# Research model card · Observatory of baseline classifiers

## Models
- Multinomial Naive Bayes (token-count-based deterministic baseline)
- Word TF-IDF centroid
- Character TF-IDF centroid
- Hybrid Word/Character TF-IDF centroid

These four research baselines are trained by `source_modelops_engine.js` on 842 synthetic-curated sentences. The current dashboard reads their **precomputed predictions** and does not train models when the page loads. The supplied regeneration script performs the same deterministic training in Node.js.

## Intended use
Educational demonstrations of model evaluation, performance shifts, error analysis, reproducible benchmarking and model-governance documentation. Suitable as a portfolio research tool.

## Not intended for
Real-world document assessment, personnel or education decisions, deployment certification, bias/fairness conclusions, real-time monitoring, or unverified inferential claims about unseen scientific corpora.

## Test results
See `README.md` and `evaluation_snapshots.json`. Report results by split and sample size. Controlled IID metrics should not be substituted for challenge performance.

## Model limitations
Synthetic-curated data, template structure, possible annotation shortcuts, 54-case challenge size, narrow task design, and no systematic external domain transfer validation within this project. Probability calibration is **not** established. The dashboard shows accuracy and macro-F1, not calibrated confidence.