The primary vulnerability of Large Language Models in the legal domain is citation hallucination: inventing plausible-sounding statute numbers, mixing up article provisions, or citing overruled precedent.

Dike eliminates this risk at the infrastructure layer through its Grounding Verification Engine.

1. Closed-World Citation Constraint

When you send a query to POST /v1/reason, the system constructs an ephemeral, closed-world context containing the top_k top-ranked legal chunks from the retrieval engine.

The LLM is prompted under strict syntactic constraints:

1. It may only state legal assertions that are directly supported by the provided context.
2. Every legal claim must be explicitly followed by a reference tag identifying the source: [[REF:canonical_id#article_no]].
3. It must not extrapolate, invent statute numbers, or cite external legal memory.

2. Real-Time Token Interception & Validation

During generation (both synchronous and SSE streaming), the Dike gateway monitors the token stream:

The Set Membership Check

Let C be the candidate set of valid (canonical_id, article_no) tuples retrieved for this query:

Valid(ref) = (ref.canonical_id, ref.article_no) ∈ C

• If Valid(ref) == true:
  • The token is verified.
  • The gateway assigns a sequential marker index (1, 2, 3...).
  • The raw tag [[REF:eg-leg-1948-131#163]] is rewritten to [1].
  • The full metadata and snippet of the source article are added to the citations array.
• If Valid(ref) == false (Hallucination):
  • The token is immediately stripped from the text output.
  • The hallucination is logged in internal metrics.
  • grounding.dropped_unverified_refs is incremented.

3. Grounding Telemetry in Responses

Every /v1/reason response returns precise grounding metadata:

{
  "answer": "يلتزم المستأجر بسداد الأجرة في المواعيد المتفق عليها [1]...",
  "citations": [
    {
      "marker": 1,
      "canonical_id": "eg-leg-1948-131",
      "article_no": "586",
      "citation_string": "المادة 586 من القانون المدني رقم 131 لسنة 1948",
      "snippet": "يجب على المستأجر أن يقوم بوفاء الأجرة في المواعيد المعينة...",
      "verified": true
    }
  ],
  "grounding": {
    "retrieved": 12,
    "used": 1,
    "dropped_unverified_refs": 0
  }
}

• grounding.retrieved: Total legal chunks supplied to the model.
• grounding.used: Count of verified citations cited in the final answer.
• grounding.dropped_unverified_refs: Number of fabricated citation markers intercepted and stripped by the hallucination guard.

4. Weak Grounding & Refusal Sentinels

If the model is asked a question that cannot be conclusively answered by the indexed corpus:

1. Weak Grounding (X-Dike-Grounding: weak):
   • If the answer text contains 0 verified citations (grounding.used: 0), Dike automatically attaches the HTTP header X-Dike-Grounding: weak.
   • Your frontend application should use this header to render a disclaimer or warning badge.
2. Explicit Refusal Sentinel:
   • Rather than inventing an answer, the model will output an explicit refusal sentinel (e.g. "لا أجد سنداً قانونياً كافياً في المصادر المفهرسة للإجابة على هذا السؤال").

Next Steps

• Explore the Search & Fusion Pipeline that powers chunk retrieval.
• Review the POST /v1/reason API Reference.