GPT 6 document reconciliation with visible evidence

Traceability in GPT 6 document reconciliation

GPT 6 document reconciliation has a published financial review case. OpenAI’s September 3 Legora story describes a tie-out across 41 documents in one agent run. Legora reported finding all four planted errors and nearly 40% improvement on that financial-statement workflow; the reported improvement across its broader benchmark was about 3%. These different measures should not be conflated or treated as results for your documents. The process here is original guidance and has not been run against a live GPT-6 system. It is designed to preserve differences for a person to decide, especially when values, dates, units, or footnotes affect the meaning.

Begin by naming the documents, versions, dates, and intended purpose. Choose a source of truth for each field or declare that no source of truth exists. Never ask for a seamless merge as the first operation. A neat merge can hide a conflict that should have gone to a reviewer.

Write a GPT 6 document reconciliation prompt

Compare [DOCUMENT A] and [DOCUMENT B] for [PURPOSE]. Align equivalent sections and fields. Return additions, deletions, changed values, changed wording, moved content, and unresolved conflicts. For each item include the location in both documents, a neutral paraphrase, materiality, and a suggested reviewer. Preserve units, dates, footnotes, and qualifiers. Do not infer missing values. Label assumptions and unknowns.

If the documents are long, reconcile by section and retain stable identifiers. Ask for an intermediate field map before asking for a final report. The map should show how a field in one document corresponds to a field in the other. When correspondence is uncertain, use “possible match” and send it to review.

Illustrative worked example

This example is fictional and is not a financial conclusion. Input A says “Revenue, Q2: $100,000, three months ended June 30”; Input B says “Revenue, six months ended June 30: $190,000.” A useful output does not merge the figures. It records a possible period mismatch, preserves both values and locations, and asks the finance owner whether the comparison is quarter to quarter or year to date. Acceptance means the ledger shows both periods, units, source locations, materiality as pending, and a named reviewer. The model’s explanation remains draft language until that owner confirms the comparison basis.

Evidence ledger

Use columns for document, version, location, field, old value, new value, status, materiality, source note, and reviewer decision. “No change found” is different from “not checked.” Record the pages or headings inspected. Keep an audit copy of the inputs under the approved retention policy.

Numbers deserve a separate pass. Check sign, currency, scale, rounding, period, and whether a number is a subtotal or a total. Text deserves a meaning pass: a changed qualifier can alter an obligation even when the headline sentence looks similar. Footnotes and tables can carry the decisive context.

Failure modes

A model may align fields by similar names while missing a unit change. It may treat a deleted paragraph as irrelevant or paraphrase a caveat away. It may also produce a confident materiality label without the business context needed to justify it. Require locations and evidence, then make materiality a human decision when the consequence is significant.

Do not use a customer story as a universal performance claim. The cited Legora article reports one organization’s experience. It does not promise identical accuracy, speed, or coverage for another document set.

Escalation in GPT 6 document reconciliation

Review every high-materiality change against the source document. Have the accountable finance or legal owner decide conflicts involving recognition, disclosure, obligation, or regulatory language. Test a sample of low-materiality changes for false negatives. If the model cannot locate a passage, mark the item unresolved.

The final report should contain a decision summary, evidence ledger, unresolved list, and reviewer sign-off. Keep generated prose separate from approved language until the owner accepts it. The document formatting example can follow this stage without changing the substantive record.

Reusable stopping rule

Stop when every in-scope section has a checked status, all material conflicts have an owner, and unknowns are visible. More rewriting cannot resolve a missing source. Return to the document owner or obtain the missing version. The GPT 6 guide gives a broader provenance checklist.

Preserve the audit trail

Keep original files read only where possible and assign stable names to the comparison run. Record extraction method, page count, detected tables, and pages that could not be read. Scanned pages, charts, handwriting, and unusual encodings may need a separate review. Do not report “no difference” when a section was unreadable.

Use a two pass review for consequential numbers. The first pass finds candidate changes. The second checks each candidate against the rendered page and, where needed, an independent calculation or system of record. Have the accountable finance or legal owner decide materiality. If reviewers disagree, preserve both views and escalate instead of averaging them.

After approval, identify the output as a reconciled version. Include source versions, comparison date, reviewer, and unresolved items. When a later version arrives, start a new run or document the delta from the approved baseline. This keeps a model assisted comparison from becoming an unexplained overwrite.

Record GPT 6 document reconciliation decisions

Not every difference requires the same response. A typographic change may be low materiality, while a changed period, unit, or obligation may require an accountable owner. Define categories before reviewing: accept source A, accept source B, preserve both pending evidence, or escalate. Let the reviewer choose the category and record the reason. Do not let a model convert a missing decision into a merged sentence.

If the documents come from different systems, confirm that extraction has preserved tables, signs, and page references. Keep a sample of source and extracted text for difficult sections. When a reviewer corrects an alignment, update the mapping record so the same error is easier to spot later. The purpose of automation is a traceable first pass that leaves the final judgment visible. For a team handoff, include the source register, field mapping, unresolved list, and the person authorized to decide each conflict. Keep a sample of accepted changes and rejected changes to calibrate later runs. If a source is revised, preserve the previous comparison and create a new dated record. This provides a defensible history without pretending that the model made the decision. Keep the prompt and evidence ledger together. When the model proposes an alignment, store the source locations that justify it. When a reviewer rejects it, record the reason and the correct mapping. Over time this creates a clearer review protocol while keeping each conclusion tied to the documents that were actually available. Finally, ask whether the reconciliation answered the business question it was created for. If the ledger is complete but the owner still cannot decide, identify the missing context. If the answer is concise but cannot be traced to a source, return it for revision. A useful comparison reduces uncertainty while keeping its remaining uncertainty visible. The point of reconciliation is a defensible decision. Preserve the evidence even when the result is that no change should be made. A reviewer should be able to see what was compared, what was unreadable, which conflicts were escalated, and who accepted the final status. Before delivery, read the summary without the model context and ask whether a new reviewer can find the source for every material conclusion. If not, improve the evidence ledger or shorten the claim. Traceability should survive the handoff. That discipline makes the result useful beyond one review. A later owner can understand the baseline, reproduce the comparison, and see exactly which items still need judgment.

A concrete GPT 6 document reconciliation run

Input: two dated statements, a trial balance, a consolidation schedule, and the expected reporting period. Ask GPT 6 document reconciliation to produce a field map before comparing values. Each row should contain document A location, document B location, normalized label, value, unit, period, and match status. Keep a separate row for footnotes and qualifiers.

Output review: sample ten rows against the rendered documents, recalculate one total independently, and inspect every unresolved or high-materiality difference. The GPT 6 document reconciliation result should distinguish added, deleted, changed, moved, unreadable, and possible-match items. A finance or legal owner decides materiality; the model supplies traceable candidates.

If extraction loses a table or sign, mark affected rows unverified and rerun that section with a better source representation. If two values agree but units differ, preserve the conflict. This is the practical value of GPT 6 document reconciliation: it reduces the search space while keeping the final decision attached to evidence.