DATA_PROVENANCE.md

Data provenance and lineage

• Source project: ModelOps Research Lab v2.0 by Rodrigo Pantoja.
• Source dataset: Academic Rhetorical Intent Mini v1.1, synthetic-curated, CC BY 4.0.
• Task: English, Spanish and Portuguese academic rhetorical-intent classification into six classes: background, objective, method, result, limitation, implication.
• Original split sizes: 842 train; 130 controlled IID test; 54 novel-pattern challenge; 1,026 records in total.
• Monitoring artifact: evaluation_snapshots.json with 736 rows = 4 model methods × (130 + 54).
• Input artifacts: train.json, iid_test.json, challenge.json, and source_modelops_engine.js carried forward unchanged from the earlier ModelOps work.
• Rebuild command: node build_artifacts.js; no external data or network required.
• Integrity checks: unique split/model/id keys, source-split ID separation, consistent sample IDs for each method; SHA-256 digests in benchmark_manifest.json.

Do not describe the two splits as monthly checkpoints or claim to be streaming production logs. The samples are deliberately small and synthetic-curated. Any external dataset must supply its own provenance, annotation quality assessment and usage permissions.