Scans & scheduling
You decide what gets scanned and how often. Incremental scanning keeps recurring runs fast and cheap.
Scope — what gets scanned
Scope is opt-in, per space. Nothing is scanned until an admin explicitly selects a space — there is no automatic, instance-wide crawl. You add and remove spaces in Settings → Scope, where each space shows an estimated page count and its model tier.
Each space can use one of two model tiers:
| Tier | Model | Use it for |
|---|---|---|
| Standard | Claude Haiku | The cost-efficient default — most spaces. |
| Deep | Claude Sonnet | High-stakes spaces where accuracy matters most, such as runbooks and policies. |
Schedules
For each scope you can choose how often it runs:
- Scan now — run a scan on demand for a chosen space. The picker confirms the estimated pages and budget impact before it fires.
- Weekly — a recurring scan once a week.
- Monthly — a recurring scan once a month.
- Off — no automatic scans; run manually whenever you choose.
Incremental scanning
The first scan of a space is a full scan. After that, recurring scans are incremental — they don’t re-analyze everything every time. Each incremental run re-checks:
- Pages that changed since the last scan.
- A rotating 10% sample of unchanged pages. This catches drift — when the world changed but the page didn’t. A page can become wrong because a system was renamed elsewhere, even though no one edited the page.
- Pages near new contradiction candidates, so conflicts surface as your content evolves.
Incremental scanning is why steady-state runs stay fast and inexpensive, and why your open-findings count should fall scan over scan as the team fixes pages.
The scan ledger
Every scan is recorded in the Scans tab of the dashboard: date, scope, pages planned and completed, findings produced, tokens used, a cost band, and status. If a scan ran into trouble on specific pages, expand the row to see the per-batch errors. Scans are designed to be resilient — a page that repeatedly fails analysis is marked as failed in the ledger and the scan moves on, so one bad page never wedges the whole run.
Partial scans and coverage
If a scheduled run can’t finish everything in its window — or scanning is paused because the monthly budget was reached — the app publishes what it has and tells you honestly. You’ll see a banner such as “Latest scan covered 62% of ENG Runbooks (resumes next window).” Partial results are flagged as partial, never presented as complete. For how the budget interacts with scanning, see Settings & budget.
Scale
The app supports up to 25,000 pages per installation with full coverage. Larger instances run in a “rolling coverage” mode that shows an honest coverage percentage on the dashboard rather than promising completeness. If you’re planning to point Evergreen AI at a very large estate, start with your highest-value spaces and expand from there.