Reviewing Digitised Files
Working through digitised files from first pass to the moment you put your name to them
How the queue is ordered
The review queue is sorted by confidence ascending: the items the classifier was least sure about come first. That is the opposite of a normal inbox, and it is deliberate: the fastest way through six thousand documents is to spend your attention where the machine was weakest and to let the strong end of the distribution be handled in bulk.
Every item in the queue shows the page image beside the extracted fields, the suggested document type, and the box and file reference from your manifest, so you can check the extraction against the paper without leaving the row.
Confidence bands
Each item is scored and lands in one of three bands. The band boundaries are set centrally and shared with the rest of our document verification, rather than being invented for archive work, so an item in a given band means the same thing here as it does anywhere else on the platform.
- Failed: the extraction could not be trusted at all. These items sit at the head of the queue and are never confirmed in bulk: each one is opened and confirmed, corrected or refiled on its own.
- Needs review: the extraction is plausible but not strong enough to accept unseen. These come next in your queue, and are the ones worth your reviewer’s time.
- Auto-accepted: the extraction was strong. These items are accepted without per-item review, but they are not taken on trust; a sample of them is pulled for quality assurance before the batch can be signed off.
Bulk confirm
Where a set of items is plainly right (one document type, one client, one box, all scoring well), you can tick them and confirm the selection in a single action instead of clicking through it row by row. Bulk confirm is what makes a back-book tractable.
- Filter, tick, then confirm: the action applies to exactly the items you ticked, never to whatever a filter happens to match when you press it, and the count is stated before you commit
- Items in the failed band are never eligible for bulk confirm. Selecting a whole page leaves them out, and a selection that includes one is refused as a whole, with nothing confirmed
- Every confirmation is attributed to the person who made it, with a timestamp, and appears on the batch record
- Confirming does not lock the item; a mistake found later is corrected in the normal way and the correction is recorded
Field-level correction
Where an item is right in the main but wrong in one detail (the date read from the wrong stamp, a name transposed, the document type close but not exact), you correct the individual field rather than rejecting the whole item.
Both values are kept. The original extracted value and your corrected value are stored together as an audit trail, so the record shows what the machine proposed and what your firm decided. That trail is also how we measure accuracy honestly: corrections are the measurement, which is why the throughput and accuracy figures we eventually publish will come from real consignments rather than from a demonstration.
If a document was filed against the wrong client, or read as the wrong type of document, refile it from the queue instead: choose the client it belongs to, the correct document type, or both. What was read from the page stays as it was.
The quality-assurance sample
What gets sampled
A sample is drawn from the auto-accepted items (the ones nobody looked at) and put in front of a reviewer: each sampled item is marked as part of the QA sample in the review queue, and a filter shows the sample on its own. The sample is 2% of the auto-accepted items or 100 of them, whichever is the greater (all of them, if there are fewer than 100), selected at random on our side. You cannot choose the sample, and the selection parameters are held centrally rather than set per batch.
Why it exists
“The classifier was confident” is exactly how a golden-rule medical certificate ends up filed as ordinary correspondence, and nobody notices until it is the document a disclosure request turns on. A sample is also what an evidential-weight standard such as BS 10008 generally expects: evidence that quality checks were carried out, rather than an assertion that they were unnecessary.
What the sample decides
Corrections made during the spot check produce an observed error rate for the batch. Below the threshold, the batch proceeds to sign-off with the sample size and error rate recorded. At or above it, the batch cannot be signed off at all: it is routed to full review or sent back for a re-scan. The threshold is configured on our side, so it cannot be waved through under deadline pressure.
What sign-off commits you to
Your firm attests, we do not
Batch sign-off is an attestation by a member of your firm with the authority to make it. We record that attestation; we never make it. We never saw your paper, so we are not in a position to say that an image faithfully represents it. Keeping that statement with the record owner is what an evidential-weight standard expects in any case.
What the record captures
- Who signed off, when, and from where
- The consignment reference, the source of the paper and the operator who scanned it
- The scanning device and the resolution used
- Declared and received counts, and the reconciliation result
- The sample size and the observed error rate
- The disposition of the original paper
The same content is rendered as a certificate and filed to your firm’s vault alongside the documents themselves, so the chain of custody travels with the records rather than living in an email.
Sign-off is append-only
The attestation cannot be edited after the fact. If something is found to be wrong later, the correction creates a new version on top of the record; the original attestation stays exactly as it was signed. A record you can quietly rewrite is not a chain of custody.
What the certificate does not say
Related topics
Can’t find what you’re looking for?
Our support team is here to help. Contact us and we’ll get back to you as soon as possible.
Contact Support