Needs Attention¶
Needs Attention gathers everything that needs a person into one queue, worst first. It has three
tabs: Needs Attention, All Alerts and Alert Rules. Press g then a to get here, or
click the bell in the header.
The queue¶
The queue combines four sources:
| Source | Examples |
|---|---|
| Alerts | Active alerts on a model's performance, drift or safety |
| Risks | Open entries in the AI Risk Register |
| Reviews | Review requests waiting for a decision |
| Deadlines | Audit dates, regulatory dates and review or risk-treatment due dates from the compliance calendar |
Filter it¶
- Source and Severity, each with a count
- Deadlines due within: 7 to 90 days
Save a filter combination you use often with Saved views. See saved views.
Act on a row¶
Depending on what the row is: Acknowledge, Assign to me, Open review, Open risk, Open dossier or Open compliance. Acknowledging and resolving alerts are recorded in the audit trail under your name.
All Alerts¶
Every alert for your organization, with Total Alerts, Active, Acknowledged and Resolved counts, a search box, and status and severity filters. Acknowledge and Resolve are on each alert.
Alert Rules¶
Alert rules describe the conditions you want to be warned about: a metric, an operator (gt, gte,
lt, lte, eq, ne) and a threshold, with a severity. The tab lists each rule's Name,
Type, Severity, Status, Conditions, Times Triggered and Created date.
Create Alert Rule opens the rule form.
When rules are evaluated¶
Active rules for a model are evaluated automatically whenever new evidence arrives for it:
| Event | Values a rule can match |
|---|---|
Performance metrics are logged (client.metrics.create / create_batch) |
Each metric by its name, for example accuracy |
| A drift report is created | drift_score, drifted_features_count, and the report's numeric metrics as drift_<name> |
| A bias audit completes | overall_fairness_score, bias_detected (1 or 0), disparate_impact_ratio, demographic_parity_diff |
A rule whose conditions are met opens an alert, unless the rule is within its throttle window. New alerts appear in this queue and under All Alerts, are counted on the rule's Times Triggered, and are available through the API and MCP server. Escalation policies route alerts to your notification channels.
A rule's metric_name must match the name the value arrives under. Test a rule before relying on it
with client.alerts.evaluate_rule(rule_id, metric_values={...}) or the
whitebox_alerts_evaluate_rule MCP tool, which checks it without opening an alert.
Choose thresholds yourself. WhiteBoxXAI doesn't present a default threshold as a regulatory requirement.
Notifications¶
Delivery channels (Slack, email, webhook, SMS and PagerDuty) are set up by an administrator in Settings → Integrations, where each channel has Send test. See Settings.