# Hallucination Testing Checklist

**System:** _______________  **Date:** _______  **Tested by:** _______________

> Do not test "is it true?" — that requires someone who already knows the answer. Test
> **"is it grounded?"** — does every claim trace to something the system was actually given.

## 1. Setup

- [ ] Retrieved context captured alongside every output — grounding cannot be checked without it
- [ ] Model and version pinned
- [ ] Temperature recorded
- [ ] Test set includes questions the corpus **can** and **cannot** answer

## 2. The core test — source attribution

- [ ] Output split into individual factual claims
- [ ] Each claim checked against the supplied context
- [ ] Unsupported claims counted → **this is your fabrication rate**
- [ ] Judge model is **different** from the generator
- [ ] Judge asked a narrow question ("is this claim supported: yes/no"), not a broad one
- [ ] Judge calibrated against human review on a sample

## 3. The no-answer test — highest value, usually skipped

- [ ] Questions included whose answer is **genuinely absent** from the corpus
- [ ] Correct behaviour defined: say so explicitly
- [ ] **Refusal rate measured**
- [ ] Refusal instruction present in the prompt ("if the passages do not contain the answer, say so")

> 🔴 A system scoring 95% on answerable questions and fabricating on 100% of unanswerable ones is
> dangerous precisely where the user is most dependent on it.

## 4. Identifier verification — cheapest, catches the worst errors

- [ ] Every number in the output checked against source
- [ ] Every date checked
- [ ] Every code, reference and identifier checked
- [ ] Every proper noun checked
- [ ] Percentages in prose checked against a source that actually contains them

## 5. Consistency

- [ ] Same question asked repeatedly (5+ runs)
- [ ] Answers compared for contradiction
- [ ] Inconsistency treated as **strong evidence of fabrication**
- [ ] Consistency not treated as proof of correctness

## 6. Context manipulation

- [ ] **Poisoned context** — a deliberate falsehood in a document. It *should* be repeated (proves grounding)
- [ ] **Empty context** — nothing supplied. Anything returned is coming from training data
- [ ] **Irrelevant context** — passages that do not answer. Must refuse
- [ ] **Contradictory context** — two sources disagree. Must surface both

## 7. The counterweight

- [ ] **Over-refusal measured** — refuses when the answer *was* present
- [ ] Tracked alongside groundedness

> Eliminating fabrication by refusing everything is not a fix. Teams that optimise only for
> groundedness drift there without noticing.

## 8. Where fabrication concentrates — verify each

- [ ] When retrieval returned nothing useful
- [ ] Specific identifiers and citations
- [ ] Numbers appearing in prose
- [ ] Bridging between two passages
- [ ] Attribution ("according to the policy…")
- [ ] **Long outputs** — fabrication rate rises with length

## 9. Metrics

| Metric | Value | Target |
|---|---|---|
| Groundedness (claims traceable) | | primary |
| Refusal accuracy | | high |
| Identifier accuracy | | ~100% |
| Consistency across runs | | directional |
| Over-refusal rate | | low |

## 10. Mitigations verified in place

- [ ] Explicit refusal instruction in the prompt
- [ ] **Sources shown in the interface** — makes remaining errors catchable
- [ ] Retrieval quality addressed before prompt tuning
- [ ] Output length constrained
- [ ] Precise values extracted programmatically, not generated

## Sign-off

| | Name | Date |
|---|---|---|
| Tested by | | |
| Reviewed by | | |