---

name: data-evaluation description: "Use OpenCDA's dumped YAML vehicle records for offline trajectory augmentation, evaluation-output interpretation, plotting, and non-destructive debugging." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Data evaluation

Route offline work on OpenCDA data dumps, trajectory labels, evaluation outputs, and debug plots here. This sub-skill is applicable when the input is a directory of frame YAML files; it does not start a simulator or alter the source tree.

• Read data-formats.md before interpreting vehicle records, trajectory tuples, or observation/prediction horizons.
• Read evaluation-and-plotting.md for the evaluation lifecycle and saved debug artifacts.
• Read troubleshooting.md when keys, horizons, output paths, or plotting backends fail.
• For safe augmentation, run scripts/generate_prediction_yaml.py with explicit --input-root and --output-root. It writes YAML only to a separate output tree by default; use --in-place only when source mutation is explicitly approved.

Boundaries and external limits

This skill covers deterministic, offline post-processing and inspection. It requires Python with PyYAML for the bundled helper and optionally matplotlib for plots. EvaluationManager also imports CARLA and consumes live/in-memory vehicle-manager state; a CARLA 0.9.12 client import was checked, but no CARLA server was verified. SUMO, ScenarioRunner, torch, and YOLOv5 runtimes are not part of the verified environment. Do not infer a successful simulator, co-simulation, detector, or model evaluation from an offline YAML check.