Problem it solves

Low-confidence output is presented as confident output. Users make decisions based on AI responses without knowing the model's uncertainty level.

When to use

Whenever model confidence drops below the operator-configured threshold on a consequential output.

When not to use

For confidence thresholds on low-stakes, exploratory output. Not every low-confidence response requires a governance event.

Governing principle

When confidence falls below threshold, the UI must surface a visual breach state, name the source of uncertainty, and offer concrete recovery options. Auto-proceeding on low-confidence output is not acceptable.

Required Components

Interaction Flow

1

Output generated

The model produces a response and returns its confidence score.

2

Threshold check

The system compares the confidence score against the operator-configured threshold.

3

Breach state surfaces

If the score is below threshold, the Confidence Threshold Warning activates a visual breach state on the output.

4

Uncertainty source named

The Uncertainty Indicators component identifies what is uncertain: data recency, conflicting sources, ambiguous input, or model limitation.

5

Recovery options presented

The user can send for human review, adjust the threshold for this output, or override with documented justification.

6

Decision logged

The user's choice is written to the governance record, including what threshold was active and what justification was provided.

Governance requirements

Threshold breach events are governance events. Each breach must be logged with the confidence score, threshold at time of breach, uncertainty source identified, and the user's recovery action.

Accessibility notes

Threshold breach states must be announced via role="status" (not role="alert" unless the breach blocks the workflow). Visual breach indicators must pass WCAG contrast requirements.