# Agent Schematic Evaluations

The Codex 5.6 Sol agents ran against the Minecraft reconstruction task in each `task_environment/` and produced an `output.schem`. The output is compared with the ground-truth `answer.schem` by `score_schematics.py` using exact block-state IoU.

## Runs

- `codex-gpt5.6sol-high/`: started with the `/goal` command at high reasoning effort. Score: **20.32%**.
- `codex-gpt5.6sol-xhigh/`: run at xhigh reasoning effort. Score: **28.84%**.

Each completed run includes the trajectory evidence as a `rollout-*.jsonl` file, plus `trajectory_stats.md` and `score_report.md`. The candidate and ground-truth visualizations are provided as `output.mov` and `answer.mov`; the corresponding `.schem` files can also be viewed with [Schemat.io](https://schemat.io/view).

## Incomplete Run

A Fable 5 high run did not complete after more than two hours and remained stuck while reading the videos. Evidence for that run is not attached.