Sort every documentto the right queue.

It sorts every incoming document to the right queue. A routing rule pack you control matches each document on fixed signals — filename, keyword/text, sender, an amount threshold, a date present — assigns a document type and a destination, and evaluates the rules in priority order, first-match-wins. Each document gets a clear verdict: ROUTE (sent to a destination), QUARANTINE (held for a human), or UNMATCHED (no rule claimed it). It's the classify-and-route front door before extraction, validation, or matching — it decides where a document goes; it doesn't move or open it.

Because of the ambiguity gate. If a document matches two rules that point to different destinations, it's QUARANTINEd regardless of which rule has higher priority — the higher-priority rule does not silently win. That's deliberate: silently first-match-winning a conflicted document is exactly how a signed contract that happens to mention an invoice lands in accounts payable and a lawyer never sees it. A genuine cross-destination conflict is a person's call, so the machine refuses to guess and hands it to a human.

They're two different problems, kept deliberately distinct. QUARANTINE means the document matched but can't be safely auto-routed — either two rules sent it to two different destinations (a conflict), or its single match rested on fewer signals than your confidence floor (too weak). UNMATCHED means no rule matched it at all — a gap in your rule pack. Quarantine says 'your rules disagree about this one'; unmatched says 'your rules don't cover this one.' One you resolve by deciding; the other by adding a rule.

Neither — it's fully deterministic. There's no model, no probabilities, and no opaque confidence score: a rule either matches the document's signals or it doesn't, the same way every time, so the same document always routes the same place. That makes every routing decision auditable and explainable — you can point at the exact rule and signals that produced it. (For model-based field extraction from a document, that's the AI Document Extraction Kit, which runs upstream of this.)

The confidence floor is the minimum number of independent signals a single match must rest on before the kit will auto-route it. Set it to 1 and any single matching signal routes the document; raise it to 2 and a rule that fires on just one weak signal is QUARANTINEd instead, so a human confirms it. It lets you require corroborating evidence — a filename and a keyword, say — before a document is auto-sorted, without changing any of your routing rules. It's separate from the ambiguity gate (which is about conflicting destinations, not signal strength).

Make tiered rules mutually exclusive. If you want a high-value invoice to go to a manager and everything else to the standard AP queue, give the lower tier a disjoint condition (e.g. amount under $5,000) so each invoice matches exactly one tier — otherwise a large invoice matches both, and because their destinations differ, the ambiguity gate quarantines it as a conflict. The gate is meant for genuine cross-department conflicts (invoice vs legal), not deliberate priority refinements within one workflow. The included rule-design playbook walks through this.