Skip to content

AI uses for records

Build a reconciliation explanation dataset that catches forced balances

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

A reconciliation explanation dataset should grade every reconciling item on four separate questions: does the arithmetic tie, which source record supports it, is the accounting treatment right, and who approved it. A zero difference proves little on its own. GnuCash, Firefly III, hledger, Beancount and QuickBooks Online can each turn the remaining gap into an entry, so flag every forced item and never label it explained.

Agent run

Reconciliation explanation

This page covers build a reconciliation explanation dataset that catches forced balances.Illustrative

Key takeaways

  • Arithmetic, evidence, treatment and authorization are four grades; a pass on one never offsets a fail on another.
  • Five accounting tools can book a remaining difference automatically or in one step, so a zero difference may be a plug.
  • Each forced-balance mechanism leaves a searchable marker; count the hits and route them to explanation review.
  • Federal guidance asks for adjustments traceable to documents, and GAO still flags intragovernmental balances after differences shrank.
  • Store one row per reconciling item with evidence, treatment basis, three role tokens and exact timestamps.

What does a reconciliation explanation dataset grade that a total cannot?

Four things for every reconciling item: whether the arithmetic ties, which source record supports the item, whether the accounting treatment is appropriate, and who authorized any adjustment. A total that ties answers only the first question. The other three are what a finance model has to learn, and they are exactly what a forced balance leaves blank.

Auditors have long treated adjustments as the place where both judgment and misconduct hide. PCAOB AS 2401, the auditing standard on fraud, explains that material misstatements often run through the financial reporting process, either as inappropriate or unauthorized journal entries or as adjustments that never become formal entries, such as consolidating adjustments, report combinations and reclassifications. It directs auditors to test journal entries and other adjustments and lists traits of suspect ones, including postings to unrelated, unusual or seldom-used accounts. A dataset that keeps only reconciled balances throws away the records that standard tells auditors to examine.

This page answers a narrower question than the live guides on correction histories and variance commentary: how to tell an explained difference from a forced one when building a reconciliation explanation dataset. It reads five accounting tools for the mechanisms that make a gap vanish, then gives a casebook of tied totals that still fail, a four-grade reviewer rubric, item-level fields, and the request fields for sourcing such cases through SourceX.

Five ways accounting software makes a difference disappear

Each of five accounting tools, from desktop bookkeeping software to plain-text ledgers, can turn a remaining difference into an entry, automatically or with one click, and none of them asks why. We read the four open-source tools at fixed commits on October 27, 2026, and Intuit's help pages for the fifth; a short script prints the matching source lines so anyone can repeat the reading from the named commits.

GnuCash balances an unbalanced transaction during its scrub routine by adding a split to an account named Imbalance followed by the currency code, creating that account if it does not exist; splits with no account at all are moved to a matching Orphan account. The scrub sets the split's value and writes no memo of its own. Firefly III's reconciliation screen offers to clear the selected transactions and create a correction adding or removing the difference. The correction is stored as a reconciliation-type transaction dated at the end of the period, titled Reconciliation with the period's dates and marked reconciled, and the screen adds that a correction can always be edited or deleted later.

The plain-text tools do it through syntax. In hledger, a balance assignment leaves out a posting amount, which is then calculated so the balance assertion holds; the manual's own example records untracked spending as a generic expense, and it warns that forcing balances this way can hide errors and makes data less trustworthy in an audit. Beancount's pad directive inserts a transaction carrying the P flag and a narration stating the padded balance and the difference, so that the next balance check passes. That check compares the total for one currency, not individual positions, and no padding is inserted when the gap is within a tolerance. QuickBooks Online lets a user finish a reconciliation that still shows a difference, creating an adjusting entry for that amount, and Intuit advises speaking to an accountant first.

What each mechanism writes down, and what it leaves for a person to supply:

Five ways accounting software makes a difference disappear
ToolMechanismWhat it recordsWhat it leaves blank
GnuCashScrub adds a balancing splitA split in an Imbalance account named for the currencyCause, supporting document, approver
Firefly IIICorrection during reconciliationA reconciliation-type transaction for the difference, dated at period endReason, and who accepted the gap
hledgerBalance assignmentA posting amount computed to satisfy the assertionThe amount itself until a report runs, and any cause
Beancountpad directiveA P-flagged transaction naming the balance and the differenceWhy the gap existed, and any approval
QuickBooks OnlineFinish with a differenceAn adjusting entry for the differenceAn explanation, unless someone adds one

The marker a forced balance leaves behind

Every mechanism above leaves a mark a dataset builder can search for, which turns a hidden plug into a countable exception. Run the search before anyone labels a case, then route every hit to explanation review instead of letting it pass as reconciled.

The mark is not a verdict. A rounding residue padded at period end may be acceptable under the company's policy; an expense booked to clear an unexplained bank gap may not be. What the mark proves is that software, not a person holding evidence, made the total agree, so the case needs a reviewer to supply the cause, the document and the approval, or to label it unexplained.

Count the marks and publish the rate per account and period with the dataset. A period full of forced items is a period whose totals carried little information, and a buyer should know that before treating its reconciliations as ground truth.

Search rules a supplier can run on an export, one per mechanism:

  • Accounts whose names begin with Imbalance- or Orphan- (GnuCash), plus whatever suspense or discrepancy accounts the company's own chart uses.
  • Transactions of the reconciliation type, or whose description begins with Reconciliation and a date range (Firefly III).
  • Postings that carry a balance assignment, an equals sign with no amount before it, in the source journal (hledger).
  • Transactions with the P flag or a narration beginning with Padding inserted (Beancount).
  • Adjusting entries created when a reconciliation was finished with a difference still showing (QuickBooks Online).

What federal reconciliation guidance and audits say about shrinking differences

That shrinking a difference is not the same as explaining it. Treasury's Fund Balance with Treasury guidance makes agencies responsible for investigating and resolving differences with Treasury's records and for reporting any required adjustments, and its guidance on the reconciliations agencies perform says every adjustment should be researched and traceable to supporting documents, with corrective action aimed at the root cause where possible.

GAO's audit of the fiscal year 2025 and 2024 consolidated financial statements of the U.S. government shows why the distinction matters. It credits Treasury with working with agencies since fiscal year 2023 to reduce intragovernmental differences each year to amounts Treasury judged immaterial to the consolidated statements. It still lists the government's inability to adequately account for intragovernmental activity and balances among three impediments to an opinion, and says it is uncertain the improvement can be sustained given underlying deficiencies at federal entities. The totals moved; the explanations behind them remained the open question.

Authorization belongs in the same chain. A Treasury Office of Inspector General audit report (OIG-17-050), published by the Fiscal Service's Administrative Resource Center, describes a control in which a difference that needs correcting is posted by an accounting technician or another accountant who neither prepared nor reviewed the reconciliation. A dataset that records who prepared, who reviewed and who posted each adjustment can test that separation; one that keeps only the adjusted balance cannot.

Tied to the cent and still wrong: a finance exception casebook

Every case below ties exactly and still fails at least one of the three grades that matter. The cases are illustrative, built from the mechanisms above and common close patterns, and describe no real company. A supplier's sample should include cases like these, labeled, because a model that never sees them learns that tying is the same as being right. Graded in order, all seven pass arithmetic, three still stand after the evidence grade, two after treatment and none after authorization.

Seven cases, each with the evidence that would turn the failure into an explained item:

Tied to the cent and still wrong: a finance exception casebook
Illustrative caseWhy the total tiesWhat failsEvidence that would explain itLabel
Bank fee gap cleared by a scrubBalancing split posted to an Imbalance accountEvidence and treatmentStatement line showing the fee; entry moved to bank chargesforced
Two misposts that cancelA receipt applied to the wrong customer offsets a missing depositEvidenceRemittance advice and deposit slip matched line by lineoffset_masked
Correction booked at period endReconciliation transaction for the full differenceEvidence and authorizationThe cause, and approval by someone who did not prepare the reconciliationforced
Untracked spend assigned to expenseBalance assignment books the gap to a miscellaneous accountTreatmentReceipts, or a policy that allows the write-off and sets its limitasserted
Timing item open for three periodsListed as in transit at every month endEvidenceThe item clearing the bank, or its reversal with a reasoncarried
Small write-offs under a limitEach gap sits inside the tool's tolerance or the company's limitAuthorization once the period total passes the limitWrite-off policy and the period's cumulative totaltolerated
Adjustment approved after postingEntry is correct; the approval is dated laterAuthorizationAn approval timestamp earlier than the posting timestamplate_approval
Four horizontal bars show the seven illustrative casebook cases graded in order: seven pass arithmetic, three remain after evidence, two after treatment and none after authorization.
The finance exception casebook graded grade by grade. Illustrative cases; no measured values.

Four grades that never average: the reconciliation reviewer rubric

Grade each reconciling item on arithmetic, evidence, treatment and authorization separately, and never let a pass on one offset a fail on another. Label an item documented only when all four pass. Arithmetic alone earns nothing, because every case in the casebook passes it.

Treatment needs a qualified accounting reviewer, not a general annotator. Whether a difference belongs in bank charges, a suspense account or a prior-period correction depends on the company's policies and the applicable accounting standards, and careful reviewers can disagree. Record each reviewer's role and the policy or standard section cited, and send a share of items to a second reviewer to measure agreement before the labels are used for training or grading.

The rubric, grade by grade:

Four grades that never average: the reconciliation reviewer rubric
GradePassPartialFail
ArithmeticItem amounts sum to the difference with no residualResidual within a stated toleranceResidual outside the tolerance
EvidenceEach item links to a source record outside the ledgerEvidence described in a note but not attachedNo evidence, or only a forced-balance marker
TreatmentAccount and period fit the cited policy or standardPlausible, but no policy or standard citedAccount or period contradicts the policy
AuthorizationApprover differs from preparer and approved before postingApprover recorded, timing unclearNo approver, self-approval, or approval after posting

Item-level fields that tie each adjustment to its evidence and approver

Store one row per reconciling item rather than one per reconciliation, so each adjustment carries its own evidence and approval. These fields are the minimum, and each should point to a record identifier instead of restating the record:

  • item_id and reconciliation_id: the item and the account-period reconciliation it belongs to.
  • amount_relation: how the item relates to the difference (equals it, is part of it, offsets another item), kept as a relation where amounts are sensitive.
  • forced_marker: the mechanism that produced the item, such as an Imbalance split, a reconciliation transaction, a P flag, a balance assignment or a discrepancy account, or none.
  • evidence_refs: identifiers of the bank lines, remittances, invoices or statements that support the item.
  • treatment_basis: the account chosen and the policy or standard section cited for it.
  • preparer_token, reviewer_token, poster_token: stable pseudonyms for the three roles, so separation of duties can be checked.
  • approved_at and posted_at: full timestamps compared exactly, never truncated to dates, so a same-day late approval still shows.
  • grades and label: the four rubric grades and one casebook label, with the reviewer's role.

Is a tied total ever enough?

Sometimes, for a bounded class of items, and the exception should be written down rather than assumed. Tools build tolerance in on purpose: Beancount inserts no padding when a gap is within its tolerance, and small rounding residues on currency conversions or interest lines are routine. A policy that sets a write-off limit, names who may approve it and caps the period total can make such items acceptable.

The cost of treating every tied total as explained is easier to see. A model trained on forced items learns that posting the difference is the fix, and an evaluation set built from them rewards it for proposing exactly that. Keep tolerated items labeled tolerated, keep forced items labeled forced, and let the buyer decide whether either belongs in training.

Requesting reconciliation cases through SourceX, and the limits of this method

Describe the item-level task, the evidence systems and the four grades in the request, and ask for the forced-balance markers by name. On the request form at sourcex.si/request-data, choosing Workflow datasets opens a workflow specification whose workflow type field lists reconciliation among its examples; put the casebook labels in the required outcome labels field, and the rubric with a second-reviewer agreement target in the evaluation criteria field.

SourceX finds matching records at businesses after a buyer describes the need; it holds no finance datasets in stock, and a request does not mean matching cases exist or that a sample will pass. The company that holds the records approves the scope and terms and signs before anything is delivered, and de-identification is done by SourceX directly or through its processing partner. Ask for bank account numbers, counterparty names and staff names to be replaced with stable tokens so the evidence links survive. Removing identifiers does not settle whether the records may be licensed; that turns on the company's contracts and is reviewed separately.

The method has limits. The code reading covers four open-source tools and one help page; enterprise ERPs have their own clearing and write-off functions, and settings vary by company. Nothing here measures how often forced items occur in real books. That rate, reported per account and period, is the next evidence a pilot sample should produce.

Frequently asked questions

What is a reconciliation explanation dataset?

A set of reconciling items, each linked to the source record that supports it, the accounting treatment chosen and the person who approved it, with separate grades for arithmetic, evidence, treatment and authorization. It is built to teach or test why a difference exists, not only that the totals agree.

Why is a zero reconciliation difference not proof that a reconciliation is right?

Because software can produce the zero. GnuCash posts unbalanced amounts to an Imbalance account, Firefly III and QuickBooks Online can book the remaining difference as an entry, and hledger and Beancount can compute or insert the amount needed to make a balance check pass.

How can a dataset builder find forced balances in an export?

Search for the marks each mechanism leaves: Imbalance or Orphan accounts, reconciliation-type transactions, P-flagged padding entries, balance assignments without amounts, and adjusting entries created at a reconciliation's finish. Route every hit to explanation review and report the rate per account and period.

Who should grade accounting treatment in a finance dataset?

A qualified accounting reviewer who records the policy or standard section relied on. Treatment depends on company policy and applicable standards, so a second reviewer should grade a share of items to measure agreement before the labels are used.

Can SourceX supply reconciliation cases with approvals attached?

SourceX sources records from businesses after a buyer describes the need, so a request can ask for item-level evidence, approver tokens and forced-balance markers. A request does not mean matching cases exist today, and the supplying company approves scope and terms and signs before any delivery.

Sources

  • GnuCash's scrub code balances an unbalanced transaction with a split in an Imbalance account named for the currency, creating it if needed, and moves account-less splits to an Orphan account. Source
  • Firefly III can create a reconciliation-type transaction for the remaining difference, dated at the period end and marked reconciled. Source
  • Firefly III's interface offers to create a correction for the difference and says a correction can always be edited or deleted later. Source
  • hledger's manual says balance assignments calculate the posting amount to satisfy an assertion and warns that forcing balances can hide errors and make data less trustworthy in an audit. Source
  • Beancount's pad plugin inserts a transaction naming the balance and difference so a later balance check passes, checks totals per currency, and pads only when the gap exceeds a tolerance. Source
  • Beancount marks transactions created from padding directives with the flag P. Source
  • QuickBooks Online lets a user finish a reconciliation with a difference, creating an adjusting entry for it, and Intuit advises consulting an accountant first. Source
  • PCAOB AS 2401 describes fraud through inappropriate or unauthorized journal entries and through adjustments not reflected in formal entries, and lists traits of suspect entries. Source
  • Treasury makes agencies responsible for investigating and resolving Fund Balance with Treasury differences and reporting required adjustments. Source
  • Treasury's reconciliation guidance says adjustments should be researched and traceable to supporting documents, with corrective action addressing root causes where possible. Source
  • GAO's FY 2025 and FY 2024 audit lists intragovernmental activity and balances among three impediments to an opinion despite Treasury reducing differences to amounts it judged immaterial. Source
  • A Treasury OIG audit report describes a control in which corrections are posted by someone who neither prepared nor reviewed the reconciliation. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies