# NLP and text-quality workflows

Run `python nlp_cases.py` from the repository root. The lab creates only original synthetic text. Names, order IDs, emails, phone numbers, tickets and reviews are invented for testing; they are not learner, customer or business records.

The verifier passes twenty checks covering split-safe TF-IDF, word/character tokenization, stemming and lemmatization collisions, duplicate screens, sentiment label contracts and hard slices, entity spans, labelled similarity pairs, retrieval, topic inspection status, multilingual slices, truncation, validation-selected incremental training, an explicit unexecuted-transformer record, structured extraction, per-field metrics, review themes, drift, privacy redaction and annotation agreement. Full outputs are saved in `nlp-verification.json`.

These compact cases are teaching references rather than production NLP benchmarks. Human topic and annotation review cannot be authenticated by code and remains explicitly pending where applicable.