AI Engineering from Scratch（英文原版）
Given a training dataset used by a deployment, check compliance against California AB 2013 and EU TDM opt-out.
Produce:
- AB 2013 coverage. Fill the 12 fields. Flag any missing or placeholder-only fields. Note that the summary becomes binding once published.
- Opt-out compliance. Does the dataset respect machine-readable opt-out signals (robots.txt, C2PA "No AI Training", TDM.Reservation)? Pre-collection filter must be in place.
- DPA jurisdiction mapping. For each jurisdiction the data subjects belong to, identify the applicable DPA and the 2025 legitimate-interest position (Irish DPC, Cologne Higher Regional Court, Hamburg DPA, UK ICO, Brazilian ANPD).
- Irreversibility audit. If the dataset contains PII, what unlearning or remediation procedure is in place? Acknowledge that no procedure fully remediates training data.
- Provenance-chain completeness. Is there a signed chain from the data source to the training pipeline? If the dataset is derived (crawled + filtered), document the derivation.
Hard rejects:
- Any deployment that cites AB 2013 without per-dataset 12-field summaries.
- Any deployment that does not respect robots.txt or equivalent opt-out signals.
- Any remediation claim that assumes surgical removal of data from trained weights.
Refusal rules:
- If the user asks whether a specific dataset is "safe to train on," refuse without jurisdiction-by-jurisdiction analysis.
- If the user asks for a universal compliance strategy, refuse — jurisdictions differ materially.
Output: a one-page check filling the five sections, identifying the highest-risk compliance gap, and naming the single most urgent remediation. Cite California AB 2013 and EU Copyright Directive TDM exception once each.