Skip to content

Systems and records

Reconciliation correction history dataset: why the fixes teach AI

By SourceX Editorial · Updated

10 sections

Short answer

A balanced ledger shows only the end state. The correction history shows how a finance team found a difference, diagnosed its cause, chose a fix, got it approved and confirmed later that the fix held. Linked to source documents and outcomes, that path is the evidence a finance AI model needs to learn reconciliation judgment; a balanced period with no linked corrections mostly teaches that the numbers eventually matched.

Record timeline

Reconciliation correction

This page covers reconciliation correction history dataset: why the fixes teach AI.Illustrative

Key takeaways

  • Corrections carry the reasoning that a balanced ledger hides: what was wrong, how it was found and what fixed it.
  • A useful correction record links the starting difference, the evidence, the rule applied, the correcting entries, the approver and a later outcome check.
  • A posted status is not an outcome; the outcome is whether the next reconciliation confirmed the fix.
  • Corrections that were themselves reversed, or cleared to a suspense account, need their own labels rather than counting as resolved.
  • Accounting systems already keep much of the lineage, such as reversal entries, corrective credit memos and change logs, but rarely link it into one case.
  • Bank account numbers, counterparty names and staff identities are removed before any history leaves the company, while the structure is kept.

Why might reconciliation corrections be more useful evidence than a balanced ledger?

Because the correction is where the judgment happened. A ledger that ties to the bank statement at period end records an outcome; it does not record that an accountant noticed one unmatched entry, traced it to a payment posted twice, reversed the duplicate and checked that the next statement agreed. A model trained only on balanced periods sees answers without the work that produced them.

Reconciliation corrections also come with a built-in test. A correction makes a claim about what caused a difference, and later records either confirm that claim or contradict it. That gives each case a verifiable outcome, which is exactly what training and evaluation data for finance agents needs and what most operational records lack.

Keep the scope precise. Most reconciliation corrections are routine in-period fixes, not the formal error corrections of accounting standards. FASB's SFAS 154, now codified as ASC 250, defines an error in previously issued financial statements and treats its correction as a restatement, separate from accounting changes. A dataset should label which of the two each case is, because a restatement involves a different approval chain and different evidence.

Correction cases are also where finance processes differ most between companies. Two businesses on the same ERP may reconcile identical bank feeds, but how they investigate a difference, which threshold triggers an approval and when they accept a timing explanation reflect their own policies. Those differences are what a general model cannot infer from accounting textbooks and what a correction history records.

What does a reconciliation correction look like in the source system?

It is spread across several record types that the system keeps but does not join. Microsoft Dynamics 365 Business Central, whose documentation is public, is a useful reference because it shows each piece clearly; other ERP and accounting systems have equivalents under different names.

At a glance

What does a reconciliation correction look like in the source system

The starting state is the reconciliation itself. In Business Central, lines that cannot be matched carry a value in the Difference field and remain on the Bank Acc. Reconciliation page after posting, as discrepancies to resolve. Microsoft's documentation lists typical causes: a ledger transaction missing from the bank statement, a bank transaction with no posting in the system, and matching transactions whose details differ too much to match automatically.

The fix leaves its own records. A reversing entry in Business Central is the same as the original with the opposite sign, keeps the same document number and posting date, can be made only once, and must be followed by the correct entry. Cancelling a posted sales invoice creates and posts a corrective credit memo that stays in the audit trail. Each of these is a link in the lineage, if the dataset keeps the document numbers that tie them together.

  • Reconciliation record: statement date, statement ending balance, matched and unmatched lines, difference values.
  • Ledger entries: the original posting, any reversing entry and the replacement entry, linked by document number.
  • Corrective documents: credit memos or cancellations that reference the original invoice.
  • Change log entries: who changed which field, when, with old and new values.
  • Supporting evidence: bank statement lines, remittance advice, vendor statements, email approvals.
  • Next-period reconciliation: the record that confirms or contradicts the fix.

What should the correction lineage and outcome labeling example include?

One case packet per correction, with ten fields that connect the difference to its verified outcome. Each field should point to a record identifier rather than restate the record, so a reviewer can open the source and check it.

Business Central's change log is a good model for the actor and timing fields: it records the user ID, date and time, table, field, type of change, and the old and new values for tracked fields. Where a system does not log a field, record the gap rather than inferring the value.

What should the correction lineage and outcome labeling example include?
FieldWhat it holdsTypical source recordRequired?
case_idStable identifier for the correction caseAssigned during extractionYes
starting_stateAccount, period, unmatched lines and difference type at discoveryReconciliation record and its unmatched linesYes
evidence_refsIdentifiers of documents consultedBank statement lines, invoices, remittances, vendor statementsYes
rule_appliedThe policy or matching rule that decided the fixClose checklist, accounting policy, matching rule setupYes, or marked unknown
diagnosisThe cause the preparer recordedReconciliation notes or journal descriptionYes
correction_entriesDocument numbers of reversing, replacement or corrective entriesGeneral ledger, credit memosYes
actorsPseudonymous preparer and approver IDs with timestampsChange log, approval workflowYes
approval_authorityWhether the approver held authority for this account and size bandApproval matrixRecommended
outcome_checkThe later record that confirms or contradicts the fixNext reconciliation, later reversalsYes
outcome_labelOne label from the outcome set, with evidence referenceDerived during labelingYes

Worked case: an unmatched ledger entry traced to a duplicate posting

This case is illustrative; the steps follow documented system behavior, but the account, counts and durations are invented. It shows what one complete packet looks like and where each label comes from. Amounts are described by relationship rather than value.

Notice what the packet adds to the raw records. The reversal alone shows that something was undone; the packet shows why it was undone, which rule required it, who approved it and that the following month confirmed the diagnosis. That chain is what lets a model learn to recognize a duplicate payment from the pattern of an unmatched ledger entry, a second journal batch and a vendor statement showing one invoice.

The same case can also be turned into an evaluation task. Give a model the starting state and the evidence records, hide the correction and the outcome, and grade whether it proposes reversing the duplicate payment. Because the outcome is verified, the grade does not depend on a reviewer's opinion of what the model should have done.

Worked case: an unmatched ledger entry traced to a duplicate posting
StepRecordWhat it showsPacket field
1. DiscoveryBank reconciliation for the operating account, month 3One bank ledger entry has no matching statement line; the difference equals one vendor paymentstarting_state
2. InvestigationVendor statement and payment journalThe same vendor invoice was paid in two journal batches 2 days apart; only one cleared the bankevidence_refs, diagnosis
3. RuleClose checklist item on duplicate paymentsUnsent duplicate payments are reversed, not left openrule_applied
4. FixReversing entry on the duplicate paymentSame document number and posting date, opposite signcorrection_entries
5. ApprovalApproval record for the reversalController approved 1 day after discoveryactors, approval_authority
6. OutcomeBank reconciliation, month 4Posted with no unmatched lines for this vendor; the reversal was not itself reversed within 60 daysoutcome_check, outcome_label: resolved_verified

Which outcome labels can a reviewer trust?

Only labels backed by a later record. A correction marked posted, closed or approved tells you the workflow ended, not that the diagnosis was right. The outcome label has to come from something that happened after the fix, usually the next reconciliation or the absence of a later reversal within a stated window.

At a glance

Which outcome labels can a reviewer trust

Use a small, fixed label set and record the evidence for each label in the packet. Six labels cover most cases without forcing reviewers into fine distinctions they cannot verify.

State the window once for the whole dataset and apply it to every case. A correction labeled resolved_verified on the strength of one later reconciliation and another labeled on three later reconciliations are not the same label. If cases near the extraction cutoff cannot meet the window, label them resolved_unverified rather than shortening the window for them.

  • resolved_verified: the next reconciliation agrees and the correction was not reversed within the stated window.
  • resolved_unverified: the difference cleared, but no later record confirms the diagnosis, for example the period has not closed yet.
  • reversed_again: the correction was itself reversed or replaced, which usually means the first diagnosis was wrong; keep both cases linked.
  • cleared_to_suspense: the difference was posted to a suspense or miscellaneous account without a documented cause.
  • open_at_cutoff: the difference was still unresolved at the extraction date.
  • restatement: the correction concerned previously issued financial statements and followed the formal error correction process.

Counterexample: a cleared difference that teaches the wrong lesson

A difference can disappear without anyone learning why. In this illustrative counterexample, three small unmatched statement lines on a card account are posted to a general expense account during a busy close, the reconciliation then balances, and the lines are never investigated. Labeled as resolved, the case would teach a model that posting unexplained differences to expense is a correct fix.

The same risk appears with AI-assisted matching. Microsoft's documentation for Business Central's Copilot bank reconciliation assist says its matches and suggested accounts can be incorrect or incomplete and that users must review them before keeping them. A history that includes such suggestions should record whether a person reviewed each one and whether a later record confirmed it; otherwise the dataset can quietly teach one model the earlier mistakes of another.

The right label here is cleared_to_suspense, or resolved_unverified where the account used was a real expense account but no evidence supports the choice. Keep these cases in the dataset with their labels: negative and uncertain examples are valuable, as long as they are not presented as correct work.

What should an independent reviewer check in each correction packet?

They should be able to, and the packet is built so they can do it without asking the finance team. A reviewer checks three things: that every evidence reference opens a real document that supports the diagnosis, that the approver had authority for that account and size band when they approved, and that the later financial state matches what the correction claimed.

At a glance

What should an independent reviewer check in each correction packet

Audit documentation standards set a useful bar. PCAOB AS 1215 requires auditors' documentation to show who performed the work and when, and who reviewed it and when. A correction packet that cannot answer those four questions for its preparer and approver is incomplete, even if the numbers tie.

  • Documents: every evidence_refs entry resolves to a record in the extract, and the diagnosis follows from them.
  • Authority: the approver appears in the approval matrix for that account and band on the approval date.
  • Final state: the outcome_check record exists, falls inside the stated window and agrees with the outcome_label.
  • Lineage: the correction_entries chain has no gaps, and any later reversal links back to this case.

Which record proves each type of fix, and which proves it held?

Different fixes leave different proof, so the packet should name the proving record by correction type. The decision record shows what the team chose; the outcome record shows whether the choice was right. Both are needed, and they are rarely the same record.

Which record proves each type of fix, and which proves it held?
Correction typeRecord that proves the decisionRecord that proves the outcomeCommon gap
Duplicate postingReversing entry linked to the original document numberNext reconciliation with no unmatched line for itReversal description does not cite the duplicate
Missing bank fee or interestNew journal entry citing the statement lineFee or interest recurring correctly in later periodsEntry posted without the statement reference
Wrong customer or vendor appliedUnapply and reapply entriesCustomer or vendor statement agreeing after the changeNo record of who requested the reapplication
Invoice errorCorrective credit memo applied to the original invoiceOpen balance for that invoice reaching zeroCredit memo not applied, left open
Timing differenceReconciliation note marking it as in transitItem clearing the bank in the following periodItem still open several periods later

What should be removed before a correction history leaves the company?

Remove anything that identifies a bank account, a counterparty or a person, and keep the structure that makes the case useful. Bank account numbers, card numbers, vendor and customer names, remittance details and staff names should be replaced with stable codes so links between records survive.

Keep relationships rather than raw values where values are sensitive: whether the difference equaled one payment, whether two postings were days apart, which approval band applied. Free-text reconciliation notes are often the richest field and the riskiest; review a sample for names, account numbers and personal details before deciding to include them.

Rights are a separate question from preparation. Removing identifiers addresses privacy, but whether the company may license records that include customer and vendor transactions depends on its contracts, which counsel should review.

Check the change log extract separately. Old and new values in tracked fields can contain bank details or names that were overwritten in the live record, so a field that looks clean in the current table may still appear in its history.

Where correction histories fit in a SourceX review

SourceX helps companies license selected business records as AI datasets and handles rights review, contracting, preparation, delivery and payment. Correction histories carry the most value when they span several years of one accounting system with enough cases to show recurring patterns; as a general guide, buyers look for two or more years of consistent history and tens of thousands of records, though smaller sets can qualify when the work is rare, and each deal is assessed on its own.

Packets do not need to exist before that conversation; qualification is judged from a short company assessment without any data.

Frequently asked questions

Is a reconciliation correction the same as an accounting error correction?

Not usually. Most reconciliation corrections are routine in-period fixes such as reversing a duplicate payment or recording a bank fee. A correction of an error in previously issued financial statements is a restatement under ASC 250 and follows a different process, so a dataset should label the two separately.

Which systems keep enough history to build correction packets?

Many ERP and accounting systems keep reversing entries, corrective documents and reconciliation records, and many offer a change log for selected fields. What they rarely do is link these into one case, so extraction usually means joining records by document number and date.

How long a window should the outcome check use?

Long enough to include at least the next completed reconciliation for the account, and long enough to catch a later reversal of the correction. Teams often state a window such as one or two closes; the key is to state it in the packet and apply it consistently.

Should corrections made by AI matching tools be included?

They can be, if the packet records that the match or suggested account came from a tool, whether a person reviewed it and what the later outcome was. Without those fields, tool-generated fixes are hard to distinguish from reviewed human judgment.

Are cases that were never resolved still useful?

Yes, when labeled honestly. Open, reversed-again and cleared-to-suspense cases show where reconciliation is hard, which is useful for evaluation. They should never be labeled as correct fixes.

Sources

  • Business Central keeps unmatched reconciliation lines with a Difference value on the reconciliation page after posting and lists typical causes of differences. Source
  • A Business Central reversing entry mirrors the original with the opposite sign, keeps its document number and posting date, can be made once and must be followed by the correct entry. Source
  • Cancelling a posted sales invoice in Business Central creates and posts a corrective credit memo that remains in the audit trail. Source
  • The Business Central change log records user ID, date and time, table, field, type of change, and old and new values for tracked fields. Source
  • Microsoft says Copilot bank reconciliation matches and suggestions can be incorrect or incomplete and must be reviewed by users before they are kept. Source
  • SFAS 154, codified as ASC 250, defines an error in previously issued financial statements and treats its correction as a restatement rather than an accounting change. Source
  • PCAOB AS 1215 requires audit documentation to show who performed the work and when, and who reviewed it and when. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies