# 9. Can every consequential claim in an AI scouting report be traced to evidence?


**Interview objective:** Evaluate the report as a collection of checkable claims, not as fluent prose.

**Put to:** Sarah Rudd, Arielle Dror and the supplier’s AI product lead.

**Why it matters to buyers:** NIST’s Generative AI Profile identifies confidently presented false content and fabricated supporting logic or citations as risks. Its guidance is cross-sectoral, not a football-specific accuracy certification. ([nvlpubs.nist.gov](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf))

**A strong answer should cover:**

- Which statements come from retrieved records, calculated metrics, scout observations or model inference.
- Source identifiers, match timestamps, dates and model versions sufficient for verification.
- Treatment of conflicting sources and deliberately unavailable information.
- Whether a citation actually supports the claim rather than merely mentioning the player.
- Human review responsibilities and correction propagation.

**Sharp follow-up:** Ask the system a plausible question that the licensed evidence cannot answer. Does it abstain, identify the gap or invent a conclusion?

**Request this artefact:** An AI Scouting Report Evidence Check with these proposed labels:

| Label | Review rule |
|---|---|
| Supported | The underlying evidence substantiates the claim |
| Inference | Reasoning is identified, bounded and distinguishable from observation |
| Unsupported | Evidence is absent or does not substantiate the statement |
| Contradicted | Available evidence conflicts with the statement |
| Unverifiable within access | The reviewer cannot inspect the claimed support |

Report factual support separately from football judgement. A verified event count does not, by itself, validate a broad claim about a player’s mentality or future performance.

Proposed expert interview; not yet conducted. No expert quotation is attributed.