CopilotForgeDraft

Templates › Engineering standards checklist

Engineering standards checklist

Template id: engineering-standards-checklist

Responsible skill: ai-engineering:ai-engineering-standards. Sections draw on ai-engineering:ai-solution-repository, ai-engineering:secrets-and-identity, ai-engineering:ai-api-and-tooling, ai-engineering:ai-solution-testing, ai-engineering:evaluation-harness, ai-engineering:ai-ci-cd, ai-engineering:ai-telemetry, ai-engineering:ai-performance-engineering and ai-engineering:ai-resilience.

Use this checklist during build of a custom AI solution and attach it to the design review. Every fact the organisation must supply is marked [To confirm: ...]. Example entries are fictional. Tick an item only when the evidence is attached or referenced. "N/A" needs a reason. An item that is mandatory at the solution's tier and is not met needs an approved exception in section 12, or the checklist result is "Not conformant".

1. Solution

FieldEntry
Solution name[To confirm]
Risk tier[To confirm: tier and date set]
Solution owner[To confirm: name and role]
Engineering lead[To confirm: name and role]
Repository[To confirm: location]
Standards version used[To confirm: version and date]
Checklist resultConformant / Conformant with exceptions / Not conformant

2. Repository and ownership

StandardMandatory from tierEvidenceStatus
Prompts stored as versioned files, one per prompt1[To confirm]Met / Not met / N/A
Model deployments and versions declared per environment1[To confirm]Met / Not met / N/A
Tool definitions and retrieval settings versioned1[To confirm]Met / Not met / N/A
Evaluation datasets versioned or referenced by immutable version2[To confirm]Met / Not met / N/A
Infrastructure as code for all resources2[To confirm]Met / Not met / N/A
Ownership file routes prompt and threshold changes to owners1[To confirm]Met / Not met / N/A
Main branch protected; review and passing checks required1[To confirm]Met / Not met / N/A
Readme, decision records and runbook present1[To confirm]Met / Not met / N/A

3. Identity and secrets

StandardMandatory from tierEvidenceStatus
Managed or workload identities for service-to-service calls1[To confirm]Met / Not met / N/A
No secrets in source, configuration files or pipeline variables1[To confirm: scan result]Met / Not met / N/A
User context for user-scoped data and tools1[To confirm]Met / Not met / N/A
Separate identities per environment1[To confirm]Met / Not met / N/A
Least-privilege role assignments, reviewed2[To confirm]Met / Not met / N/A
Secret scanning and push protection enabled1[To confirm]Met / Not met / N/A
Secrets register with owners and rotation dates2[To confirm]Met / Not met / N/A

4. Tools and APIs

StandardMandatory from tierEvidenceStatus
Narrow tools, one operation each, with clear descriptions1[To confirm]Met / Not met / N/A
Contracts defined in machine-readable specifications and versioned2[To confirm]Met / Not met / N/A
Server-side validation of every argument1[To confirm]Met / Not met / N/A
Write tools separate, idempotent and confirmed upstream1[To confirm]Met / Not met / N/A
Rate limits and timeouts declared and enforced2[To confirm]Met / Not met / N/A

5. Testing

StandardMandatory from tierEvidenceStatus
Unit tests with stubbed model responses1[To confirm]Met / Not met / N/A
Contract tests for tools and output schemas1[To confirm]Met / Not met / N/A
Failure-path tests (timeout, throttling, filter block, malformed output, tool error)1[To confirm]Met / Not met / N/A
Permission tests with users of different access2[To confirm]Met / Not met / N/A
Adversarial tests including injection through inputs, documents and tool results2[To confirm]Met / Not met / N/A

6. Evaluation

StandardMandatory from tierEvidenceStatus
Automated harness testing the deployed configuration2[To confirm]Met / Not met / N/A
Thresholds agreed by the business owner before results were seen2[To confirm: date agreed]Met / Not met / N/A
Model-graded evaluators calibrated against human labels2[To confirm]Met / Not met / N/A
Zero-tolerance failures defined and reported separately2[To confirm]Met / Not met / N/A
Results history stored with configuration per run2[To confirm]Met / Not met / N/A

Fictional example result: "Full set 240 cases; correctness 91% against threshold 88%; groundedness 96% against 95%; zero-tolerance breaches none; grader agreement with human labels 0.85 on 50 cases."

7. CI/CD

StandardMandatory from tierEvidenceStatus
Production deployment only through the pipeline2[To confirm]Met / Not met / N/A
Evaluation gate on prompt, model, tool and retrieval changes2[To confirm]Met / Not met / N/A
Production approval by named approvers per release policy2[To confirm]Met / Not met / N/A
Rollback covers code, prompts, model configuration and infrastructure2[To confirm]Met / Not met / N/A
Model version register with retirement dates2[To confirm: check source records for dates]Met / Not met / N/A
Progressive exposure with monitoring3[To confirm]Met / Not met / N/A

8. Telemetry

StandardMandatory from tierEvidenceStatus
End-to-end tracing with correlation ids2[To confirm]Met / Not met / N/A
Model, prompt and tool versions recorded per request2[To confirm]Met / Not met / N/A
Tokens, latency, errors and filter outcomes recorded2[To confirm]Met / Not met / N/A
Content capture decision agreed with the DPO or data owner2[To confirm: decision and date]Met / Not met / N/A
User identifiers pseudonymised in operational telemetry2[To confirm]Met / Not met / N/A

9. Performance

StandardMandatory from tierEvidenceStatus
Latency targets agreed at a stated percentile3[To confirm]Met / Not met / N/A
Per-step latency and token budgets3[To confirm]Met / Not met / N/A
Capacity plan against expected peak3[To confirm]Met / Not met / N/A
Load test results against targets3[To confirm]Met / Not met / N/A

10. Resilience

StandardMandatory from tierEvidenceStatus
Timeouts on every external call2[To confirm]Met / Not met / N/A
Bounded retries with backoff and jitter on idempotent calls only2[To confirm]Met / Not met / N/A
Fallback deployments or models evaluated before use2[To confirm]Met / Not met / N/A
Degraded mode and user messages designed with the business owner2[To confirm]Met / Not met / N/A
Failure tests run in a test environment3[To confirm]Met / Not met / N/A

11. Operations hand-over

ItemEntry
Support team and route[To confirm]
Monitoring signals and alert routes[To confirm]
Runbook location[To confirm]
Next review date[To confirm]

12. Exceptions

StandardReasonCompensating controlRisk acceptedApproverReview date
[To confirm][To confirm][To confirm][To confirm][To confirm][To confirm]

13. Sign-off

RoleNameDecisionDate
Engineering lead[To confirm][To confirm][To confirm]
Design authority reviewer[To confirm][To confirm][To confirm]

Credit: Template from CopilotForge by Simon Foote (https://github.com/SimonFoote-greensparx/CopilotForge), licensed under CC BY 4.0. Keep this credit when you share or adapt the template; in your own completed document you may keep it as a footer or remove it.

Download the template (Markdown)