Skip to main content
A syntactically valid definition can still ask the wrong question. Test the population, comparison and explanation, then assess whether your connected sources can support the conclusion.

Use four evidence cases

For the encryption example in Create a custom control: Also inspect mixed populations: a known failure can make the control fail while other subjects remain unknown. Failure does not imply complete evidence coverage.

Check the boundaries

  • Try the threshold itself and values on either side. “At most five” includes five.
  • Confirm Boolean, numeric and text values are not being confused.
  • Inspect excluded and filtered-out subjects, not just the ones that remain.
  • Review observation time, source eligibility and stale evidence handling.
  • Check the same subject is used across sources; a similar display name is insufficient.
  • Review client overrides and the disabled state of the control and its parent standard.

Roll out in a reviewable sequence

1

Prepare an editable draft

Copy a suitable library template or create a standard. Library copies start disabled; clear Enabled when drafting an individual custom control.
2

Check evidence coverage

Use a test workspace or an agreed representative client context. Confirm the relevant integrations are connected, mapped and supplying the required predicates.
3

Inspect the definition and settings

Review scope, filters, expectations, thresholds and autonomy with the person responsible for the standard.
4

Enable deliberately and run checks

Account for the standard’s scope and client overrides before enabling it. Use Run checks and inspect the resulting assessments, evidence and evaluation time.
5

Review the wider impact

Investigate unexpected failures and gaps before relying on the standard across clients. Document intentional exceptions and identify who handles follow-up.
Use real observations in a test workspace or an agreed client context. Keep synthetic examples out of production evidence.

Recheck after changes

Repeat the relevant cases when predicates, source normalisation, filters, operators, defaults or client overrides change. A past result answers the previous question using the evidence available then. Keep automated assessment and remediation separate. An enabled control does not guarantee a supported automatic fix. See remediation and approvals.