Controls

A control is the unit DialogAssert is built around. Controls are grouped into an approved, versioned control set. It is not a test an engineer wrote; it is an approved statement of what the business requires, with an owner, a version and an approval record.

control_set:
  id: lost-card-block
  owner: Cards Operations
  version: 7
  status: approved
  approved_by:
    - Cards Product
    - Customer Operations Risk
  approved_at: 2026-08-21

controls:
  - id: confirm-before-block
    expectation: Agent obtains customer confirmation before blocking a card.
    severity: high
    evidence_required:
      - conversation
      - action_sequence

  - id: selected-card-only
    expectation: Only the card selected by the customer may be blocked.
    severity: critical
    evidence_required:
      - entity_selection
      - action_argument
      - post_condition

Why the metadata matters

A finding is only as defensible as the expectation behind it. owner, version and approved_by turn “the tool says this failed” into “the agent did not do what Cards Operations approved on 21 August, in version 7 of the control”.

Every report cites the control version it evaluated. If the control changes, previous results are not silently reinterpreted.

Controls

A control is one thing that must be true. Keep them singular — a control that combines two requirements produces a finding nobody can act on.

Field	Purpose
id	Stable identifier. Findings, baselines and reviewer decisions all key on it.
expectation	One sentence, in business language. This is what appears in the report.
severity	critical, high, medium or low. Critical controls gate the release.
evidence_required	What must be observable for this control to be evaluated at all.

Evidence requirements are a first-class field

evidence_required is what makes inconclusive results possible. If a control needs post_condition and the environment cannot supply it, DialogAssert reports an assurance gap rather than guessing.

Declaring the evidence up front also means discovery can tell you which controls this environment can evaluate before you run anything.

What a control set is not

• It is not derived from the agent’s source code, prompts or traces. Expectations come from the business; discovery only supplies identifiers.
• It is not a prompt for a judge model. Most controls are settled deterministically.
• It is not a compliance attestation. It is a technical expectation that a human approved.

Related

• Journeys — the conversation a control set is exercised through.
• Evidence — what can be observed, and how strongly.
• Baseline comparison — how a control’s result is compared with the last approved release.