can retry scoring without invoking the model again. Older sessions without that receipt cannot be recovered through this button.
• Corrected Python 3.12 scorer shutdown, approval and saved-state observation, paginated payment verification, and asynchronous recovery measurements. Premature payment-walk termination is scored from recorded vendor evidence instead of being mistaken for missing benchmark infrastructure. Existing scoring thresholds remain unchanged.
• Swarm terminology now describes mixed local and cloud nodes.

Installation and updates

Download the Apple Silicon macOS DMG below and drag Goose.app into Applications. Save active work before replacing or restarting the app. Existing settings and session data are stored outside the application bundle.

Compatible stable releases are also available through Settings → App → Updates. The release ZIP and latest-mac.yml assets serve that updater.

Benchmark runtime setup currently targets Apple Silicon macOS. A configured local model endpoint is supported separately from automatically attaching managed MLX models to a Swarm; this release does not claim every provider or mixed-node configuration was exercised by the Gemini acceptance run.

Benchmark prompt, methodology and results · Goose Swarm features and installation manual