# Evidence Pack outline

Use this outline before go-live. Fill it in your environment. It is not a certificate and not legal advice.

Print or save to PDF from the browser if you need a shareable draft for engineering and counsel. The fillable HTML UI is at [/evidence-pack](/evidence-pack); this Markdown twin is the same section outline for tools and offline drafts ([/evidence-pack.md](/evidence-pack.md)).

## 1. System inventory

Name the LLM surfaces in scope: model or API class, retrieval corpus, tools or agents, environments, and owners. State what is in and what sits outside this pack.

- Product surface and user-facing behaviour
- Model or provider path (no secrets in the shared draft)
- Data sources and residency notes
- Owners for product, eng, risk, and ops

## 2. Golden set

Document the labeled cases that define “good enough.” Include coverage notes, label rules, and last refresh date.

- Case count by task type
- Pass criteria and severity labels
- Freshness vs current production behaviour

## 3. Eval results

Record scores for task success, groundedness or faithfulness, refusal or policy, and any adversarial pack you run. Always publish method limits next to the numbers.

| Metric | Score | Threshold | Status |
| --- | --- | --- | --- |
| Task success | — | — | — |
| Groundedness / faithfulness | — | — | — |
| Refusal / policy | — | — | — |
| Cost / latency (p95) | — | — | — |

## 4. Logs and monitoring

Show how production traffic is sampled into evals, where traces live, and which alerts fire on quality or cost spikes.

- Sampling rate and retention
- Trace or prompt store location
- Alert owners and on-call path

## 5. Human handoff

Define when the system must stop and escalate: confidence rules, refusal paths, and the human queue that receives the case.

- Handoff triggers
- Queue and SLA
- Audit trail for escalations

## 6. Kill switch

Name the control that disables or degrades the LLM feature fast. Record who can pull it, how long it takes, and what users see afterward.

- Control location and authority
- Time to effect
- Fallback behaviour

## 7. Go, remediate, or no-go

One page: verdict, top risks, effort to clear, and the decision owner. Attach the engineering detail behind it. Do not treat this outline as regulatory clearance.

- Verdict (go, hold, or kill)
- Top three risks with severity
- Effort band to clear hold items
- Decision owner and date

---

Support for EU AI Act readiness, when purchased as an add-on, means technical testing and documentation inputs for your counsel. It is not legal advice, not a conformity assessment, and not a guarantee under Regulation (EU) 2024/1689, MDR, or IVDR.

**Next:** [Request audit scope](/#scope) · [LLM Quality Audit](/llm-quality-audit) · [AI Readiness Check](/ai-readiness-check)
