Exceptions & human review
An exception is the machine saying "I am not sure." Everything about the review loop — who sees what, how much gets checked, what counts as done — is configuration, not code.
What raises an exception
Triggers are always an array — a single document can raise several at once, on several stages. A document with any open exception is in_review and is not delivered until it is cleared.
A stage that fails — OCR crash, connector timeout — produces a job error, visible in Jobs → Errors. An exception means the stage succeeded but the output needs a human. Do not route them to the same place.
Work queues
A work queue is a saved live query over the document catalogue, not a folder. Documents are never moved; they match a queue or they do not. System queues are generated per pipeline and exception type; custom queues are anything you can express as a filter.
The workstream. Stage tabs carry the open exception count; the Up next rail shows what the queue hands you after this one.
Access profiles
A profile defines what a person may see, independently of which queue they are working. Effective visibility is always queue ∩ profile. A document a profile excludes is invisible, not greyed out — the API returns 404 rather than 403, because the existence of a record is itself information.
Coverage rules
Not everything needs checking twice. Coverage is expressed per person as a flat list of rows — a condition and a percentage — and a correction's coverage is the highest percentage among matching rows.
| Condition | Coverage | Effect |
|---|---|---|
| Flat — all documents | 5% | Baseline spot-check for an experienced operator |
| New operator, first 30 days | 100% | Every correction verified during ramp |
| Total ≥ $50,000 | 100% | Money threshold, regardless of who worked it |
| Class = Credit memo | 25% | Higher scrutiny on a historically error-prone class |
The operator loop
Three flagged fields cleared in sequence — watch the remaining count fall from 3 to 1. Hosted externally.
- Claim — The operator opens a queue and is handed the next matching document. Claims are exclusive and time-boxed.
- Correct — Only fields that raised a trigger are focused. Each correction records the before value, the after value, the actor and the original confidence — this ledger is what trains the next model revision.
- Submit — The document either delivers immediately or, if a coverage rule matched, goes to a verifier as a second pass.
- Park or skip — Parked documents wait on an external answer and stay in the queue with a reason. Skipped documents are handed to someone else and counted separately in quality metrics.
What gets measured
- Wait — time from exception raised to claim. The number that tells you if you are staffed.
- Handling — active time per document. The number that tells you if the UI is working.
- Edit count — corrections per document. A rising average on one class means the model or the schema needs attention.
- Straight-through rate — documents delivered with no exception. The number the business actually cares about.