Skip to content

Systems and records

Why the fix goes unlinked in an accounting correction history dataset

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

An accounting correction history dataset should keep three linked records for every fix: the original error, the entry that reverses or adjusts it, and the replacement that states the right amount. Overwriting keeps none of the pair. In two open-source ledgers with a built-in reversal, read at pinned commits, the reversal points back at the error but no field ties the replacement to it, so a model sees negations, not corrected answers.

Record timeline

Accounting correction

This page covers why the fix goes unlinked in an accounting correction history dataset.Illustrative

Key takeaways

  • Treat a correction as three records, the original, the reversing or adjusting entry and the replacement, plus the links between them, never as one edited row.
  • Akaunting rewrites a posted banking transaction in place unless it is reconciled or part of a transfer, and creates no transaction history table, so the error is gone.
  • Medici and Formance Ledger keep the error and point the reversal back at it; neither has a field tying the re-entered amount to the error it fixes.
  • Kill Bill's item adjustments are deltas linked to the original item, so the corrected amount is the original plus its adjustments: a complete pair.
  • Score detection and correction separately, as grammar-correction benchmarks do, and keep reviewed entries that needed no fix.

What should an accounting correction history dataset keep for each fix?

Three records and the links between them: the entry as first posted, the entry that reverses or adjusts it, and the replacement that states what should have been recorded. Each record keeps its own posting date, effective period and creating user, and each link says which record corrects which. A dataset that stores only the final balance, or only the edited row, has thrown away the error a model is supposed to learn to catch.

At a glance

What should an accounting correction history dataset keep for each fix

Why a figure changed, and whether the stated reason holds up, is a separate question; this page is about how the fix itself is stored and whether the trio can be rebuilt from an export. The fields that make it usable, defined so that two reviewers would fill them the same way:

  • Original record ID: the identifier of the entry as first posted, never reused for the corrected version.
  • Correction method: overwrite, reversal, adjusting entry or void and reissue, taken from the system's own record type rather than inferred from amounts that cancel.
  • Correcting record ID: the reversal or adjustment, with the name of the field that points back at the original.
  • Replacement record ID: the re-entered amount, or a note that none exists because the adjustment already carries the corrected value.
  • Link type: a stored field, a shared group key or a reviewer's match, recorded as such, because only the first two are evidence.
  • Dates: posting date and effective period of each record, plus the date the error was found.
  • Chain position: first correction, correction of a correction and so on, so that a reversed reversal is not counted as two errors.
  • Reviewer verdict: error, routine business reversal or estimate update, with the reviewer's role and the records they relied on.

Overwrite, reverse or adjust: what each method leaves in the ledger

The method decides what a dataset can recover. An overwrite leaves one row holding the new value; a reversal leaves the error, its mirror image and usually a fresh entry; an adjusting entry leaves the error and a delta that brings it to the right amount. Only the last two keep the error, and only some implementations keep the link.

TigerBeetle's documentation states the case for adding rather than rewriting. Its transfers are immutable, fixes are made with further transfers that reverse or adjust earlier ones, and as a result "The log of events includes the original error, when it took place, as well as any attempts to correct the record and when they took place." A reporting layer can still show only the corrected figure. Going the other way, from a rewritten record back to the error, is impossible.

The four methods compared on what each leaves for a model to learn from:

Overwrite, reverse or adjust: what each method leaves in the ledger
MethodRecords left behindOriginal value recoverable?What a model can learnWhere it shows up
Overwrite in placeOne row with the new value and a later update timeNo, unless a separate log captured itFinal answers only, with no error examplesTools that let users edit posted transactions
Reverse and re-enterThe error, a mirror-image reversal and a new entryYesError detection; the right answer only if the new entry is linkedLedgers with a void or revert action
Adjusting entryThe error and a delta to the corrected amountYesDetection, the size of the error and the corrected value by summingBilling item adjustments and manual adjusting journals
Void and reissue a documentThe voided document and a replacement documentYes, if the void is keptDocument-level errors, if a reference field ties the twoInvoicing and tax reporting workflows

We read the code and documentation of five open-source ledgers at commits pinned on October 27, 2026, asking what happens to a posted entry when someone fixes it. One overwrites, one appends and leaves linking to convention, two reverse with a stored back-pointer, and one adjusts with a linked delta. Neither of the two with a built-in reversal has a field that ties a replacement entry to the error it replaces.

Akaunting's job for updating a banking transaction saves the submitted values over the stored row; its own checks refuse the update only when the transaction is reconciled or belongs to a transfer. Its migrations create history tables for bills, invoices, documents and modules, but none for transactions, so after an edit the table holds the new amount and a newer timestamp. The earlier amount survives only if something outside the table kept it.

TigerBeetle never edits: corrections are new transfers. Its recipe suggests, without enforcing, a transfer code that marks a correction or its reason, and reusing the same user_data_128 value on the correcting transfers. Follow both, give any replacement transfer the same key, and the original, the correction and the replacement share one identifier; skip them and the link has to be rebuilt from amounts and timing.

Medici and Formance Ledger build the reversal in. Medici's void marks the original journal voided and records a void reason (given none, the new journal's memo becomes the original memo with a [VOID] prefix), then posts opposite entries whose _original_journal field names the journal being voided. Formance's revert posts a compensating transaction carrying revertsTransactionId, stamps the original with revertedByTransactionId and revertedAt, accepts optional metadata and answers a second revert with a conflict. Both can date the reversal at the original's timestamp instead of today: use_original_date in Medici, atEffectiveDate in Formance. Neither gives an ordinary new entry a field naming the error it replaces, so the re-entered amount is a stranger to the record it fixes unless someone writes the link into free-form metadata.

Kill Bill's invoice adjustments take the third route. An item adjustment or a repair adjustment is a new invoice item whose linked_item_id names the item it adjusts, so the corrected charge is the original amount plus its linked adjustments. Repair adjustments default to the description "Adjustment (subscription change)", a reminder that the same mechanism records ordinary business changes as well as mistakes. The comparison, ledger by ledger:

Five open-source ledgers, read for the links a fix leaves
Ledger (commit)How a fix is storedCorrection points to errorError points to correctionReplacement tied to error
Akaunting (67baffd)Posted transaction updated in placeNo separate correction existsNo separate correction existsNo: the error is overwritten
TigerBeetle (0450c32)New transfers reverse or adjust; originals are immutableBy convention: a shared user_data_128, suggested not enforcedBy the same conventionOnly if the convention is followed
Medici (54fa40b)Void flags the journal and posts opposite entriesYes: _original_journal on each reversing entryFlag only: voided and void_reasonNo field
Formance Ledger (e745536)Revert posts a compensating transactionYes: revertsTransactionIdYes: revertedByTransactionId and revertedAtNo field
Kill Bill (15177ae)Adjustment item added to the invoiceYes: linked_item_idNo fieldNot needed: corrected amount is original plus adjustments
Five ledgers in rows against three records, original, correction and replacement: Akaunting keeps only the overwritten row, TigerBeetle links all three by a shared key if the convention is followed, Medici and Formance link the reversal to the original but not the replacement, and Kill Bill links an adjustment that makes a replacement unnecessary.
Which links survive a fix in five open-source ledgers, read at pinned commits on October 27, 2026. Source: each project's code and documentation.

How a payroll tax form keeps the original, the change and the reason

The IRS form for correcting an employer's quarterly payroll tax return, Form 941-X, is a working model of a linked correction. Its instructions put the corrected total in column 1 and, in column 2, ask filers to "enter the amount you originally reported or as previously corrected", with the difference in column 3; they add, "Be sure to explain the reasons for this correction on line 43."

Two design choices carry over to a ledger dataset. The original and corrected figures sit side by side with the difference computed, so nobody has to rebuild the pair from scattered rows. And "as previously corrected" makes the form chain-aware: a second correction starts from the first correction's result, not from the original, which is the rule a dataset needs to avoid counting a reversed reversal twice.

Financial statements apply similar logic one level up. IAS 8, as issued for 2026, requires material prior period errors to be corrected retrospectively in the first financial statements authorised for issue after their discovery, by restating comparative amounts for the period in which the error occurred, and asks for disclosure of the nature of the error and the correction for each affected line item. The period an error belongs to and the date its fix is posted can therefore differ, and the back-dating options in Medici and Formance show how easily a posting date can be set either way. A correction case needs both clocks. Restated reports kept version by version are covered on the financial restatement page.

What can a model learn from an error and its fix?

Two skills, which should be scored apart: spotting that an entry is wrong, and producing the right replacement. Grammatical error correction research already separates them. The ERRANT toolkit's M2 format stores the original sentence, then each edit's span, error type and correction, and its scorer can evaluate detection "(i.e. ignoring the correction)". Its noop edit "explicitly indicates an annotator/system made no changes to the original sentence", which keeps reviewed-and-correct examples in the data.

A reversal-only pair supports the first skill and not the second. An error and its mirror image tell a model which entries were wrong; without a linked replacement they say nothing about what the entry should have been. Adjusting entries do better, because the corrected amount is the original plus the delta, and a replacement tied to the error by a stored link does best.

Edit histories can also teach preferences. Gao and coauthors (arXiv 2404.15269) infer a user's latent preference from the edits the user makes to an agent's output and use the inferred description to shape later responses, judging methods by how much editing users still have to do. The finance analogue is a model that learns a team's correction habits, which helps only if the corrections it learns from were themselves right.

The finance version needs negatives as well: entries that a reviewer examined and left alone. Without them a detector learns that every entry near a month-end close, or every entry by one clerk, is suspect.

Eight correction histories and the label each one earns

The casebook collects the shapes a correction history takes in an export, including the ones that should not be labeled errors. Every case is Illustrative, invented to show a pattern, with amounts in plain units; none describes any company's records. The label set is fixed so reviewers can agree, and the last column names the record that settles each case:

Eight correction histories and the label each one earns
Case (Illustrative)What the export showsWhat is missing or wrongLabelRecord that settles it
Vendor bill keyed at 4,500 instead of 450One bill row at 450 with a recent update timeThe keyed error was overwrittenUnrecoverable pairA separate audit log or the bill image, if either survives
Journal reversed and re-entered to another accountThe original, a reversal pointing back, a new entry the same dayNo field ties the new entry to the originalLinked reversal, unlinked replacementMatching amount and document plus the reviewer's match note
Reversal of a reversalAn entry voided, then the void itself undoneA simple count scores two errors for one entryChain with no net correctionChain position field and the second action's reason
Accrual reversed on the first day of the next monthAn entry and its scheduled oppositeNothing was wrong; the reversal is routineBusiness reversal, not an errorAccrual policy and the reversing schedule
Overbilled line reduced by an adjustment itemThe original item and an adjustment linked to itNothing; the corrected amount is the sumComplete pair, delta formThe adjustment's linked item ID and description
Reversal back-dated into a closed monthA reversal carrying the original's timestampThe fix landed in a period already reportedPeriod conflictClose calendar and the poster's authority on that date
Revised bonus estimate booked as a correctionA reversal and a new entry for the revised amountAn estimate update recorded as an errorEstimate updateThe estimate method and the date the new information arrived
One error replaced by three smaller entriesOne reversal and three new entriesWhich entries replace the error is a judgmentReconstructed replacementSource document totals and the reviewer's mapping

Pair integrity rubric: five checks before a case enters the set

Run five checks on every candidate case and record each result separately. A case enters the training set only when the first two pass; the other three shape its label. The first three checks can run automatically against a data dictionary that names each system's link fields, while the last two, and every casebook label, need a qualified accountant who knows the close calendar and the policy in force. Two reviewers working blind on a sample, with agreement reported per check, show whether the labels hold.

At a glance

Pair integrity rubric: five checks before a case enters the set

The checks, in the order they should run:

  • Original present: the entry as first posted is in the export with its own ID, amounts and dates. If not, the case is a final-answer example, not a correction.
  • Correction typed: the method comes from the system's record type, such as a void, a revert or an adjustment item, not from two amounts that happen to cancel. If not, label it unconfirmed.
  • Link recorded: the correction points at the original through a stored field or a shared key. A reviewer's match is allowed only when labeled reconstructed.
  • Replacement resolved: a linked replacement exists, or the corrected value is the original plus linked adjustments, or the case states that nothing replaced the original.
  • Chain closed: every later action on the same entry is included, and the final state recomputed from the chain equals what the ledger shows today.

Does a reversal that nets to zero prove there was an error?

No. A reversal proves that someone undid an entry, not that the entry was wrong or that undoing it was authorized. Routine accrual reversals, cancelled orders and plan changes use the same mechanics, which is why Kill Bill's repair adjustments default to a subscription-change description and why the casebook keeps a business-reversal label.

IAS 8 defines prior period errors as omissions and misstatements arising from a failure to use, or the misuse of, reliable information that was available. A figure revised because new information arrived later does not fit that definition, so a dataset that labels every reversal an error teaches a model to treat estimate updates and normal reversals as mistakes.

Arithmetic cannot settle the label either. A ledger that balances after the fix, and a correction whose amounts net exactly, say nothing about whether the treatment was right or whether the person who posted it could. Those calls belong to a qualified accountant working from the records named in the casebook, and the dataset should record who made each one.

Requesting correction chains from a company's ledger

Ask for chains, not rows: every entry that shares a correction link or key, exported with original IDs intact and identities replaced consistently, so that links survive de-identification. User accounts, names typed into memos and counterparties need the same pseudonym every time they appear, or the chain breaks at the first replaced value.

A useful requirement names the correction methods in scope, the minimum share of cases whose link is stored rather than reconstructed, the label set and the acceptance test, for example the five pair integrity checks run on a sample the buyer scores. It should also say which periods are closed, since back-dated reversals are only visible against a close calendar.

SourceX finds datasets for buyers on request instead of keeping a catalog, and the company that holds the records approves the scope and terms and signs before anything is delivered. A requirement written this way is where that search starts; whether any given ledger keeps the links is checked case by case, and nothing here implies that a particular company's records will meet it.

Methods note: commits read, rules applied and limits

The ledger comparison comes from bin/correction_links.py, which reads text files from shallow clones made on October 27, 2026 and runs nothing from any repository: akaunting/akaunting at 67baffd, tigerbeetle/tigerbeetle at 0450c32, flash-oss/medici at 54fa40b, formancehq/ledger at e745536 and killbill/killbill at 15177ae. For each check it searches named files for stated regular expressions and prints every matching line with file and line number. A no-field verdict means that searches for any other replacement link (for Formance, fields named correctsTransaction, correctedBy, supersedes or replacesTransaction; for the others, names containing replace, supersede or corrects in the models and API schema) also found nothing; for Akaunting, the history-table verdict means no migration creates a table whose name contains both transaction and histor, and the script prints the four history tables it does find.

The IRS instructions, IAS 8 and the edit-learning paper were read through search excerpts from their publishers' sites on October 27, 2026, and ERRANT's format in its repository at commit 2840b84. None of the five ledgers was run, and their code can change after the pinned commits.

Not measured: how often real companies use each method, what share of reversals in practice are errors rather than business events, and reviewer agreement on the rubric. Those need authorized company ledgers, and no results are reported here. Accounting points are general information, not accounting advice.

Frequently asked questions

What is an accounting correction history dataset?

It is a set of correction cases in which each case keeps the original entry, the entry that reversed or adjusted it, any replacement entry and the links among them, with dates and a reviewer's label. It differs from a ledger export, which shows final balances, and from a change log, which shows that a field changed but not which entries make up one fix.

Is reversing an entry better than editing it for AI training data?

Usually, because the error survives as its own record and can be used as an example. A reversal on its own still lacks the right answer, so the replacement entry must be linked to the error, or the fix must be an adjusting entry whose amount brings the original to the corrected value.

How do you link a replacement entry when the system has no field for it?

Match it during preparation using the same source document, counterparty and amount within a short window after the reversal, then label the link as reconstructed and record who matched it. Keep reconstructed links separate from stored ones so a buyer can test both.

Should reversed accruals count as corrections?

No. A scheduled accrual reversal is a routine business event that uses the same mechanics as an error correction. Label it as a business reversal, or a model will learn that every month-start reversal marks a mistake.

What can finance data teams borrow from grammar-correction datasets?

Store the original, each edit, its type and its correction; score detection separately from correction; and keep examples a reviewer judged correct with no change. The ERRANT toolkit's M2 format does all three for text.

Sources

  • Akaunting's transaction update job overwrites the stored row and refuses only reconciled or transfer transactions (commit 67baffd). Source
  • Akaunting's migrations define history tables for bills, invoices, documents and modules, none for transactions (commit 67baffd). Source
  • TigerBeetle transfers are immutable and are corrected with further transfers; the recipe suggests a correction code and a shared user_data_128 (commit 0450c32). Source
  • Medici voids a journal by flagging it, storing a void reason and posting opposite entries that carry _original_journal; use_original_date is optional (commit 54fa40b). Source
  • Formance Ledger's revert creates a compensating transaction linked both ways (revertsTransactionId, revertedByTransactionId, revertedAt), with optional metadata and atEffectiveDate (commit e745536). Source
  • Kill Bill's invoice_items table carries linked_item_id, used by adjustment items to name the item they adjust (commit 15177ae). Source
  • Kill Bill repair adjustments take a reversing item ID and default to the description Adjustment (subscription change) (commit 15177ae). Source
  • Form 941-X instructions: corrected total, amount originally reported or as previously corrected, the difference, and an explanation of reasons on line 43. Source
  • IAS 8 requires retrospective correction of material prior period errors, disclosure of each affected line item, and defines prior period errors. Source
  • ERRANT's M2 format records the original sentence and each edit's span, type and correction, supports detection-only scoring and noop edits (commit 2840b84). Source
  • Gao et al. (2024) learn a user's latent preference from their edits to agent output and use it to shape later responses. 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