Newton Data Prep — Clean → Split → Featurize

A pre-modeling pipeline for time-series sensor data. Three composable building blocks that take a raw multivariate dataframe and hand back the (X, y, metadata) arrays a downstream KNN / Isolation Forest classifier expects.

Origin. The three vendored scripts originated from work by Lucas (Solutions Engineering) and have been used end-to-end in real prep pipelines. The repo copy is the source of truth — ping Lucas if you need to pull in a newer revision.

When to Apply

• User has raw multivariate sensor CSVs (1+ sensors, irregular timestamps, NaN gaps) and is about to embed them with atai-newton-omega-model, and needs to clean the data first.
• User asks "should I drop rows with NaNs?" / "how do I handle gaps?" / "what's a good train/test split for time-series?"
• User's classifier is suspiciously good or suspiciously bad and you suspect temporal leakage — DataSplitter(mode='oot') is the fix.
• User has per-sensor embeddings (from the Omega model) and needs to fold them into a single feature matrix for KNN — that's FeaturePreparer.
• User wants the "joint state" pattern described in atai-newton-omega-model in code form.

Use the external omega-1-4-preflight checks instead when: you want a read-only go/no-go gate before committing to a run. Preflight makes no changes to the data. This skill makes changes — block-splitting, imputation, dimensionality reduction. The two are complementary: preflight tells you whether the dataset is salvageable; this skill cleans it up.

Do not use this skill when:

• The task is video, image, or text (use atai-newton-fusion-model).
• Data is already pristine (regular sampling, no NaNs, no leakage risk) — the pipeline becomes a no-op and you can hand the dataframe directly to the downstream skill.