Right to Explanation API

Implements a structured API contract, typically RESTful or gRPC, to programmatically handle data subject access requests (DSARs). The interface accepts a subject identifier and a decision timestamp as input, returning a machine-readable payload containing the logic of the automated decision. This standardizes the fulfillment of GDPR Article 22 and Article 15(1)(h) obligations, moving away from manual, ad-hoc PDF generation to an auditable, automated workflow. The response schema must be deterministic and versioned to ensure consistency across legal reviews.

Goes beyond static feature importance by dynamically computing the minimal set of changes required to flip a specific automated outcome. For a denied loan application, the API doesn't just state 'income was a factor'; it calculates and returns: 'If your annual income were $5,000 higher, the decision would have been approved.' This technique, rooted in algorithmic recourse, provides actionable, human-readable insights. The API must execute these what-if simulations against a sandboxed model instance to avoid polluting production metrics while generating the explanation.

Allows the requestor to filter the breadth of the explanation to prevent information overload or exposure of proprietary model internals. The API must support query parameters to toggle between:

• Global vs. Local Explanations: The model's general logic versus the factors for a single prediction.
• Feature Ranking: Returning only the top-k most influential features.
• Attribution Method: Specifying the algorithm used, such as SHAP or LIME, to ensure the explanation meets the specific rigor required by a legal auditor or data scientist. This granularity balances the legal 'right to meaningful information' with the enterprise need to protect trade secrets.

Every API call—both the request for an explanation and the generated response—must be cryptographically logged to a WORM-compliant (Write Once, Read Many) storage system. The API attaches a cryptographic non-repudiation token and a secure timestamp (RFC 3161) to the response payload. This proves not only what the explanation was, but exactly when it was generated and that it hasn't been altered post-hoc. This integration closes the accountability loop, transforming the API from a simple query tool into a defensible legal artifact for regulatory investigations.

Recognizes that 'meaningful information' varies by user literacy. The API must be content-negotiable to serve explanations in multiple formats from a single endpoint:

• Structured JSON: For integration with downstream compliance dashboards and automated auditors.
• Natural Language Text: A plain-language summary generated by an LLM, translating complex feature weights into readable sentences.
• Visualization Data: Base64-encoded images or raw data points for rendering waterfall charts, force plots, or decision trees directly in a privacy portal. The Accept header drives the serialization logic, ensuring the API serves both machines and humans effectively.

Executes explanation queries without exposing raw training data or the complete model state to the requestor. The API architecture must incorporate a differential privacy budget check before releasing any statistical aggregation. If generating a counterfactual requires referencing a protected cohort (e.g., 'people like you'), the API injects calibrated noise or uses k-anonymity thresholds to prevent membership inference attacks. This ensures compliance with data minimization principles while still fulfilling the right to explanation, preventing the API from becoming a vector for model inversion.