# How we score an LLM Quality Audit

Scores are agreed with you before the final readout. Each area uses a 1–5 scale.

| Score | Meaning |
| --- | --- |
| 5 | Meets the agreed threshold with headroom. Regressions gated. |
| 4 | Meets the threshold. Minor gaps documented. |
| 3 | Borderline. Go-live only with explicit risk acceptance. |
| 2 | Fails on material cases. Remediate before scale. |
| 1 | Unsafe or unmeasured. No-go until evidence is rebuilt. |

## Six areas

1. **Golden set** — coverage, labels, freshness vs the live surface.  
2. **Groundedness and faithfulness** — answers tied to allowed sources.  
3. **Refusal and safety** — policy breaches, jailbreaks, unsafe tool use.  
4. **CI and regression** — evals on change, or only on a laptop.  
5. **Production sampling** — live traffic feeding the suite, or a frozen demo set.  
6. **Cost and latency** — unit cost and delay under realistic load.

## What the money buys

Engineering report with severity and engineer-day sizing. One-page exec summary. Repeatable suite. CI harness when repo access allows. 60-minute review with the person who ran the work.

From €8,000 for two weeks on an agreed surface. [LLM Quality Audit](/llm-quality-audit) · [24h PoC Spike](/ai-poc-sprint) · [Open proof](/#open-proof) · [Request scope](/#scope)
