atai-omega-encoder-agent-data-prep
Installation
SKILL.md
Omega Encoder Agent Data Prep — Recordings → Role Files
The Omega Encoder agent embeds whatever numeric columns it is given, a window at a time, and returns one 768-dimensional embedding per channel per window. It does not score, label or check anything. So the shape of what goes in decides whether what comes out can be used: a label column is embedded as a sensor, a segment that is not a whole number of windows loses its tail (or, in serving, fails), and two files whose time ranges overlap can't be told apart in a batch session's output.
This skill turns recordings into role files that avoid all of that:
index.csv ─► prepare.py ─► prepared/ ─► build_roles.py ─► roles/ ─► preflight_roles.py
(segments, one cut per exact grid, z-score (library), library/ validation/ every rule,
roles, segment, fixed rows one file per segment, test/ delivery/ PASS / FAIL
groups, aligned on labels held back, labels/ manifest.json
labels) an event distinct times