── ── Tax & compliance
Reading a Return Top to Bottom Finds the Errors You Already Suspected
August 17, 2026 · 7 min read · By Brad Ju
Different error classes hide from different kinds of attention, so a pre-delivery review works better as five separate passes with five different questions than as one careful read. Identity, prior-year comparison, internal consistency, source traceability, and reasonableness — in that order.
The errors that reach clients are almost never exotic. They're a carryover nobody re-checked, a state that quietly dropped off, a name spelled two ways, an election made last year and forgotten this year, and a number that is technically correct but obviously wrong sitting next to last year's.
None of those is hard to catch. They survive review because of how review is usually conducted: one careful read of the return from top to bottom, hunting for mistakes. That method has a structural weakness — it finds what the reviewer already suspects, and very little else.
Why one pass can't work
Different error classes hide from different kinds of attention. Checking whether a schedule ties out and noticing that last year's rental has vanished are not the same cognitive act, and a reviewer doing the first is not doing the second. So the review is five separate passes with five different questions, run in order. This is the second of four CPA/EA skills we published as open source, and the sweep structure is the whole of it.
| Sweep | The question it asks | What only it catches |
|---|---|---|
| 1 · Identity and posture | Is this the right taxpayer, year, status, and are the required signatures and authorizations in place? | The embarrassing errors, plus due-diligence gates that carry a penalty on their own |
| 2 · Prior-year comparison | What changed since last year, and why? | Disappearing line items and untraced carryovers |
| 3 · Internal consistency | Does the return agree with itself? | Schedules that don't tie, figures that differ between two places |
| 4 · Source traceability | For each material number, can I point at the document? | Numbers whose only provenance is that the software brought them forward |
| 5 · Reasonableness | Does the whole picture match this client's actual year? | The correct-looking wrong answer |
Sweep 2 is the one that isn't optional
Open last year beside this one and require a one-sentence explanation for every material change. Every line present last year and absent this year — where did it go? Account closed, property sold, or oversight? Every new line — what happened? Every line that moved materially — why?
There's a gate worth stating plainly: a disappearing line item outranks a wrong number. Wrong numbers get caught by review. Missing items get caught by nobody, because there is nothing on the page to look at. This is also where carryovers get traced from the prior return rather than retyped, where prior-year elections get confirmed as still in effect and still appropriate, and where the same set of state filings as last year gets verified — one added or dropped needs a reason on the record.
On an 1120-S, the entity-to-owner handoff belongs here too, and it's the most commonly missed item in the whole review: where a shareholder claims a loss, takes a distribution, disposes of stock, or receives a loan repayment, the basis form belongs on that shareholder's individual return. Each preparer assumes the other one did it.
Sweep 4 has a gate people argue with
For each material number, point at the document. If you can't, that is the finding — not a task for later. The gate: if a number's only source is that the software brought it forward, it has not been verified this year. Carryforward is a convenience, not evidence.
The specific trap in this sweep is payment-settlement information returns. They get reconciled to gross receipts, not added on top. Double-counting there is common, produces a plausible-looking return, and the client will not catch it.
Sweep 5 is the one that gets skipped for being soft
Look at the whole return the way a stranger would. Does the picture match what you know about this client's year? Does income support their evident circumstances? Large round numbers where you'd expect precision? Anything conspicuously out of proportion? And the question that compresses the entire sweep into one line: would you be comfortable explaining every material position out loud to someone who asked why?
Any "no" stops delivery until it's resolved. It reads as subjective, and it is judgment — which is what a reviewer is for. But the test is operational, not aesthetic: either you can say it out loud or you can't.
Two sweeps, one error
An 1120-S in its second year, prepared by an experienced preparer. Sweeps 1, 3, and 4 came back clean. Sweep 2 flagged a repairs-and-maintenance line at roughly six times prior year with no explanation — not wrong on its face, since the client had bought a building. Sweep 5 then asked the stranger question, and the work turned out to include a roof replacement: a capital item sitting in a deduction line.
Neither sweep alone gets there. The comparison sweep finds the anomaly; the reasonableness sweep refuses to accept the first story that fits it. That's the argument for keeping them as separate passes instead of merging them into general care.
Write the findings down
A silent fix teaches the preparer nothing and leaves no record. Each finding gets a location, a severity — blocker, correction, query, or note — and a disposition. Nothing gets filed with an open blocker or correction; a query either resolves or converts into a documented acceptance with a stated rationale. And the note items are free planning: they're next year's chase list, collected while the context is still in your head.
Use it or fork it
Pre-Delivery Return Review is one of four CPA/EA practice skills we open-sourced under MIT at github.com/deciqAI/knowledge-skills — copy the folder into .claude/skills/, or paste the Markdown into any model. It covers 1040 and 1120-S, and it deliberately contains no thresholds, rates, or dates: every one of them expires, and a remembered figure is the failure mode the review exists to catch. Educational content — not tax advice.
FAQ
Why review a tax return in five passes instead of one careful read?
Because different error classes hide from different kinds of attention. Checking that a schedule ties and noticing that last year's rental disappeared are different cognitive acts, and a single top-to-bottom read mostly confirms what the reviewer already suspected.
Which review sweep catches the most errors?
The prior-year comparison. It's the only pass where last year's return does the work for you, and it's the one that surfaces disappearing line items — which no other check can find, since there's nothing on the page to look at.
What should a reviewer do when a number's only source is the software's carryforward?
Treat it as unverified. Carryforward is a convenience, not evidence; carryovers should be traced from the prior return rather than accepted because they appeared.
Is the reasonableness sweep too subjective to be useful?
It's judgment, which is what a reviewer is for, and the test is operational: can you explain every material position out loud to someone who asks why? It's the only sweep that catches a return that is internally perfect and wrong at the source.
Should review findings be logged or just fixed?
Logged, with location, severity, and disposition. A silent fix teaches the preparer nothing and leaves no record, and the lower-severity findings are next season's planning and document-request list.