Hallucination Detection

The hallucination detection spots the ungrounded claims in the production output (L336) — the hallucination (the fluent untruth: the confident claim with no basis, L336), the signals (the uncertainty L139, the contradictions, the missing citations L192, L336), and the guard (the checks: the NLI L336 and the LLM-judge L343, the citations L192, and the fallback L336) — the RAG's (L280) untruth watch (L336).

The one-sentence interview answer: "The hallucination detection spots the ungrounded claims (L336). The hallucination (L336): the fluent untruth (L336) — the confident claim (L336) with no basis in the retrieved context (L189) or the model's knowledge (L336) — the model (L278) generates what sounds right (L336). The signals (L336): the uncertainty (L139) — the low probability (L336); the contradictions (L336) — the answer vs the context (L189); the missing citations (L192) — the ungrounded claims (L336) lack the sources (L192). The detection (L336): the NLI (L336) — the natural language inference: does the answer follow from the context (L337); and the LLM-judge (L343) — the second model (L343) scoring the groundedness (L337). The guard (L336): the citations (L192) — the grounded answers (L280) cite the sources (L192); the check (L336) — the low-groundedness (L337) flagged (L336); and the fallback (L336) — the flagged answer (L336) refused or corrected (L336). The AI shape (L173): the RAG (L280) — the ungrounded claim (L336) in the answer (L336) — the detection (L336) and the guard (L336) — the fluent untruth (L336), caught (L336)."

Think of the hallucination detection as the fact-checker at the newsroom. The reporter (the model, L278) writes the story (the answer, L328) — fluent (L336) but sometimes invented (L336): the sources (the retrieved context, L189) checked (L336). The fact-checker (the detector, L336): the source check (the citations, L192) — the un-cited claims (L336) flagged (L336); the contradiction check (the NLI, L337) — the story vs the sources (L336); and the editor's review (the LLM-judge, L343) — the second opinion (L343). The flagged stories (L336) — the low-groundedness (L337) — are withheld (the fallback, L336): the correction (L336) or the refusal (L336). The newsroom works because the sources are cited, the claims are checked, and the flagged stories don't print (L336).

   the fact-checker (the detector, L336)
   ┌────────────────────────────────────────────────────────┐
   │ the sources (the context, L189) · the citations        │
   │ (L192)                                                 │
   │ the checks (L336): the NLI (L337), the judge (L343)    │
   │ the withheld (the fallback, L336) — the correction,    │
   │ the refusal (L336)                                     │
   └────────────────────────────────────────────────────────┘

The mental model is the fact-checker: the sources, the checks, and the withheld (L336).

   the answer (L328)
        │  the fluent claim (L336)
        ▼
   ┌────────────────────── THE SIGNALS (L336) ──────────────────────────┐
   │  the missing citation (L192) · the low probability (L139)         │
   │  the contradiction vs the context (L189, L337)                    │
   └──────────────────────────┬───────────────────────────────────────┘
                              ▼
   ┌────────────────────── THE DETECTION (L336) ────────────────────────┐
   │  the NLI (L337): the answer follows from the context? (L336)      │
   │  the LLM-judge (L343): the groundedness score (L337)              │
   └──────────────────────────┬───────────────────────────────────────┘
                              ▼
   ┌────────────────────── THE GUARD (L336) ────────────────────────────┐
   │  the grounded (L337) → the answer with the citations (L192)       │
   │  the ungrounded (L336) → the fallback (L336): the correction,     │
   │  the refusal (L336)                                               │
   └──────────────────────────────────────────────────────────────────┘

The flow is the detection: signals → detection → guard (L336).

• The hallucination (L336). The fluent untruth (L336): the confident claim (L336) with no basis (L336).
• The signals (L336). The uncertainty (L139), the contradictions (L337), the missing citations (L192).
• The detection (L336). The NLI (L337) — the natural language inference; the LLM-judge (L343) — the second model (L343) scoring the groundedness (L337).
• The guard (L336). The citations (L192), the check (L336), and the fallback (L336) — the correction (L336) or the refusal (L336).

• A RAG platform (L280). The grounded answers (L337) with the citations (L192) — the ungrounded (L336) flagged (L336).
• A chat product (L162). The detection (L336) on the free-form answers (L328) — the fallback (L336) for the ungrounded (L336).
• A medical or legal copilot (L371). The high-stakes (L371) — the detection (L336) mandatory (L336), the refusal (L336) the default (L336).
• A search summarizer (L336). The summaries (L336) checked against the sources (L189) — the NLI (L337).
• Anything RAG (L280). The untruth watch (L336) — the fluent claim (L336), caught (L336).

The through-line: the watch is the untruth's — the signals, the detection, and the guard (L336).

Say it in four moves:

1. The hallucination. "The fluent untruth — the confident claim with no basis (L336)."
2. The signals. "The uncertainty (L139), the contradictions (L337), the missing citations (L192)."
3. The detection. "The NLI (L337) and the LLM-judge (L343)."
4. The guard. "The citations (L192), the check (L336), the fallback (L336)."

• The hallucination is the fluency's trap (L336). The model (L278) generates what sounds right (L336) — the fluent untruth (L336) is the risk (L336).
• The grounding is the RAG's fix (L280). The grounded answer (L337) — the retrieval's (L189) context (L280) — the hallucination (L336) reduced (L336).
• The citations are the evidence (L192). The cited answer (L192) — the user (L162) verifies (L192) — the un-cited claim (L336) flagged (L336).
• The judge is the scalable check (L343). The LLM-judge (L343) — the groundedness (L337) scored (L343) at the scale (L343) — the human review (L341) sampled (L341).
• The fallback is the trust (L336). The flagged answer (L336) refused (L336) — the product's (L173) honesty (L336).

• The fluency trusted (L336). The confident answer (L336) taken at face value (L336) — the fluent untruth (L336) ships (L336).
• The detection missing (L336). The ungrounded (L336) unflagged (L336) — the RAG (L280) answers from the air (L336).
• The citations absent (L192). The answer (L328) without the sources (L192) — the verification (L192) impossible (L336).
• The fallback missing (L336). The flagged answer (L336) still delivered (L336) — the trust (L336) broken (L336).
• The judge un-validated (L343). The judge's (L343) bias (L343) unchecked (L343) — the L343 validation (L343) is the judge's (L336).

• Ground the answers (L280) — the retrieval's (L189) context (L280).
• Cite the sources (L192) — the evidence (L192) in the answer (L336).
• Check the groundedness (L337) — the NLI (L337) and the judge (L343).
• Fallback on the low (L336) — the correction (L336) or the refusal (L336).
• Validate the judge (L343) — the agreement with the humans (L341).

Q: Walk me through the hallucination detection.

A: Spotting the ungrounded claims (L336). The hallucination — the fluent untruth (L336). The signals — the uncertainty (L139), the contradictions (L337), the missing citations (L192). The detection — the NLI (L337) and the LLM-judge (L343). And the guard — the citations (L192), the check (L336), the fallback (L336).

Q: Why do models hallucinate?

A: The generation (L336): the model (L278) predicts the next token (L336) — what sounds right (L336) — not what's true (L336). The fluency (L336) and the confidence (L336) don't imply the grounding (L337). The RAG (L280) reduces it (L336): the grounded answer (L337) from the retrieved context (L189).

Q: How do you detect it in production?

A: Per answer (L336): the NLI (L337) — does the answer follow from the context (L189); and the LLM-judge (L343) — the groundedness (L337) scored (L343). The signals (L336) — the missing citations (L192), the contradictions (L337) — plus the checks (L336), and the low-groundedness (L337) flagged (L336).

Q: What do you do with the flagged answers?

A: The fallback (L336): the correction (L336) — the grounded re-answer (L280); or the refusal (L336) — the honest "I don't have the source" (L336). The high-stakes (L371) — the medical and the legal (L371) — refuse by default (L336).

• What's the hallucination (L336)?
• Why do models hallucinate (L336)?
• How do you detect it (L336)?
• What do you do with the flagged (L336)?
• What's the NLI (L337)?

The senior read: the eval measures; the detection guards — the two groundedness checks (L336).

// The hallucination detection (L336) — the production guard (L336).
// 1 · THE GROUNDED ANSWER (L280) — the citations (L192).
async function groundedAnswer(query) {
  const chunks = await retrieve(query);          // the retrieval (L189)
  const answer = await model.invoke({ query, context: chunks });
  return { answer, chunks };                     // the evidence (L192)
}

// 2 · THE DETECTION (L336) — the groundedness check (L337).
async function checkGroundedness(answer, chunks) {
  // THE NLI (L337): does the answer follow from the context?
  const nli = await nliScore(answer, chunks);    // L337

  // THE LLM-JUDGE (L343): the groundedness score (L337).
  const judged = await judge.groundedness({ answer, chunks });  // L343

  // THE SIGNALS (L336): the citations' coverage (L192).
  const cited = chunks.some((c) => answer.includes(c.citationId));  // L192

  return { score: (nli + judged.score) / 2, cited, grounded: judged.score >= 0.85 };
}

// 3 · THE GUARD (L336) — the fallback (L336).
async function guarded(query) {
  const { answer, chunks } = await groundedAnswer(query);
  const check = await checkGroundedness(answer, chunks);
  if (!check.grounded || !check.cited) {
    return { fallback: 'refusal', message: 'I could not verify this from the sources.' };  // L336
  }
  return { answer, citations: chunks.map((c) => c.source) };   // L192
}

What the reader must SEE — the watch, applied:

  retrieve + context        → the grounding (L280, L189)
  nliScore + judge          → the detection (L337, L343)
  citations coverage        → the evidence (L192)
  grounded < 0.85 → refusal → the fallback (L336)

  The fluent claim checked, the ungrounded refused (L336).

• The check is the answer's latency (L336). The NLI (L337) and the judge (L343) — the seconds (L336) per answer (L336) — the sampling (L341) for the interactive (L162).
• The judge is the cost (L343). The second model (L343) — the tokens (L332) — the sampling (L341) bounds it (L336).
• The citations are the zero-cost evidence (L192). The retrieval's (L189) sources (L192) — no extra cost (L336).
• The fallback is the trust's cost (L336). The refusal (L336) — the missed answers (L336) for the honesty (L336).

• The hallucination detection = the untruth's watch (L336): the hallucination, the signals, the detection, the guard.
• The hallucination: the fluent untruth (L336).
• The signals: the uncertainty (L139), the contradictions (L337), the missing citations (L192).
• The detection: the NLI (L337) and the LLM-judge (L343).
• The guard: the citations (L192), the check (L336), the fallback (L336).

HALLUCINATION DETECTION = spotting the ungrounded claims

THE HALLUCINATION (L336)
  the fluent untruth (L336) — the confident claim (L336)
  with no basis (L336) — what sounds right (L336)

THE SIGNALS (L336)
  the uncertainty (L139) · the contradictions (L337)
  the missing citations (L192)

THE DETECTION (L336)
  the NLI (L337) — the answer follows from the context? (L336)
  the LLM-judge (L343) — the groundedness score (L337)
  per answer (L336) — the production (L307) check (L336)

THE GUARD (L336)
  the citations (L192) — the evidence (L192)
  the check (L336) — the low-groundedness (L337) flagged (L336)
  the fallback (L336) — the correction (L336) or the refusal (L336)

INTERVIEW, 4 MOVES
  1 hallucination "the fluent untruth (L336)"
  2 signals      "the uncertainty, the contradictions, the citations (L336)"
  3 detection    "the NLI and the judge (L337, L343)"
  4 guard        "the citations, the check, the fallback (L336)"

You now hold the watch: the signals, the detection, and the guard — with the sources cited and the ungrounded withheld. The newsroom checks the stories — and the fluent untruth doesn't print (L336).

Next: does the answer follow from the retrieved evidence? — Groundedness Evaluation (L337).