A shadow ledger earns the right to become the system of record only after it matches the legacy field by field across a full close cycle, inside tolerances a controller has signed in advance.
Published dual-ledger validation guidance puts the duration floor at one clean quarter end, and a full year end for a core replacement.
The cutover decision is not a project milestone. It is a signed reconciliation. What follows sets the parallel-run length by system criticality, names the line items a controller countersigns on the tie-out matrix, and defines pass/fail thresholds concrete enough to argue in a steering committee.
What the Shadow Ledger Is Supposed to Prove
A shadow ledger is a parallel record of accounting data kept beside the official general ledger, running the same transactions through the new ERP while the legacy system remains the book of record. In a migration context it exists to prove three things before cutover: that every legacy transaction landed in the new ledger, that each posted to the correct account with the correct amount, and that the new system closes a period on the same numbers the legacy system closes on.
What it is not: a demo environment, a test tenant, or a sandbox run against synthetic data. A shadow ledger consumes production source data (bank feeds, invoices, payroll files, intercompany journals) and posts in real time, so the comparison with the legacy GL is being made on the same economic events, not on representative samples. If the intercompany workflow is in scope, the parallel also has to post the matched sides — see our write-up on repairing intercompany balances after cutover for what happens when it does not.
How Long the Parallel Run Should Actually Run
Duration is set by what the system has to prove, not by project calendar pressure. The upper bound should cover the longest routine accounting event the ledger will handle: for a general ledger that means a month end and at minimum one quarter end; for a revenue system that handles ASC 606 schedules or ASC 842 lease rolls, a full year end is often required to exercise annual true-ups and audit confirmations.
The second constraint is that the parallel must run long enough to see repeat cycles of the same workflow. A bank reconciliation that passes once proves nothing; a bank reconciliation that passes four consecutive weeks on a daily feed starts to look like a control. Where live feeds are new, the bank feed architecture itself needs time to prove it is complete, not just correct.
The Tie-Out Matrix a Controller Signs
The tie-out matrix is the single artifact that converts a parallel run from an engineering exercise into an auditable control. Each row is a balance the controller signs to; each column is the evidence it was produced the same way in both systems. At minimum the matrix covers:
- Trial balance — every GL account, by period, legacy vs shadow, with variance in currency and percentage.
- Subledger-to-GL tie — AR aging, AP aging, inventory valuation, fixed asset net book value, each reconciled to its control account.
- FX revaluation — unrealized gain/loss by currency pair, CTA roll, and the rate source used; our note on FX revaluation in a multi-currency close covers the common breaks.
- Tax — sales/VAT output, use tax accrual, withholding, and the filing-period cut-off applied to each.
- Accruals and deferrals — prepaid amortization, accrued payroll, revenue deferrals, reconciled schedule by schedule, not just balance to balance.
- Cash — bank reconciliation by account, with unreconciled items aged and categorized.

Each row carries an owner, a tolerance, a method, and a signature. Design the matrix before the first entity is migrated — our piece on designing the first multi-entity close argues the matrix is the migration plan, not a deliverable produced at the end of one.
Thresholds That Separate a Pass From an Investigation
A reconciliation threshold is a predefined tolerance used to decide whether a difference between two balances is acceptable or must be investigated, as one working definition puts it. The question is where to set it, and the SEC's own guidance is unambiguous: in its Staff Accounting Bulletin on materiality, the Commission states that exclusive reliance on a numerical threshold such as 5 percent has no basis in accounting literature or the law, though it may serve as an initial screening assumption.
Which means a defensible threshold schedule is tiered and qualitative on top of numerical:
- Zero tolerance on cash, intercompany, tax payable, and any account that posts to a regulated filing. A one-cent variance here is a research item.
- Rounding tolerance (sub-dollar) on FX revaluation where the rate source differs by rounding rule, documented with the rate source used.
- Materiality screen (commonly 1 to 5 percent of account balance, never the only test) on accrual estimates and allowance accounts, with any directional pattern across periods escalated regardless of size.
- Count tolerance on subledger transactions: record counts must match exactly, even where dollars tie within a screen.
Thresholds are not confidence intervals; they are the line at which a variance stops being noise and starts being a research item with an owner and a deadline.
Why Field-Level Comparison, Not Just Totals
Aggregate totals can hide offsetting errors: two accounts wrong by equal amounts in opposite directions produce a perfect total, as the same dual-ledger validation framework cited above notes. A trial balance that ties does not prove the ledger is right; it proves the two sides sum to the same number.
Field-level comparison is the control that closes this field-level mismatch. For every posted transaction the parallel keeps, the matrix should compare the amount, the account coded, the posting date, the entity, the cost centre, the currency, the FX rate used, and the document reference. Open-item files (AR invoices, AP vouchers, bank items) are matched by document key, not by summed balance. Where the shadow ledger is driven by an agent, the posting fields are exactly what should be logged and reviewed; the maintenance burden that implies is covered in the note on monthly maintenance for an AI agent.
The reason to run field-level reconciliation daily, not just at month end, is that daily diffs stay small enough to investigate. A month's worth of variances arriving together is an escalation, not a reconciliation.
The Signature That Makes a Shadow Ledger the Book of Record
Cutover is a decision that something has already been proved, not an event that proves it. Before signing, the controller should be able to point at: a tie-out matrix clean for the full duration without a repeating variance, a threshold schedule with every exception resolved or escalated by owner, a field-level reconciliation whose diffs are below screen on the final period, and a close dry run on the shadow ledger completed inside the target calendar. Vendors claiming AI-driven close automation should be asked how each of these lines is produced; our view on what that reconciliation workflow actually looks like is in how an agent runs the month-end close, and the implementation playbook, including the signed matrix, sits inside ERP migration with reconciliation proof.
The one question worth putting to any implementation partner is this: show the parallel-run exit criteria you will sign to, with the thresholds, the owners, and the evidence each line produces. A good answer is a document, not a methodology slide.
