Automated Triage & Confidence Thresholds
How Dori uses confidence scoring, severity levels, and the Attention Model to filter noise and surface 'Needs You' items.
One of the biggest failures of traditional notification systems is alert fatigue. Every incoming message, email, or capture triggers a ding, forcing humans to act as manual triage routers.
Dori solves this with an Automated Triage & Confidence Threshold Engine. Instead of notifying you about every raw event, Dori categorizes incoming captures along strict confidence thresholds.
1. The Attention Model & Severity Levels
Dori evaluates every situation across three severity states:
quiet: High-confidence routine actions (e.g. automatically transcribing a note, linking a file to a project, updating indexes). Runs 100% silently in the background.waiting: Asynchronous tasks executing on your behalf (e.g. background document parsing, research bundle generation).needs_you: Items that require explicit human judgment, confirmation, or decision approval.
2. Confidence Threshold Routing
When a workflow or action produces a result with a numeric confidence score (0.00 to 1.00) — based on entity matching, project affinity, and filing patterns — Dori’s ConfidenceGuard checks it against a single default threshold (currently 0.6, not yet configurable per workflow):
Pass ($\ge 0.6$): Proceeds Automatically
If confidence clears the threshold, the guard passes silently and the result proceeds — e.g. filed into your Markdown vault, or routed through the normal quiet / waiting / needs_you attention states without extra friction.
Warn ($< 0.6$): Flagged for Review
If confidence falls short, the guard returns a warn verdict instead of blocking outright, and the item is surfaced for you to review before it’s trusted to proceed automatically — rather than being silently auto-filed.
Separately, in the portal UI, ambiguous entity matches (e.g. a decision note that could belong to two client accounts) are shown with simplified categorical confidence labels (high / medium) rather than raw scores, so you can approve or correct the match quickly (⌘+Enter) from the Needs You strip.