# Scientific and technical limitations

- **Synthetic** examples are not evidence of generalization to authentic academic abstracts, policy texts or student writing.
- Typo, synonym and appended-context edits are deterministic heuristics; semantic label invariance was **not** verified by human annotators.
- A spell-change model flip can be informative, but may not constitute an actual error in naturally occurring writing.
- Destructive truncation can change a statement's rhetorical function; no accuracy is reported for it.
- Dictionary substitution has low row-change coverage on some splits; a deceptively high subset score is not an overall performance improvement.
- Conditional subset scores compare on matched changed IDs. They cannot be substituted for full-sample accuracy.
- The fixed examples are small, particularly cross-language subgroups; bootstrap resampling omits training and annotation uncertainty.
- Naive Bayes posteriors are not proven calibrated. Similarity margins are **not probabilities**.
- Research release gates are arbitrary adjustable checklist thresholds, NOT safety, compliance, fairness or deployment certification.
- No real-time monitoring, real users, neural models, learned adversarial generator or external validation is implemented.
- Models are transparent pedagogical lexical baselines; they are not SOTA, and no production deployment is recommended.