Forward it. Drop it. Photograph it.
It arrives understood.

No, and the order of operations is the reason. A free rules classifier runs first, then a deterministic ladder tries five rungs for each expected field, structured metadata paths, labelled templates, regular expressions, validating parsers, and rules learned from your own corrections. A language model is only asked about the fields none of those resolved, and it is asked once, in a single batched call. A document that is fully machine readable can be processed with no model call at all, and re-processing a document never buys the same facts twice. The design point is that reading every document has to be cheap enough to do to every document, not just the important ones.

The revision question is answered at the moment the document is filed, against the drawings already on the job. If the bytes are identical to something already there, it is a true duplicate and chains under the original. If the new drawing orders as newer than the one on file, it supersedes it, and the superseded version is retained rather than deleted. If an older drawing arrives late, it files itself into the chain as superseded instead of unseating the current one. If two carry the same revision label, that is flagged as a possible duplicate rather than archived, because two halves of a set can legitimately share a label. Anything the comparison cannot order stays a suggestion for a person to settle. Auto-action only fires on a strong drawing-number or reference match, never on a similar title alone, and a council-stamped approval set is treated as an approval artefact that neither supersedes nor is superseded.

Amber is a candidate, a value Oryn extracted that is waiting on a person. Green is committed, a value a human accepted. Committing a fact writes it into the job record, records the provenance of the change including where it came from and who accepted it, and flips the fact to committed. The rule underneath it is that fields carrying money or a legal identifier stay candidates structurally, regardless of how confident the extraction was. Confidence describes how cleanly a value was read, not whether it is correct, so it is never treated as authorisation. A high confidence score on a wrongly located approval number is exactly the failure this design exists to catch.

Naming follows a convention that starts with the job number, then the document label, then the revision where a revision is meaningful, then the document date. The job number is read from your own database rather than extracted from the page, so the part of the name that has to be right never depends on a read. Renaming only happens while the document still carries the filename it was uploaded with, so a name you chose is never overwritten. Filing goes to the folder the document type belongs in, within a standard project folder structure covering estimating, plans, specifications, engineering, supplier quotes, takeoffs, contracts, approvals, selections, construction, variations, claims, invoices, compliance and handover. If you have moved that kind of document before, your correction becomes the preference and beats the default.

Yes. A review is a specific document at a specific version, assigned to a named reviewer with a due date, and the project manager is pre-selected as the default reviewer. Each reviewer gets their own review and their own notification. The states are deliberately few, requested, then approved, approved as noted, rejected or cancelled, and nothing moves twice. A resubmission is a new review on the new version rather than a reopened old one, which keeps the history honest. Every transition writes an append-only event. If a review sits unanswered past your chase setting, Oryn chases it automatically, at most once a day, and never once a decision has been made.

Neither, and that is deliberate. Search runs over the words actually in your documents, including scanned pages that were transcribed, plus the document names and Oryn’s own summaries and tags. Results come back with the matching snippet and the page it sits on. What you get is recall grounded in your own material rather than a similarity score you cannot inspect, and when nothing relevant is on file the honest answer is that nothing is on file.

Yes. Automatic filing and automatic superseding are governed by the same organisation-level setting, and turning it off leaves Oryn suggesting rather than acting while the reading and extraction continue as normal. Everything the automation does is reversible in any case. Filing a document to the wrong project is fixed by reassigning it, and superseding a drawing flips a status and adds a forward link rather than removing anything, so the original document and every version of it remain.

It banks them, and a person promotes them. When you override a value Oryn extracted, the correction is recorded privately to your organisation. Nothing reads that bank automatically. An administrator can see which fields are being corrected most, write a rule against the real corrected examples, test it against them, and promote it only once it demonstrably extracts the right thing. A rule that matches none of the banked examples is refused. The pipeline only ever improves through a promoted rule, which means learning cannot quietly change what your documents say. Filing preferences work the same way, the folder you moved a document kind into becomes the preference for that kind next time.