Every release, dated and tagged.

Runs now reserve the crew's recent P90 run cost plus a variance buffer at admission and settle at each checkpoint. A run that exhausts the wallet parks in awaiting_credit for 72 hours and auto-resumes on top-up rather than failing. Side-effect Actions are withheld while a wallet is short.

Sampling regimes now report reviewer agreement, median decision time and override rate per reviewer. Autonomy graduation is blocked when reviewer agreement falls below the configured bound.

Change plans now show the manifest diff, the validation result and the sample test run on one screen, with per-hunk accept. Previously accept was all-or-nothing.

Clause library refreshed to 31 July 2026 currency. Twelve new evaluation cases drawn from disputed tenancy matters. Change notice sent to entitled tenants.

Exports over 10,000 trace records truncated silently at the page boundary. Records were never lost, but the export was incomplete. Affected exports can be regenerated.

A support view could render the last four characters of a tenant-managed secret reference identifier. No key material was ever exposed. Reported through the disclosure programme, fixed within 48 hours.

Every declared external effect must state what happens if it needs to be undone. Undoable actions declare compensation: none explicitly, which is a decision rather than an omission.

Route a configurable share of production traffic to a candidate release with automatic reversion on evaluation-gate breach. Professional and above.

Publish verified components to your own organisation without publishing to the world. Lineage and verification tier are tracked identically.

Superseded by runs.search with cursor pagination. Removal no earlier than v1.4.0 or 30 November 2026, whichever is later.

When a source expires or materially changes, every artefact that relied on it is identified and its owners notified, with the affected claims listed. Previously staleness was visible only on the Knowledge Pack itself.

Where sources of comparable authority disagree, both positions are carried into the evidence record and surfaced to the reviewer. The system no longer silently prefers the higher-ranked retrieval.

Adds barramundi and grouper knowledge packs, revised dissolved-oxygen safety rules, and 40 evaluation cases built from real incident reports.

Failures now print the full path from untrusted input to the ungated effect, rather than naming only the offending node.

Three moderate advisories in the code-execution sandbox base image. No exploitation observed. Patched image rolled to all regions within the maintenance window.

Escalating an approval task could notify the original assignee twice. The Correction Record and the audit trail were always correct; only the notification duplicated.

Per-node credit estimates now split platform and model meters, and respect a BYOK configuration when one is attached.

Four-stage graduation with risk-weighted sampling, published confidence bounds and automatic reversion on breach. Regulated crews can be pinned to the permanent full-review floor.

The validator now proves statically that no path runs from untrusted input to an external side effect without an intervening approval gate. If the property does not hold, no release is produced.

Draft, Tested, wGrow verified, Tenant verified and Deprecated, each with published criteria. Compatible updates may be recommended but are never installed silently.

Deployed crews are callable as MCP tools with policy and approval enforced by the runtime. Approved external MCP servers are consumable as Tools, with per-tool admission control.