Data preprocessing

itis_sumo.preprocess -- an optional layer that sits before training and after prediction.

DataPreprocessor (data_preprocessor.py)

Handles, in one reusable object:

• Variable mapping -- standardized names (x1, x2, ... for inputs; y1, y2, ... for outputs), so downstream Dakota confs never see raw, possibly-unsafe user variable names.
• Normalization / denormalization -- min-max or z-score scaling, applied before training and inverted after prediction.
• Sign switching -- restorable sign flips (useful for maximize-vs-minimize framing without re-deriving a surrogate).
• Configuration persistence -- the applied transform is serialized to JSON alongside the run, so predictions can be inverse-transformed later by a different process/session without re-deriving the mapping.

required_completed_jobs (models.py)

Minimum completed-sample-count check for building a Dakota (surfpack) GP surrogate -- max(floor, len(input_vars) + 1), since Dakota aborts surrogate construction when given <= len(input_vars) training points.

Verified behavior

(Verification & Validation Category H) -- transform/inverse-transform round-trips to 1e-10 for z-score, min-max, and sign-switch, including on a 1000-row x 20-variable dataset; and normalization measurably improves surrogate accuracy on badly-scaled inputs (e.g. f(x) = 1000x + 1) where an unnormalized fit degrades.