A library gives you functions. OreML gives you the judgement calls that decide which function to use, when, and why — the part that normally takes years of experience to get right, made explicit and repeatable.
Which correlations are real?
A binned dependency coefficient that reads numeric-to-categorical pairs a linear correlation simply cannot — with the trivial self-diagonal masked out of every ranking, so a column's perfect correlation with itself never tops the list.
Does this actually cause that?
The effect is identified through a back-door, instrumental or front-door route, estimated by several independent estimators, then attacked by five separate refutations — placebo, random common cause, subset, bootstrap, unobserved-confounder sensitivity — and graded on what survives.
What counts as an outlier here?
Not a textbook constant. The fence scales from 1.5× to 3.0× the interquartile range with the measured skew, and elongates on the heavy tail only — so a genuinely skewed column isn't punished for its shape.
Which column is worth predicting?
Candidates are ranked on completeness, class balance, cross-validated label consistency and the causal weight carried forward — then an admissibility veto rejects identifier-shaped columns outright, however well they score. An identifier is never an outcome.
How should a gap be filled?
The missingness mechanism is classified before anything is filled, and the strategy is chosen per column — a measured dependency on other columns routes to neighbour-based or iterative imputation; otherwise a normality test decides mean against median.
Is this row impossible, or just unusual?
Around twenty-two detectors score every row. Their scores are rank-normalised and averaged within six families — density, isolation depth, reconstruction error, probabilistic, histogram, rank — so six similar detectors can't outvote one well-calibrated one.
Did that new feature leak the answer?
Engineered features are measured against the target, one by one. Anything built in a process that could see the target is held to a stricter threshold than a feature that never saw it — leakage is detected, not merely avoided.
Did any of it actually help?
Feature sets and projections compete under paired cross-validation against the untouched original — and if nothing beats it by a statistically significant margin, the original wins. Balancing faces the same test from an independent judge, and when it made things worse the report says so in as many words.
Twenty-five steps, twenty-five methodologies. Not one of them is a default setting.
