Model cards

Every optional FreshData model artifact ships with a model card covering: purpose, non-goals, runtime behavior, training data sources, license summary, PII policy, teacher-model usage, human-review process, evaluation metrics, known limitations, failure modes, the protected-column guarantee, calibration version, artifact hash, and FreshData version compatibility.

The authoritative cards live in the repository and are copied into each packaged artifact (dist/artifacts/<model_id>/model_card.md):

• fd-col-encoder-v1 — column/value encoder + semantic-type role head
• fd-intent-v1 — context-sentence intent head (optional evidence for the deterministic parser)
• calib-v1 — isotonic confidence-calibration tables (plain JSON)

Every card states explicitly:

• there is no generative repair model
• there is no runtime LLM
• there is no automatic model download during cleaning
• model output is evidence, not authority — the policy gate decides
• the protected-column guard remains absolute

Install status and hashes are always inspectable locally:

import freshdata as fd
fd.models.status()