AI Assurance Evidence Pack

One evidence package for deciding whether an AI system can enter production and remain under control. Complete it for a specific system version — not for a generic “AI policy”.

How to use it: fields with a line can be edited in the browser before printing. Attach links or identifiers for real artefacts: test repositories, reports, logs, dashboards and decisions. Every result should identify the model, prompt, data and code versions.

System inventory and owner card

System name / use case

Business owner

Technical owner

Stage and environment

Version covered by this decision

Suppliers and integrations

Purpose, users and decisions supported

Input data, sensitive data and permissions

AI Act risk classification

Record the preliminary qualification and its basis. Route uncertainty to compliance or legal counsel.

Question	Yes / no / n.a.	Rationale and evidence	Owner
Does the system perform a prohibited practice?
Is it a high-risk use case under Annex I or III?
Do Article 50 transparency obligations apply?
Organisation role: provider, deployer, importer or distributor?
Other regimes: GDPR, sector rules, cybersecurity, IP?

Golden set and acceptance criteria

Golden-set scope

Version and location

Normal, edge and refusal cases

Metric / scenario	Acceptance threshold set before testing	Consequence if missed	Decision owner
Correctness / task success
Grounding / citations
Security / data
Cost and latency

Evaluation plan and results

Test	Method / set	Result	Threshold	Status	Artefact
Answer quality
Hallucination and grounding
Prompt injection / red team
Permissions and data leakage
Fallback and integration failure
Change regression

Material test limitations

Logs, versioning and monitoring

Input, output and operation time
Model identifier and prompt version
Retrieval sources and passages
Tool calls and action outcomes
Decisions, refusals, fallbacks and escalations
Unit cost and limits
Access control and log retention
Alerts and incident owner

Signal	Threshold / SLO	Source / dashboard	Response and owner
Quality
Security
Cost / volume
Availability / fallback

Human handoff

Handoff triggers

Who takes over and when

What the human sees

How feedback returns to the system

Kill switch and rollback

Stop scenario	Who can disable	Mechanism	Safe mode / fallback	Last tested
Security incident
Quality degradation
Invalid agent action

Model, prompt, data and code rollback plan

Go/no-go decision

☐ GO
thresholds met

☐ CONDITIONAL GO
limitations and remediation date

☐ NO-GO
production blockers

Decision rationale

Approved use scope

Limitations and reassessment date

Remediation plan and sign-off

Gap / risk	Priority	Action	Owner	Due date	Closure evidence

Business owner — decision / date

Technical owner — decision / date

Security / privacy — opinion / date

Compliance / legal — opinion / date

Important: this template organises technical evidence and operational decisions. It does not replace legal advice or a formal conformity assessment, does not guarantee AI Act compliance, and proves no quality without attached results, logs and approved criteria.