# Fast-WAM Task 9

Independent GELLO specialist initialized from the pinned Fast-WAM Base LIBERO checkpoint. This package contains a LoRA adapter and trained pose10 input/output projections; the original base model and VAE are also required.

Task prompt: `pick up the purple object and put it in the black saucepan`.

Use this folder with the repository's `inference/serve_fastwam.py`, its own `normalization.json`, and its exact `prompt.txt`. Input is the current external RGB image and measured base-frame flange pose10. Output is 32 absolute next-achieved flange pose10 targets with commanded aperture, at 15 Hz target spacing. The server predicts actions without generating future video. Target spacing does not imply 15 inference requests per second.

`INFERENCE_REPORT.json` records all 40 fixed held-out windows across five episodes. The final adapter passed current-image latent parity, action normalization, finite pose10 output, orthonormal rotation, deterministic inference, HTTP query parity, and a native-versus-portable comparison on the same recorded input. These checks establish software compatibility in the recorded environment, not closed-loop physical task success.

See [INFERENCE.md](../INFERENCE.md) for installation and client usage, [FINETUNE.md](../FINETUNE.md) for the full adaptation workflow, and [GITHUB_CODE.md](../GITHUB_CODE.md) for the matching code. Keep this task's weights, normalization, and prompt embedding together.

The final held-out action flow loss is above the lowest intermediate value. This is the requested continued specialist, not a validation-selected best checkpoint. Its physical performance must be measured; additional training alone does not demonstrate an improvement.