FEXLINK AIMicrolink Research Institute · Industry knowledge display
You are here: Home / Guardrails & safety boundary · FEXLINK AI

Guardrails & safety boundary

In the display phase the platform takes a conservative safety stance: it refuses rather than emit unverified or high-risk conclusions.

Display phase Professional Q&A is in preparation

This platform currently provides knowledge and citation display only — no professional answers, selection or compliance conclusions.

1. Safety guardrails

GuardrailStatusDescription
Insufficient-evidence refusalEnabledEVIDENCE_INSUFFICIENCY_REFUSAL: explicitly refuses when evidence is insufficient.
High-risk force blockEnabledHIGH_RISK_FORCE_BLOCK: all high-risk questions are blocked and handed off.
Engineering design output disabledDisabledENGINEERING_DESIGN_OUTPUT_DISABLED: no engineering design output.
Final selection output disabledDisabledFINAL_SELECTION_OUTPUT_DISABLED: no final selection conclusion.
Compliance commitment disabledDisabledCOMPLIANCE_COMMITMENT_DISABLED: no compliance commitment.

2. Fail-closed behaviour

Professional Q&A is off during the display phase. Any call to /api/chat always returns:

{ "ok": false, "error": "PROFESSIONAL_ANSWERING_DISABLED", "citations": [], "model": { "connected": false, "professional_answering_enabled": false } }

The system never emits a professional answer; it fails closed rather than degrading into guesswork.

3. High risk & human handoff

  • Questions on high-voltage work, on-site engineering parameters or licensed judgement are always blocked.
  • Complex site conditions are referred to a human expert.
  • The platform does not replace a licensed professional.

4. No fabrication, extractive only

  • Answers come only from verified sources; no invented standard numbers or parameters.
  • Invalid citations and fabricated standards count zero.
  • UNVERIFIED content is never stated as confirmed fact.

5. Continue

See About for the capability roadmap and Citations & traceability for the citation mechanism.