agent hallucinated a policy clause that doesn't exist in the handbook: verification error

TL;DR

The agent invented a plausible-sounding clause. Require every policy verdict to cite the exact handbook section it rests on, verify the citation exists in the indexed handbook, and reject verdicts whose citations do not check out.

Error

agent hallucinated a policy clause that doesn't exist in the handbook

Steps

1. Change the agent's output contract: every policy verdict must include the handbook section number and quoted text. Expected: verdicts carry checkable citations.
2. Verify each citation against the indexed handbook text before accepting the verdict. Expected: fabricated sections fail verification.
3. On verification failure, make the agent re-derive the verdict from retrieved sections only. Expected: the second attempt grounds itself in real text.
4. Log all verification failures with the invented clause text. Expected: a measurable hallucination rate per agent version.
5. If the rate is high, constrain the agent to quote-then-judge: it must paste the relevant handbook passage before stating the verdict. Expected: grounding becomes structural, not optional.

When to use

• An agent cites policy clauses that do not exist.
• Verdicts sound authoritative but cannot be traced.

When not to use

• The agent misreads a real clause (a comprehension problem).
• The handbook itself is ambiguous (clarify the policy).

Tool compatibility

• Custom audit agents with handbook retrieval.

Variant phrasings

Agent invented a policy rule

Same failure, plain wording.

Unverifiable policy citation

The audit-trail phrasing.

Why it happens

Language models complete patterns fluently, and policy language is highly patternable. Without a citation-verification step, fluent invention is indistinguishable from real recall.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstWRHOX8j9FXGj7Gq2gO2Ig

Resolved from the public thread: https://vectle.com/posts/pstWRHOX8j9FXGj7Gq2gO2Ig

Thread query: agent hallucinated a policy clause that doesn't exist in the handbook: verification error

TL;DR: The agent invented a plausible-sounding clause. Require every policy verdict to cite the exact handbook section it rests on, verify the citation exists in the indexed handbook, and reject verdicts whose citations do not check out.