Correct it once.
It never comes back.

How review works

Review Desk / learned from 1,140 decisions

Learned from your reviewers5 rules

Prior-year W-2 offered against a current-year requirement Request the current yearapplied 212× Automatic

Bank statement missing its final summary page Accept if the balance page is thereapplied 96× Automatic

Unlabelled scan that reads as a title commitment File as title, not as an exhibitapplied 341× Automatic

Gift letter arriving without a donor signature Hold pending signatureseen once Learning

Two appraisals, same property, different dates Keep the later oneseen once Learning

Always yours to sign3 rules

Any file above the $2M authority line Named underwriternever automatic Reserved

Exceptions to the flood-zone requirement Named underwriternever automatic Reserved

First file from a new correspondent Named underwriternever automatic Reserved

01 / INPUT

Bring the mess.

Do not prep. Do not sort. Do not clean up. Just point Sheaf at the raw intake and let the engine handle the reality of operational documents.

Scans

Phone Photos

Merged Packets

Application Dumps

Wrong-order Pages

Junk Pages

Dense Tables

Signatures

02 / PROCESS

Learning your desk.

Every correction your reviewers make is a decision Sheaf keeps, and then applies. The queue gets shorter each week, and the rules it runs on are yours rather than an industry average.

Queue

Only what needs you.

Work reaches a person for one of two reasons: the evidence fell short of your threshold, or your policy puts a signature on this class of decision. Everything else is already filed.

Decide

One call, fully recorded.

Each decision keeps the actor, the pages it covered, and what the engine had proposed before anyone touched it — so the disagreements are visible, not just the outcomes.

Learn

The correction becomes a rule.

When the same situation resolves the same way twice, it stops reaching the queue. Sheaf applies the decision from then on, and cites the two that created it.

Diverge

Your shop, not the average.

What is learned here is scoped to your organization. Nothing your reviewers decide leaves it, and nothing from anyone else’s desk arrives in yours.

Technical detail How review works, and what it takes to promote a decision to a rule Read the spec →

03 / OUTPUT

Structured payload.

The file doesn't need to arrive perfect, but the data leaves that way. Ready for your existing infrastructure.

The queue comes back as work, not as a dashboard. What needed a person, what a learned rule already settled, and which decisions it was learned from.

Every automatic resolution points back at the human ones behind it, so an auditor can walk from the outcome to the people without leaving the record.

{
  "review_id": "rev_8849-2A",
  "queued_for_human": 2,
  "settled_by_rule": 11,
  "rules": [
    { "id": "prior_year_w2", "learned_from": 2 }
  ],
  "named_signer_required": true
}

How review works

Every operation draws its lines somewhere slightly different. Review is where Sheaf learns where yours are — and then stops asking about them.

The queue is the training signal.
A decision made twice stops being a question.

Review is a queue, not a stage.

Nothing waits for a human by default. Work reaches the queue for one of exactly two reasons: the evidence came in short of the threshold you set, or your policy says a person signs this class of decision regardless of how confident anything is. Everything else is filed before anyone opens the file. The queue is what is left over, not what the work has to pass through.

A decision is recorded with everything it rested on.

Not just the outcome. Each decision keeps four things:

Actor

The person who decided, or the workflow that decided under a named policy. Both are recorded the same way; neither is anonymous.

Proposal

What the engine had recommended before anyone touched it.

Scope

The exact pages the decision applied to, so it survives the document being re-cut later.

Departure

Whether the human agreed with the machine, and precisely where they didn’t.

That last one is the valuable one. A decision that matched the recommendation confirms the engine. A decision that overrode it teaches the engine something, and it is the overrides that shape what happens next.

The second identical decision stops being a question.

Observed. A reviewer settles something the engine was not sure about.

Corroborated. The same situation is settled the same way again.

Promoted. It stops reaching the queue. Sheaf applies the decision and cites the two that created it.

Retired. A later contradiction takes the rule back out, and the question returns to the queue.

Promotion needs agreement, not volume. One reviewer clicking through the same thing forty times is a single opinion repeated forty times; two reviewers arriving independently at the same answer is a policy. Only the second one promotes.

Your desk, not the average desk.

Two lenders can read the same document and reach opposite conclusions, and both can be right, because they are underwriting to different policies. So what is learned here is learned for you. It is scoped to your organization and never pooled into a shared model — nothing your reviewers decide can surface in a competitor’s queue, and nothing from theirs arrives in yours.

A rule comes back out as easily as it went in.

Every automatic decision points at the human decisions that produced it. Open one and you get the reviewers, the dates, and the files they were looking at. Retire it and the question returns on the next file, with the history kept rather than deleted — because the reason a rule was wrong is worth as much as the rule was.

Automation is a setting, not a personality. Every class of decision has a mode — always ask, ask until learned, or never ask — and the mode is yours to set.

Correct it once.It never comes back.