Guard

Guard — Hallucination Detection

Guard is a hallucination firewall. It checks whether LLM output is supported by source documents and returns a numeric verdict.

Usage

result = client.guard(
    text="The contract allows 90-day returns.",
    source_context="Our policy: returns within 14 business days.",
    mode="lexical",
)

result.verdict        # "verified" | "rejected" | "weak"
result.action         # "allow" | "block" | "review"
result.confidence     # 0.0 to 1.0
result.is_safe        # True if verdict == "verified"
result.is_blocked     # True if action == "block"
result.claims[0].reason  # "numerical_mismatch" | "partial_match" | None

Verdicts

Verdict	Meaning	Action
verified	All claims supported by source	allow
rejected	Claims contradict the source	block
weak	Partial or insufficient evidence	review

Modes

Mode	Speed	Best For
lexical	<1ms	Token overlap, fast checks
hybrid	~50ms	Token + semantic embedding
semantic	~500ms	Full embedding similarity

Mock (no API needed)

from wauldo import MockHttpClient

mock = MockHttpClient()
r = mock.guard("60 days", "14 days")
# verdict: "rejected" — mock detects numerical contradictions

Integration with any LLM pipeline

def safe_answer(question, context):
    answer = your_llm.generate(question, context)
    check = client.guard(text=answer, source_context=context)
    if check.is_blocked:
        return "Could not verify this answer."
    return answer