Version: Cloud

v1.62

Release notes

Deployment-based guardrails
Enforce your own domain-specific safety logic directly using Deployments as guardrails. Whether you need to flag sensitive content, mask PII, or hard-block non-compliant requests, deployment-based guardrails give you full programmatic control and flexibility. Bring your own models, your own rules, and your own thresholds without being locked into a predefined safety framework.
Alerts on custom metrics
Alerts can now be configured on custom metrics. Registration deployments also support custom-metric alerts. For line/bar charts an alert is raised for every x-y value pair where the y value crosses the defined threshold; for timeseries the aggregation value (sum/max/min/average) is evaluated against the threshold.
Model cards and data cards as AI suggestion sources
Model cards and data cards can now be selected as data sources when generating AI suggestions for use case templates. Documents, model cards, and data cards can also be manually (re-)processed to make them available to the suggestion generator.
Clear inference data for deployments
Request logs, prediction logs, actuals, and evaluations can now be cleared by admin users for selected deployments before a specified date.
Data cards and model cards for repository-less deployments
Data cards and model cards can now be added to deployments that do not have a linked repository.

Guardrails configuration now includes flagging, masking, and blocking request options.
Datetime abbreviations now follow the updated design standard using dotted short forms (e.g. mth., wk., min.). Notification timestamps now show relative time; deployment overview timestamps now show full words (e.g. "Updated 3 hours ago").
Improved filter behavior and resolved various small filtering issues across the platform.
SHAP tree explainer additivity check is now disabled by default.

Fixed a bug where downloading the risk class report PDF on a completed risk classification assessment produced no output.
Fixed an issue with controls A005, A006 and A007 of the AIUC-1 framework being hidden in the use case control list when they should have been visible.
Fixed tab counts in the team registry not updating when filters are applied.
Fixed an issue with periodic review status changes for certain use cases.
Fixed a bug where removing a large number of controls from a custom control framework would silently fail due to URL query parameter length limits. Controls are now sent in the request body.