Systems and records
A cross system balance reconciliation dataset that keeps both numbers
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A cross system balance reconciliation dataset should keep both systems' figures, each one's cutoff and definition, and a label for every difference: timing, definition, error on a named side, or unsupported adjustment. No system is the record of truth for every attribute. The federal DATA Act Broker shows the split: its 14 account-level checks against Treasury's GTAS data block a submission, while all 8 award-level checks against other systems only warn.
System export
Cross system balance
Key takeaways
- Authority belongs to an attribute, a level and a date, not to a whole system: the bank governs what cleared, the ledger governs what the company committed and reported.
- In the DATA Act Broker's 242 rule rows (commit bed939e, October 2026), only 6 of 25 cross-file rules block a submission; the other 19 warn and report the difference.
- Label the difference, not just the match: timing items carry forward, definition gaps get a bridge, the side that erred corrects, and plugs stay open.
- Eight Broker rules expect two figures to agree with opposite signs, a definition difference that a plain equality test would misread as an error.
- Grade five things separately: both values kept, difference class, correcting side, authorization and next-period follow-through.
Which system is the record of truth when two finance systems disagree?
Neither, as a rule. Each system is authoritative for something specific, at a stated level of detail and as of a stated time. The bank is the record of what cleared the account, the general ledger is the record of what the company committed and recognized, and a billing system is the record of what was invoiced. When their balances differ, the useful question is which side, if any, has to correct, and on what evidence.
That is the decision a cross system balance reconciliation dataset has to teach. A model that always trusts the outside system will adjust the books for a deposit that clears next week. A model that always trusts the ledger will defend a journal someone posted straight to a control account. Both behaviors end with a zero difference, and both can be wrong.
So every case needs four things: both figures with their as-of times, the definition and cutoff each system applies, a class for every reconciling item, and the name of the side that corrects. The rest of this page builds that design from a public rulebook, two regulatory precedents and an illustrative casebook, then sets out how qualified accountants should review the labels.
What 242 federal validation rules decide about two systems that disagree
It decides by comparison partner. The DATA Act Broker, the open-source validator the U.S. Treasury built for federal spending submissions, makes an agency's account-level budget figures match, exactly, the SF 133 data held in Treasury's GTAS reporting system, and lets award-level figures that come from other government systems disagree, with a warning and a reported difference.
We counted the rule list in the Broker's public repository (fedspendingtransparency/data-act-broker-backend, commit bed939e of 14 October 2026). Its sqlRules.csv holds 242 rule rows. The validator's documentation says data failing a rule marked as an error fails validation until fixed and resubmitted, while warnings are displayed and written to downloadable reports; the submission guide adds that later steps can proceed with warnings outstanding.
The split is consistent. All 14 rules comparing File A, the account-level appropriations file, with lines of Treasury's GTAS SF 133 report are fatal, and the Broker can generate File A from that SF 133 data in the first place. The 5 comparisons with the same report made at program activity and object class level, in File B, are warnings. So are all 8 rules comparing File C, built from agency financial systems, with File D1 (procurement awards) and File D2 (published financial assistance awards). D1 is pulled for a date range by each award's action date, so the two sides need not share a cutoff.
Two design choices are worth copying into any finance dataset. Every rule in the Treasury and award groups, and every warning among the agency's own files, returns a column named difference, 38 rules in all, so a failed comparison arrives with both sides and the gap instead of a bare fail. And submitters can store a written comment against each file, which the Broker's interface documentation describes as explanations or notes for that file. The table sorts every rule row by what it compares against.
| Comparison partner | Rule rows | Fatal | Warning | What a failure does |
|---|---|---|---|---|
| Treasury GTAS SF 133, account level (File A) | 14 | 14 | 0 | Blocks the submission until File A matches |
| Treasury GTAS SF 133, program activity and object class (File B) | 5 | 0 | 5 | Warns and reports the difference |
| Award systems: File C against D1 and D2 | 8 | 0 | 8 | Warns and reports the unmatched or differing amount |
| The agency's own Files A, B and C against each other | 17 | 6 | 11 | Missing shared accounts block; amount and subset checks warn |
| No second source: single-file arithmetic, formats, reference lists | 198 | 143 | 55 | Mixed by rule |
Timing, definition, error or plug: what the difference itself says
Every reconciling item belongs to one of four classes, and the class, not the size, decides who acts. Timing differences clear on their own and are carried with an expected clearing date. Definition differences never clear; they are bridged by a documented rule. Errors are corrected by the system that made them. A plug, an entry made only to force agreement, corrects nothing and stays flagged until evidence replaces it.
At a glance
Timing, definition, error or plug: what the difference itself says
Tax reporting already handles the largest definition gap in finance this way. Schedule M-3 reconciles a large corporation's financial statement income with its taxable income, and its supporting statements carry separate columns for the amount per the income statement, the temporary difference, the permanent difference and the amount per the tax return. Where the books follow GAAP, the IRS instructions say differences that GAAP treats as temporary go in the temporary column. Neither figure is wrong; the bridge is the record.
Securities rules show the other half. SEC Rule 17a-13 requires covered broker-dealers to count and verify securities each quarter, compare the results with their records and enter unresolved differences in a security count difference account within seven business days, and it bars the people who keep those records from running the count. The disagreement becomes its own dated entry instead of being absorbed by the books.
Sign conventions are a definition difference that trips automated checks. Eight Broker rules, all warnings, expect two figures to agree with opposite signs, written into the rule text itself. A model that tests plain equality would flag every one of those matching rows as an error. These four labels go on every reconciling item:
- Timing: both sides are right as of their own cutoff; record the expected clearing date and check that it cleared.
- Definition: both sides are right under their own rules; record the bridging rule and the document that states it.
- Error, side named: one side recorded something wrong or missed it; record which side corrects and the evidence.
- Unsupported adjustment: an entry made to force agreement; keep it open, never reconciled, until evidence and an approval exist.
Sub-ledger and ledger, bank and book, billing and ERP: who governs which attribute
Each of these pairs splits authority by attribute, so a dataset should record which system governs each attribute in the case, and the policy that makes it govern, rather than one source-of-truth flag per system. The sub-ledger answers which customer and which document; the ledger answers what was reported for the period.
Processor and bank show why timing has to be a field. Stripe documents a payout reconciliation report that ties each automatic payout to the payments, refunds, disputes and fees it settles, with a separate section for transactions not yet settled at the report's end date. A month-end processor balance and a month-end bank balance can both be correct and still differ by the payouts in transit.
| Pair | System A governs | System B governs | Usual timing difference | Usual definition difference |
|---|---|---|---|---|
| AR or AP sub-ledger and ledger control account | Sub-ledger: customer or vendor, document and open items | Ledger: the reported balance for the period | Documents posted after the ledger period closed | Journals posted straight to the control account; revaluation booked only in the ledger |
| Bank statement and cash ledger | Bank: what cleared and on which value date | Ledger: what the company committed, cleared or not | Deposits in transit and outstanding checks | Booking date versus value date; sweep or pooled accounts shown net |
| Billing system and ERP revenue | Billing: what was invoiced, to whom, for which term | ERP: revenue recognized and deferred for the period | Invoices raised in one period for service in the next | Invoiced amount versus revenue earned over the service term |
| Payment processor and bank | Processor: charges, refunds, disputes and fees per payout | Bank: the net amount received and its date | Payouts in transit at month-end | Gross charges versus net deposits |
Casebook: eight disagreements and the label each one earns
The cases are illustrative, built to carry one label each; none describes a real company. Each gives the two figures, the class, the side that corrects and the evidence a reviewer must see before accepting the label.
Four of the eight end with no correction at all, and a fifth must stay open. A dataset that rewards a model for driving the difference to zero would teach the wrong action in more than half of them.
| Case (illustrative) | Two figures that differ | Class | Who corrects | Evidence for the label |
|---|---|---|---|---|
| 1. Deposit recorded on the last day of the month | Ledger cash above the bank balance by the deposit | Timing | Neither; carry it with an expected clearing date | Deposit record and the next statement showing it cleared |
| 2. Monthly bank service charge | Bank balance below the ledger by the charge | Error, books (omission) | Books record the charge | Statement line and the account agreement |
| 3. Another customer's deposit credited to the account | Bank balance above the ledger | Error, bank | Bank reverses it; books unchanged | The bank's correction on a later statement |
| 4. Journal posted straight to the receivables control account | Ledger control account differs from the sub-ledger total | Error, books (bypass) | Books reverse and repost through the sub-ledger, with approval | Journal with no customer reference; the approver's authority |
| 5. Annual subscription invoiced in March | Billing shows the full invoice; ERP recognizes one month | Definition | Neither; bridge through deferred revenue | Contract term, recognition policy, deferred revenue roll-forward |
| 6. Card payout sent on the 31st | Processor and bank balances differ by the payout | Timing, plus gross versus net | Neither; bridge with the itemized payout | Payout report lines and the bank credit date |
| 7. One contract obligation in the financial and award systems | Equal amounts with opposite signs, recorded in different months | Definition and timing | Neither; explain it in the file comment | Award action date, posting date and the rule's sign convention |
| 8. Residual posted to suspense so two balances agree | Difference now zero | Unsupported adjustment | No one yet; the item stays open | None on file: needs evidence and an approver before relabeling |
Reviewer rubric: five grades that never average into one
Score each case on five grades, each pass, partial or fail, and report them side by side. An answer can get the arithmetic right and fail on authority, and an average would hide exactly that.
At a glance
Reviewer rubric: five grades that never average into one
Two accountants label each packet independently and a third settles disagreements; the same grades then apply to a model's answer:
- Grade 1, both values kept: the answer cites each system's figure, as-of time and record ID, and overwrites neither.
- Grade 2, difference class: every reconciling item is labeled timing, definition, error with a named side, or unsupported adjustment, with its evidence.
- Grade 3, correcting side: the answer names the system that corrects, or says neither, and cites the policy that makes that system govern the attribute.
- Grade 4, authorization: any proposed adjustment names an approver role with authority for that account and amount band, and the preparer is not the approver.
- Grade 5, follow-through: timing items cleared when predicted in the next period's records, or the answer flags them as aged.
- Automatic fail: the answer forces agreement through suspense or another clearing account and calls the case reconciled.
Does a forced match ever settle the question?
No. A zero difference proves the arithmetic, not the accounting treatment or the authority behind the entries that produced it. Two offsetting errors tie, and so does a plug. The strongest objection to this design, that agreement is what reconciliation exists to produce, confuses the control with its evidence.
The Broker shows the limit plainly. Its 14 fatal Treasury rules make File A equal the SF 133 figures for the period, and a rule such as A6 fails on any nonzero gap. They cannot show that the obligations behind those figures were recorded correctly in either place. Matching the designated record moves the question to that record; it does not answer it.
Sometimes forcing agreement is the policy, for example when a small timing residual goes to a clearing account and reverses next period. Then the packet should carry that policy, its threshold and its approver, so a model learns the rule rather than the habit.
Packet fields that keep both numbers and the reason
Build each case as an authorized, anonymized task packet: the two extracts as of their own cutoffs, the item-level matching, the reviewer labels and the documents behind them. Replace bank, customer, vendor and staff identifiers with stable codes so links between records survive, and keep amounts, or their exact ratios, only where the supplying company has approved it.
At a glance
Packet fields that keep both numbers and the reason
A qualified accountant should check a sample of labels against the company's written policies before any packet is used for training or evaluation, and the packet should record that review. These fields are the minimum:
- Pair and level: the two system roles, such as sub-ledger and ledger, and whether they are compared by account, document or line.
- Definitions: each system's scope, sign convention, currency basis and cutoff rule, quoted from configuration or policy.
- Snapshots: both figures with as-of timestamps and the report or extract ID that produced each.
- Item matches: matched pairs, unmatched items on each side and the matching rule used.
- Labels: class, correcting side and governing system per attribute, each with an evidence reference.
- Adjustments: entry IDs, preparer and approver roles, approval date and the policy clause relied on.
- Explanations: the comment written at the time, kept verbatim after redaction.
- Follow-through: next-period evidence that timing items cleared, or the age of items that did not.
- Review record: both independent labels, the adjudicated label and the reviewer's qualification.
Requesting cross-system cases, and what this design has not shown
A buyer who needs these cases should name the system pairs, the comparison level, the span of periods and the acceptance test in the request, for example that every packet carries both snapshots and an evidence-backed correcting-side label. SourceX does not hold finance datasets ready to ship; it looks for U.S. businesses that can supply records matching a buyer's stated scope, volume, language, format, timing and licensing needs, and the supplying company approves the scope and terms and signs before anything is delivered.
What the design does not show: how often each class occurs in real companies, how well current models assign the labels, or whether agencies resolve Broker warnings or simply accept them. The rule count describes one public rulebook at one commit, not reconciliation practice elsewhere. No results from real packets are published yet.
Methods note: the Broker counts come from sqlRules.csv and the SQL files beside it at commit bed939e6e51e5b31cac4ece7e3389a7464e6e78a (14 October 2026), read on 27 October 2026 by a script that only opens those files. Each row of sqlRules.csv is one unit, so B9 and B24, listed once per file type, count twice (240 distinct labels). A row is a Treasury comparison when its category is accuracy and its message mentions SF 133, with or without a space or hyphen; an award-system comparison when it is flagged cross-file and its file type or target file is D1 or D2; an agency-file comparison when it is any other cross-file row. The difference count tests each rule's SQL for a column aliased as difference. Rows are tested in that order and none matched two classes.
Frequently asked questions
Is the general ledger always the record of truth?
No. The ledger is the record of what the company reported for a period, but customer and document detail lives in sub-ledgers and evidence of cleared cash lives at the bank. A dataset should name the governing system for each attribute and cite the policy behind that choice.
What is the difference between a timing difference and a definition difference?
A timing difference disappears once both systems pass the same cutoff, such as a deposit that clears next week. A definition difference persists because the systems measure different things, such as invoiced amounts against revenue earned over a term, so it needs a documented bridge rather than a correction.
Should a reconciliation dataset include cases that were forced to zero?
Yes, labeled as unsupported adjustments. They teach a model to recognize a plug and to refuse to call it reconciled. A dataset without them produces a model that has never seen the shortcut it most needs to refuse.
Why does the DATA Act Broker let some disagreements through?
Its designers made account-level figures match Treasury's SF 133 data exactly, but treat award-level mismatches between agency financial files and the procurement and assistance award files as warnings. Those files use different date bases, and some rules expect opposite signs, so the Broker reports the difference instead of rejecting the submission.
Can cross-system packets be built without exposing bank or customer details?
Yes, if identifiers are replaced with stable codes before export and the links between records are kept. Amounts can be kept, scaled or expressed as ratios depending on what the supplying company approves, but the item-level structure has to survive for the labels to be checkable.
Sources
- The DATA Act Broker's rule list holds 242 rule rows with severity, cross-file flag, file type and target file; counts by comparison partner are derived from it. Source
- Data failing a rule marked as an error fails validation and must be fixed and resubmitted; warnings and errors are displayed and written to downloadable reports. Source
- The Broker's submission guide says later steps can start once errors are fixed, with warnings allowed or acceptable. Source
- File A can be generated from the SF 133 data agencies submit, with its amounts summed from named SF 133 lines. Source
- File D1 is generated from procurement award records filtered by action date within the requested date range. Source
- Rule A6 compares File A's total budgetary resources with SF 133 line 1910 and fails on any nonzero difference, returning the difference. Source
- Rule C23.1 expects File C obligations to match D1 obligations with opposite signs and reports the difference. Source
- Submitters can store comments described as explanations or notes for particular files, one per file. Source
- Treasury built the Broker as an open-source tool that ingests spending data from agency award and financial systems and validates it. Source
- Schedule M-3 statements separate amounts per the income statement, temporary differences, permanent differences and amounts per the tax return; GAAP-temporary differences go in the temporary column. Source
- Rule 17a-13 requires quarterly security counts compared with the firm's records, unresolved differences recorded within seven business days, and counts by people independent of the records. Source
- Stripe's payout reconciliation report ties each automatic payout to the transactions it settles and lists transactions not yet settled at the report's end date. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.