IXIA Guardrails

Live on Cloudflare Workers AI · TechEx Amsterdam 2026

Every enquiry answeredin seconds.The risky oneswait for a lawyer.

IXIA writes the first reply to new clients of law firms and clinics. Guardrails checks each draft against the firm's own knowledge base, Llama Guard 3 and an independent verifier model. Safe drafts are sent at once. Sensitive ones go to a review queue, and the client gets an instant acknowledgement in their own language.

Write as a client of Demo Law Amsterdamfictional firm
The gate idle

    Review queue and audit chain

    Held drafts from this browser. Approve, edit or reject them the way a firm's intake lawyer would. Every decision is a record whose SHA-256 hash covers the previous one, so changing any past entry breaks verification.

    Waiting 0

    • Nothing waiting. Try the burn-out or asylum sample above.

    Audit chain 0

    1. No records yet.

    How the gate decides

    Five stages, three models, one deterministic rule set. Any block stops the draft; any review holds it; otherwise it is sent.

    1. 01
      Inbound rules

      Prompt-injection markers and sensitive personal data (health, criminal matters, migration status, children, violence or self-harm, religion, trade union) in six languages. The draft model flags the same categories as a second opinion.

      regex + draft model
    2. 02
      Retrieve

      Multilingual embeddings pick the four most relevant entries of the firm's knowledge base. A Dutch question finds the English fee entry.

      @cf/baai/bge-m3
    3. 03
      Draft

      A grounded first reply in the client's language, with citations to the knowledge entries it used.

      @cf/meta/llama-3.3-70b-instruct-fp8-fast
    4. 04
      Safety and verification, in parallel

      Llama Guard 3 classifies the reply (S6 specialised advice and S7 privacy hold it; other hazards block it). A second model from another family lists every claim the knowledge base does not support, case-specific advice, promised outcomes and echoed sensitive details.

      @cf/meta/llama-guard-3-8b · @cf/openai/gpt-oss-120b
    5. 05
      Deterministic outbound rules

      Every price or percentage must appear in the retrieved knowledge. Phone numbers, e-mails and links must be the firm's own. Citations must exist. The reply must be in the client's language. An AI-disclosure line is always appended.

      rules

    Measured, on 40 hand-written enquiries

    Loading the latest run…

    GroupCasesSent at onceHeld or blocked

    The cases were written by the team: 14 plain FAQs that must be answered at once, 18 sensitive or hostile enquiries that must never get an automatic answer, 8 advice-seeking or bait questions where either outcome is acceptable. They are a regression test, not a benchmark. Full per-case output is in the repository under eval/results.