Colab accelerator notebooks

The Colab notebook index links the completed TPU word-sorting example and tracks the remaining TPU/GPU notebooks awaiting completed execution.

Each notebook installs EvalRX and Antigravity CLI (agy), launches the coding agent in the runtime, runs evalrx run, and displays the baseline, agent analysis, repair selection, and independent confirmation. Installation creates a separate environment in /content/evalrx-env, with its own Python 3.12 and locked packages, so Colab image updates do not change it. Where GitHub or PyPI are unreachable, as in Google-internal Colab, the notebook can install from an offline bundle instead. For setup options, see the environment section of the notebook index. Save the executed notebook to keep its outputs, generated code, tables, and figures together.

Add GEMINI_API_KEY to Colab Secrets and enable notebook access. The notebook sets agy's modelProvider to gemini; the key alone is insufficient. No interactive login is required. See agy authentication.

The target model runs locally: jax_local on TPU, hf_local on GPU. Agent reasoning uses the external API. The example's output identifies the actual validation hardware. Inference repairs leave weights unchanged; an EXPLORE gain is not a confirmed repair.

The planned hardware pairs use Gemma 4 E2B-it with the same two datasets: BBH word sorting and CRUXEval output prediction. Each pair keeps the 128 examples, seed, EXPLORE/CONFIRM split, generation budget, and repair settings identical, so the notebooks demonstrate hardware adaptation of the same experiment.

See the CLI reference and JAX backend design for supported flags and capabilities.