Insights: is the record healthy?

The Insights page answers the questions a lead asks about the record itself: how much review debt exists, how long drafts sit, how often the team changes course, and which areas have no recorded decisions at all.

What it measures

Six metrics, all computed from the event log and the review system, each with a per-node breakdown at the bottom of the page.

Review debt. Open review items by reason (review date reached, awaiting confirmation, dependency changed, condition met) and by age. At Acme Robotics, three items over three months old meant three decisions nobody owned anymore; the fix was reassignment, not more process.

Time to confirm. The median and 90th-percentile days from a draft’s creation to its confirmation, plus the drafts waiting right now. A rising median usually means agents are drafting faster than owners are reading.

Course changes. Confirmations against supersessions and reversals, by month over the last year. A workspace that never supersedes anything is not stable, it is stale. When Acme Robotics eventually outgrows AR-041 (Postgres full-text stays as the search engine), the supersession lands on the record with its reasons, and this chart is where that shows up.

Orphaned decisions. Decided records with no links and no evidence. They say what was decided but not why, and nothing connects them to the decisions they constrain.

Coverage. Decisions per node, and nodes with no decision activity in the last six months. A quiet node means one of two things: nothing was decided there, or decisions are being made without being recorded. Both are worth a conversation.

Capture mix. Where decisions enter the record: the web editor, Slack, Teams, or an agent over MCP. Every decision carries its channel from the moment it is created, so this needs no extra instrumentation.

Visibility

Every count applies the same node-visibility rules as the lists and search. A member who cannot open a restricted node does not see that node in the coverage table, and decisions on it are absent from every total, share, and trend they see. The numbers differ between people by design: an admin sees the whole record, a member sees the record they can access.

This is applied inside the aggregation, not as a filter on a drill-down. There is no way to infer restricted activity from a difference between two numbers on the page.

Reading it as a lead

The page is built to be screenshotted into a leadership deck. The tiles are the headline; the sections carry the explanation. Three patterns worth acting on: review debt older than a month (decisions without effective owners), a capture mix that is all web editor (the record is written after the fact, not at the moment of deciding), and quiet nodes that ship changes anyway (an unrecorded decision stream).