What's new in Sovel

Three AI agents now run periodic health scans over the governed knowledge base: StalenessLintAgent (flags entries past their confidence decay window), ContradictionLintAgent (surfaces skills that conflict with newer evidence), and CoverageGapLintAgent (identifies high-frequency failure modes with no governed skill). All lint findings land in the reviewer inbox — nothing is auto-corrected.

New /drift view shows work orders where field practice diverges from the mapped procedure, with drift percentages and linked WO evidence. The 64% drift on the RAS pump declogging procedure in our synthetic dataset was the first real signal this surfaced.

Interactive /graph view shows assets, experts, failure modes, issues, and entries as a relationship graph. Highlighted nodes show sole-expert assets — the "what if they retire?" view that maintenance leaders actually need. Powered by vis-network with filter controls.

Reviewer edit patterns now feed a lightweight correction-inference layer that surfaces "suggested improvements" in future AI drafts for the same extraction dimension. The reviewer remains the decision-maker — the ribbon is advisory.

When a proposed knowledge entry conflicts with an existing governed Operations Skill, the system now raises a propagation alert in the reviewer inbox. The reviewer sees both the existing skill and the new proposal side-by-side, with the conflict reason highlighted. No silent overwrite.

New leadership dashboard with knowledge metabolism, net knowledge position, capture rate, governance throughput, placement rate, staleness percentage, and expert concentration — all in one view. Designed for reliability and maintenance managers, not just technicians.

Reviewer decisions (APPROVE, EDIT, REJECT, DEFER) now require a structured reason code. Reason codes are logged with timestamps and author attribution. The audit trail is complete — every governance decision is discoverable, not just the outcome.

First versioned Maintenance Ontology schema — typed nodes for assets, failure modes, experts, and Operations Skills with typed edge relationships. Operations Skills placed through the reviewer inbox are now bound to ontology nodes, enabling graph queries and contradiction detection at placement time.

Core issue detection pipeline live. Upload a work order export; the gap engine applies six detection rules and returns a ranked issue board. First real dataset: 4,218 WOs across a synthetic water treatment plant, 31 issues surfaced, 34 assets covered. This is the foundation everything else builds on.