New
Blind-sample review. Configure a percentage of autonomous decisions to be routed for independent human re-work, with the system's own outcome hidden from the reviewer. Disagreements are collected, categorised and available as candidate evaluation cases.
This is the control we most often see missing in governance designs, and it now ships as part of the platform rather than as something each estate builds.
Cost per decision. Reported natively, combining serving, retrieval, evaluation, amortised retraining, attributed idle capacity and human review time. Available per use case and per business unit.
Structure-aware chunking by default for supported document types, carrying the section path as retrieval metadata. Fixed-window chunking remains available.
Improved
- Policy gates are versioned separately from agent definitions, with their own change-control trail.
- Idempotency keys are now required on state-changing tools rather than recommended.
- Trace retention is configurable to match the underlying business record rather than the log retention default.
- Quarantine now has a first-class expiry policy with an audit record for expired items.
Fixed
- Fixed reviewers occasionally seeing the automated outcome in the blind-sample preview pane.
- Fixed idle capacity being attributed evenly rather than to the reserving workload.
- Fixed a case where an escalated request was billed to the escalation route rather than the originating workload.
Deprecated
| Item | Replacement | Removal |
|---|---|---|
| Prompt-embedded policy rules | Runtime policy gates | 2026.9 |
| Shared agent/policy versioning | Independent policy versions | 2026.11 |
- Release
- Governance
- Retrieval
- Cost