Malformed JSON Moderation API Responses in Node.js: Schema Gates Over Adapters

Malformed JSON responses from moderation API calls in Node.js require a robust handling strategy. The first step is to separate transport success from semantic success, as a 200 response can still contain invalid JSON. Parsing only checks for syntax errors, not semantic correctness or policy compliance. For a ticket like "The blender arrived with a cracked jar," the application needs a specific schema: a decision, a reason, and possibly a confidence value.

If the provider returns malformed JSON, the application should not silently classify the message as safe. Instead, log a redacted record, retry only when the operation is safe to repeat, and route the ticket to an alternate moderation path or human review. This evaluation constraint ensures that recovery preserves the triage contract when the provider response is unreliable.

While provider-native structured-output features like OpenAI Structured Outputs, Anthropic tool use, or Gemini response schemas can improve response quality, they introduce portability issues. Each provider has unique request fields, refusal behaviors, limits, and error vocabularies. An adapter can normalize these calls, but cannot make their guarantees identical. Therefore, a schema gate provides a stable boundary after provider-specific code has finished.

The gate should perform three checks: syntax, schema, and policy. JSON parsing is the bouncer that checks syntax. If parsing fails, the object is rejected. If parsing succeeds, the schema and policy checks are performed to ensure the object conforms to the expected shape. If the object is missing required fields or violates policy, it is rejected with an appropriate error.

When handling malformed JSON, separate transport success from semantic success. A 200 response can contain truncated JSON, a fenced object, extra content, or valid objects with the wrong fields. JSON.parse alone cannot guarantee safety. Implement try/catch blocks to parse within safe boundaries.

If the JSON is malformed, reject it immediately. If it parses successfully but fails schema validation, the object is still rejected. If it passes schema validation but violates policy, the object is rejected with a policy failure. Each failure type represents a distinct operational event, so collapsing them into one catch-all error makes debugging more difficult and obscures which fallback is being triggered.

Measure the failures to distinguish between syntax failures (the bytes cannot be parsed), schema failures (the object parses but lacks required fields), and policy failures (the object is technically valid but violates policy rules). Collapsing these into a single "AI error" makes debugging harder and hides which fallback is being used.

To optimize costs and reliability, read the account budget before sending a batch of tickets. Use the budget result to decide whether the batch is allowed, and pass the approved batch into the provider adapter. The first capability's output (budget status) becomes an input to the second-stage call (sending the batch). This sharing of credentials and base URL simplifies operations.

For moderation calls, use a single credential and base URL to simplify operations. The budget read is performed before sending a batch, and the moderation call is injected where there is no dedicated moderation route. This maintains honesty while demonstrating the control boundary a worker should implement.

In summary, handle malformed JSON responses with a schema gate in Node.js. Separate transport success from semantic success, perform syntax, schema, and policy checks, and reject unusable results. Use provider-native structured output for quality improvements while maintaining a stable business workflow boundary. Treat each failure type as a distinct operational event, measure failures to guide debugging, and share credentials and base URLs to optimize costs and reliability.