Every AI deduction the scanner surfaces lands in the practitioner’s review queue with one covenant attached: nothing ever reaches a client return, an IRS transmission, or a tax-software field without an explicit accept from the preparer of record. Treasury Circular 230 §10.35 — competence, diligence, and appropriate supervision — stays with the practitioner; the model is research, never the last word. The six steps below are the working-paper sequence every engagement-level review closes against, with the print-and-take-to-engagement checklist at /checklist the same six steps in the same order.
1. Confirm the IRC §-citation is on the flag
A flagged deduction may include the IRC section and a link to IRS.gov or Cornell LII when the catalog record has that source. The reviewer verifies the authority and facts; when the engine cannot find a source, the item remains a research lead rather than a filing conclusion.
2. Confirm the preparer-imported source return is visible alongside the flag
The reviewer needs a line-for-line comparison, not the AI’s redacted view. The "Surfaces a single preparer-imported source return alongside the AI suggestion" capability exists to keep this step possible in the first place — without the source return visible on the same drawer, the AI suggestion is recommendation-without-ground-truth. Cross-tie the IRC § from step 1 to the destination line on the source return, the prior-year column, and the bank record. If any one of those three doesn’t coexist on a single working-paper trail, the audit posture fails.
3. Confirm the form/line + revision history on the engine output
The flag must carry its destination form/line on the return — Schedule C line 30 (other expenses), Form 8829 line 13 (utilities), Schedule 1 line 17 (SEHI), and so on — plus the prior-year comparison and the revision-history snapshots the engine surfaced. A deduction without a destination form is a recommendation without a return-side shelf; the revision history is what catches the prior-year anomaly the engagement should know about before the line ships. If the destination form isn’t on the same drawer as the flag, the suggestion stays on the reviewer queue until the question is answered.
4. Accept / modify / reject with a reason on reject
Every decision is an explicit click — nothing ever auto-applies; nothing ever reaches a tax-software field without a reviewer’s click first. Accept returns the line to the preparer’s review queue for one more look before save/submit. Modify lets the reviewer overwrite the suggestion inline. Reject is the load-bearing one: the reject-reason field is required, timestamped, and the "Re-surface a suggestion the reviewer explicitly rejected — Reject is final unless reopened by the same reviewer" guardrail means the reason string closes the suggestion with the audit-loop justification the engagement file needs.
5. Confirm the audit-log entry is written
Accept / modify / reject (with timestamp + reviewer identity + reason on reject) can land in the per-firm ledger. Retention, exportability, and sufficiency for an examiner depend on the configured policy and professional review; the ledger is not itself an audit defense.
6. Confirm the PTIN stays on the return
The preparer’s signature on the return is unchanged; the model never transmits to the IRS and never holds an EFIN. Treasury Circular 230 keeps the reviewer the preparer of record. The "Replace the preparer’s signature on the return" product guardrail is why this last step is non-negotiable on every engagement — no AI suggestion ships without a PTIN on the bottom of the page, and no PTIN ships without a human on the loop who clicked accept.
Use the printable checklist
Open the /checklist page in this browser. The same six steps, in the same order, with the checkboxes you can tick as you go. Ctrl / Cmd-P → Save as PDF and the page prints as a clean working-paper you can attach to the engagement file — the headlines, the rationale paragraphs, and the checkbox state all in one document. The page persists completion state in your browser, so a half-finished review picks up where you left off.
Where the scanner fits
The Taxerity.AI scanner surfaces exactly the same six-step checklist on the practitioner’s own engagement queue — pairing each AI flag with the IRC §-citation from step 1, the source-return drawer from step 2, the destination form/line from step 3, the accept/modify/reject click from step 4, the audit-log entry from step 5, and the PTIN-stays-on-the-line guardrail from step 6. It does not replace the practitioner-side review that has to close the loop — it surfaces the citation chain inline with the suggestion, so the six steps above are the natural reading order the reviewer carries into the engagement.
Closing
Solo CPAs who close an engagement against the same six-step sequence — IRC §-citation, source return, form/line + revision history, accept/modify/reject with reason, audit-log entry, PTIN-stays-on-the-line — keep Circular 230 §10.35 as the implicit default of every review, the same way the four-step canonical workflow on /about keeps the citation-grade contract as the default of every flag. Print /checklist, walk through the six steps, attach the signed checklist to the engagement file before going on the return.