Prompt Guardrails¶
Statement¶
ClaimGuard has no external generative-AI integration in current production scope, so prompt guardrails — the controls that constrain what a generative model can be asked to do and how it can respond — are not applicable to the current production configuration.
This page is the placeholder so a reviewer can see the topic is named and not overlooked. If an external generative-AI integration is introduced in the future, the guardrail design described below becomes the governing policy.
Implementation (future)¶
If an external generative-AI integration is introduced into production, the following guardrails will apply:
- Strict-scope prompt. The model is constrained to opine only on tool-grounded signals; out-of-scope topics (legal liability, coverage decisions, criminal intent, customer character) are explicitly forbidden.
- Structured-output contract. The model returns parseable JSON per a defined schema; downstream code does not consume free-form narrative.
- Data minimisation at prompt construction. The integration sends only the fields necessary for the analysis; raw uploads, credentials, PII fields beyond what's already in the claim, and unrelated claims are not sent.
- Human-in-the-loop. Every model output is presented to a human reviewer; no automated decision-making.
- Prompt-source review. Prompt files are versioned in source control and reviewed in the same PR process as application code.
- Allow-listed prompt paths. The integration loads prompts only from a fixed allow-list of source paths; arbitrary user-supplied prompt contents cannot reach the model.
Status¶
implemented — verified 2026-05-10. No external generative-AI integration in current production scope; the prompt-guardrail design is documented and ready to apply if such an integration is introduced.
Roadmap¶
- Prompt-injection regression set — to be authored if and when an external generative-AI integration is introduced.
- Held-out adversarial prompt evaluation — same.