# Study notes

This project explains the completed `codex-pilot20-v1` experiment. The experiment operator confirms that all 60 trials used GPT-6 Astra in xhigh mode. It is a small simulation study of an adaptive agent-and-tool system.

## Protocol

- Four purposively selected tasks in each of five RoboDojo capability dimensions.
- Native seeds 0, 1, and 2; the first public layout in each seed directory. The common layout index does not imply identical scenes across seeds.
- All 60 planned trials have native results. Failures were retained without replacement.
- Native scores and native full-success flags are distinct. Positive partial score is not full completion.
- The button-counting task had prior source inspection and development. The 19-task view removes all three of its trials.

## Control

GPT-6 Astra (xhigh), running through the historical Codex-in-the-Loop controller/bridge, receives standard RGB images, instructions, joint states, gripper states, and end-effector poses through an observation bridge. Public static calibration and an explicitly estimated contact-height plane can turn a selected image point into an approximate geometric target.

The agent selects move, grip, hold, and home commands. Numerical interpolation and native inverse kinematics translate targets into controls. The next observation supports corrections or a change of strategy. Feedback is at decision boundaries rather than continuously after every submitted vector.

An actual worked example is included in `data/worked-example.json`: the first ruler-lifting attempt failed; a later chunk used a closed gripper to push the ruler toward three blocks. The controller then requested return toward the initial public joint state. The native evaluator reported success.

## Information access and limitations

The bridge excludes object ground-truth poses, depth, segmentation, reference actions, future frames, and reward internals. The complete interactive session has not undergone an independent information-isolation audit. Earlier infrastructure output included a resource prim name reported as unused for control, and the button task had prior development. This is not a certified zero-leak or zero-shot study.

The model configuration for this campaign was GPT-6 Astra in xhigh mode, as confirmed by the experiment operator. That setting was not serialized into each individual trial artifact; the complete prompt history, token usage, and billing were also not archived. Concise visible decision rationales are not internal chain-of-thought traces. Robot state was recorded at decision points; the 35,646 submitted control vectors are not certified executed physics steps.

## Results

- All trials: 11/60 full successes (18.3333%), mean native score 24.4167/100.
- Excluding prior development: 8/57 (14.0351%), mean score 20.4386/100.
- Five tasks succeeded at least once; two succeeded on all three seeds.
- No selected Precision or Long-Horizon trial fully succeeded.
- The local archive records 318 decision chunks. This website publishes compact summaries and four representative clips rather than the large raw observation/control archive.

Published baseline means are recomputed for the same task names. Their episodes, layouts, adaptation, and execution budgets differ from this pilot. The frozen snapshot is dated September 5, 2026, downloaded September 8. No baseline policies were rerun and no official rank is claimed.

The experiment used simulated ARX X5 manipulation with Isaac Lab 2.3.2.post1 and Isaac Sim 5.1. It does not establish real-robot transfer or causal attribution between language reasoning, numerical tools, and accumulated context.