Review & Curation

The pipelines propose; a human decides. Review & Curation is the workspace where curators resolve open questions, approve or reject proposed changes, and repair anything that broke - so only reviewed, accepted meaning becomes the truth agents run on. Every action is a governed, versioned change.

The Learning Queue

The Learning Queue is where the semantic layer asks for help. As Colrows scans and reasons over a datasource, it raises typed questions (inquiries) a curator resolves one at a time. Open it from Review Inquiries on the home page. The queue groups open inquiries by type:

• Missing category - a coded column whose values need a governed meaning.
• Definition conflict - two or more definitions disagree on what a term means.
• Semantic ambiguity - meaning cannot be inferred from a name or its context, for example a column description that needs a human answer.

A curator opens an inquiry, reviews the details, and submits a resolution; the item moves from pending to completed.

Separately from the Learning Queue, proposed changes to an existing metric or business term are reviewed as an old-vs-new field diff you can approve, reject, or edit inline.

Needs Attention

Consensus → Needs Attention has two tabs. Invalid Entities lists assets that fell out of validity - a metric whose table was dropped, a verified query whose column changed - so a curator can repair them. Invalid assets never reach runtime; repairing one re-runs the full validator and, when it passes, restores it to a new valid version. Catalog also sends you here after a delete, so you can check the dependents it affected. Processing Failures (admins and curators) lists change-record processing that failed, per datasource, so you can retry or ignore it.

Change Explorer

Consensus → Change Explorer lets you find a change record, review its cause, and follow every downstream effect: which pipelines processed it, which entities it touched, and the change trail that led there. Use it to answer "why did this definition change?" before you edit anything.

Pipeline operations

Consensus → Pipelines is where curators and admins drive and watch the build:

• Run and schedule source scans (database metadata, column profiles, query history, analytics platforms), knowledge generation, publishing, and projection repair; pause and resume schedules.
• Save a per-database update plan that chooses the optional enrichment steps (profile columns and generate categories, generate missing descriptions, discover relationships, use query history). Metadata reading and the AI search update always run; a database update schedule reuses the saved plan.
• Watch run history live, and handle pipeline failures - retry or ignore a failed change so one bad item never stalls the rest.
• Finish onboarding a datasource: profile columns, generate missing descriptions, and publish - moving it from Draft to Ready for AI. If an optional step fails, the database still becomes available and the failure is listed on its overview.

Operation runs report their state as QUEUED, RUNNING, COMPLETED, DEGRADED (completed with issues), FAILED, or CANCELLED.

Manage via API

Review actions are driven from the console, and the same operations are available to administrators and curators on the platform API: