Compliance & Data Privacy
Before deploying to a regulated client, an architect must know exactly which features are Zero Data Retention eligible, how to prevent PHI leakage through schema caching, and how US-inference pricing affects committed capacity.
1. ZDR Eligibility Matrix
Zero Data Retention means Anthropic does not store inputs or outputs beyond serving the request. Know this table for the exam.
| Feature | ZDR Eligible? | Reason if Not |
|---|---|---|
| Adaptive Thinking | Yes | N/A |
| Effort parameter | Yes | N/A |
| Citations | Yes | N/A |
| Structured Outputs | Yes | N/A |
| Web Search / Web Fetch (standard) | Yes | N/A |
| Compaction | Yes | N/A |
| Bash, Computer Use, Text Editor | Yes | N/A |
| Message Batches API | No | Inputs/outputs stored server-side up to 29 days |
| Files API | No | Files stored server-side |
| Code Execution | No | Execution state held server-side |
| Agent Skills | No | Skill data follows standard retention |
| MCP Connector | No | Requires server-side routing |
| Web Search/Fetch + Dynamic Filtering | No | Code execution required; stores intermediate state |
2. HIPAA Safeguards: Schema Caching Risk
For healthcare-related marketing, PHI must never appear in JSON schema definitions.
- Why: Structured output schemas (tool definitions,
output_config) are compiled and cached separately from message content, they may be treated differently under your BAA. - Rule: Field names in schemas should be generic (
"patient_id"→"record_id"). Actual PHI values belong in message content only, never in schema property names or descriptions.
3. US Inference Pricing Multiplier
Setting inference_geo: "us" restricts compute to US infrastructure and incurs a 1.1x pricing multiplier on all token categories on every Claude 4.6-or-later model, including all of the 5.5 models.
- Priority Tier impact: Each token processed with
inference_geo: "us"draws down 1.1 tokens from your committed TPM capacity, the multiplier applies to Priority Tier burndown, not just standard billing. (Opus 5.5 and Sonnet 5.5 are not offered on Priority Tier.) - ZDR compatibility: Data Residency (inference_geo) is ZDR-eligible. You can combine US-only inference with ZDR for the highest data protection posture.
The 1.1x multiplier applies to every model that accepts inference_geo (Claude 4.6 and later), Haiku 5.5 and Sonnet 5.5 included; older models such as Haiku 4.5 reject the parameter with a 400. On Haiku 5.5 it also applies on top of the over-100K-token rate. When a client asks for US-only inference on a cost-sensitive workload, the multiplier is the same everywhere, so the saving comes from routing: send high-volume steps to Haiku 5.5 and reserve Opus 5.5 for strategic reasoning.
Lab Exercise: Compliance & Data Privacy Review
Self-driven lab Module9_Self_Driven_Lab.ipynbObjective: evaluate an agent architecture for ZDR, regional inference, and regulated-data risks.
- Create a feature matrix that marks each component ZDR-eligible or not.
- Add `inference_geo: "us"` where appropriate and calculate the policy/cost implication.
- Review schema caching or prompt caching choices for regulated data exposure.
- Define a human escalation trigger for privacy-sensitive or policy-uncertain actions.
A compliance review memo for the advanced agent architecture.