Reconciliation
Matching bank lines to the ledger is easy. Timing differences, partial payments, fees netted from receipts and FX each need a judgment and a note.
Exception 14: cleared, not explained.
Finance agents pass automated checks and still book the wrong number: the right amount in the wrong period, an exception cleared instead of explained. Lasker sources the signal that catches it: evals by credentialed accountants, traces of how the close is really done and licensed, de-identified workpapers.
For teams training or evaluating agents for the close, reconciliation and audit.
TomPerson B, the invoice from Northwind SupplyVendor 1 came in after cut-off.
Accrue $48,200 in Q3 and flag it for review. Bank ref 4417 2290account.
Maria ChenPerson A · Controller · +1 415 555 0142phone
Illustrative example. Names, places and numbers are made up.
✓ Cleared to leave · rights on recordTying out a balance is the easy part. These are the places a controller or an audit senior returns the work.
Matching bank lines to the ledger is easy. Timing differences, partial payments, fees netted from receipts and FX each need a judgment and a note.
Exception 14: cleared, not explained.
An expense has to land in the period the service was received, with support. The right amount in the wrong month is still an error.
Invoice dated May 2 for April service: accrue in April.
Every entry needs the right accounts, support and an approval trail. A plausible entry with no support fails review.
JE-311: no support attached.
ASC 606 and IFRS 15 for revenue, ASC 842 and IFRS 16 for leases turn contract terms into numbers. The answer is in the contract, not the invoice.
Contract 22: performance obligations not separated.
Choosing a sample, testing it and documenting the result so another auditor can follow it.
Sample size not tied to the risk assessment.
Knowing which variance matters and writing a reason a reviewer accepts.
Rent expense up 18% month on month: no explanation.
Public data has the financial statements. It does not have the reconciliations, workpapers and review notes behind them.
Automated checks confirm that numbers tie out. A controller asks whether they are right, in the right period and supported.
Workpapers belong to the companies that prepared them and hold client and personal data, so they have to be licensed and cleaned, not scraped.
For teams whose agents do close and audit work that someone has to sign off.
Review notes on real work: what was flagged, why, and how it was fixed.
Source: review notes from controllers and audit seniors
Step-by-step traces of a close task, with the reasoning at each step.
Source: workflow traces by credentialed accountants
Exceptions, partial matches and unusual entries, not clean textbook data.
Source: licensed reconciliations with exception notes
Answers tied to the US GAAP or IFRS rule they rely on.
Source: expert evals with the standard cited
Two outputs for one task, ranked by an accountant, with the reason.
Source: preference data from reviewers
Different charts of accounts, ERPs and close calendars.
Source: records licensed from companies of different sizes
Expert-made data is open for pilots now. Licensed company records are cleaned and come with a rights record. We deliver in your schema.
| Dataset | What is in a record | Used for | Format |
|---|---|---|---|
| Expert evals | Agent output scored against a written rubric by a credentialed reviewer | Evals, release gates | Scores + rationale, JSONL |
| Workflow traces | A close task done step by step, with the reasoning at each step | Demonstrations, fine-tuning | JSONL + source files |
| Preference data | Two outputs for one task, ranked, with the reason | Reward models | JSONL |
| Review notes | Reviewer comment on a workpaper, the fix and the final version | Verifiers, critique models | JSONL + workpapers |
| Licensed reconciliations | De-identified reconciliations with each exception and how it was resolved | Training, evals | CSV + JSONL |
| Close procedures | Close calendars, checklists and procedures as a company runs them | Planning, grounding | PDF + extracted steps |
An expert eval of an agent’s bank reconciliation, and a licensed reconciliation with its exceptions.
Illustrative. Field names are our draft schema, and we adapt it to your pipeline in the pilot.
{
"record_id": "ev-0412",
"task": "Reconcile operating cash to GL, April close",
"framework": "US GAAP",
"reviewer": {
"credential": "CPA",
"verified": true
},
"scores": {
"accuracy": 4,
"cut_off": 2,
"exceptions_explained": 2,
"support": 3
},
"finding": "Ties out, but a May 2 deposit clears an April gap.",
"verdict": "corrections",
"rights": {
"source": "expert-authored",
"personal_data_removed": true
}
}{
"record_id": "rc-1187",
"owner": "company-07",
"period": "2026-04",
"account": "operating cash",
"framework": "IFRS",
"exceptions": [
{
"type": "timing",
"amount": 4210.00,
"resolution": "deposit in transit, cleared May 2"
},
{
"type": "fee_netted",
"amount": 35.50,
"resolution": "bank fee booked to expense"
}
],
"rights": {
"license": "non-exclusive, training + eval",
"personal_data_removed": true
}
}The questions a controller or an audit senior asks, written down before scoring so two reviewers can agree.
| Dimension | What the reviewer checks | Evidence per finding |
|---|---|---|
| Accuracy | Amounts, accounts and signs are right | Line reference |
| Period and cut-off | Entries land in the right period | Date evidence |
| Standard applied | The US GAAP or IFRS rule fits the facts | Standard cited |
| Exceptions | Every difference is explained, not just cleared | Note per item |
| Support | A reviewer could sign it off | Missing support listed |
Reviewers hold the credential the task needs, checked with the body that issued it. Items with disagreement go to a second reviewer and are adjudicated.
Pick the process where your agent is weakest. The pilot answers whether this data moves it.
One process, such as bank reconciliation or accruals, and the framework: US GAAP or IFRS.
Week one
A small set of evals and traces, scored against a rubric you approve.
You test it in your pipeline
Full set, ongoing work or stop. Price is set per dataset after the sample.
No commitment before
Banks and regulated companies often cannot send customer or financial records outside. Our reviewers work inside your environment, on your devices or virtual desktops, and only the findings you approve leave. It suits the close, reconciliations and credit work, scored by the same credentialed reviewers.
CPAs, ACCA members, chartered accountants and internal auditors, matched to the task and checked with the body that issued the credential.
Both. The framework is set per task and recorded on every record.
Yes. Send the tasks and your agent’s outputs, and reviewers score them against a rubric we agree with you first.
From the companies that own them, under a non-exclusive license with a rights record. Client material is excluded unless the client agreed in writing, and personal data is removed before release.
Yes. Reviewers work in your environment on your systems, under your access controls. Records stay with you, and only the findings you approve leave.
The license scope is written per dataset: training, evaluation or both, and whether a model trained on it can be used commercially.
Per dataset or per batch of tasks, after the sample. We do not publish ranges until we have closed deals to base them on.
Keep finance records? Your company keeps ownership. License close and reconciliation records on your terms.
License your records →Send the process, the framework and the volume. We reply with a rubric draft and what a pilot would include.