Pre-launch · design partners
A statement reaches you as a file the applicant downloaded, saved and sent. Every step in between is a step where balances can be edited, debits deleted and transactions invented. Assay runs four classes of forensic check on every file and returns what failed, where, and the evidence behind it.
| Row | Date | Description | Amount | Balance |
|---|---|---|---|---|
| 09 | 03.04 | TRF/SALARY/APR | +840,000.00 | 1,204,880.12 |
| 13 | 07.04 | POS/RETAIL | −18,400.00 | 1,186,480.12 |
| 17 | 11.04 | ▸TRF/CONSULTING | +615,000.00 | 1,801,480.12 |
| 21 | 14.04 | POS/FUEL | −22,000.00 | 1,779,480.12 |
| 25 | 19.04 | ▸TRF/CONSULTING | +615,000.00 | 2,394,480.12 |
| 29 | 26.04 | UTIL/PWR | −31,250.00 | 2,363,230.12 |
Findings
Illustrative specimen, not a live analysis. Figures are invented.
The statement is the single most important document in the decision, and it is the one document you have no way to check. Everything downstream inherits that.
A statement arrives as a PDF the borrower downloaded, saved and sent. Every step between the bank and your credit file is a step where it could have been edited.
An analyst opens the file and forms an impression. There is no test being run, no threshold being applied, and no record of what was checked.
A doctored statement does not fail loudly. It fails months later as a loan that was written against income that never existed.
Each runs independently and reports independently. A file can fail one and pass three, and that pattern is itself informative.
Producer and creator metadata, incremental save history, embedded font sets, object-level edits. A statement exported by a bank looks nothing like one that has been through an editor, and the file says so before anyone reads a figure.
Every running balance is recomputed from the transaction column. Deleted debits and edited amounts break the chain, and the break localises to the exact row where the document stops being true.
Character spacing, baseline alignment, kerning and rendering differences across a single column. Retyped figures rarely match the surrounding glyph metrics, even when they look identical to a reviewer.
Repeated descriptions, cloned reference numbers and transaction patterns that recur too regularly to be organic — the signature of a statement padded to reach an income threshold.
A forensics product that overstates its own coverage has already failed at the thing it sells. This table is the real state of the system.
| Source | Region | Formats | Status |
|---|---|---|---|
| Tier-1 Nigerian banks | Nigeria | PDF statement exports | In development |
| Mid-tier & digital banks | Nigeria | PDF statement exports | Planned |
| Scanned & photographed statements | All | OCR pipeline | Planned |
| Kenya & Ghana banks | East / West Africa | PDF statement exports | Planned |
Nothing is marked in service yet. When tier-1 Nigerian coverage ships, that row changes and nothing else on this page does.
A single number would be easier to sell and worse to rely on. The response returns each check independently, what it found, and which rows to look at — so a reviewer can disagree with it.
curl https://api.assaycredit.com/v1/analyses \
-H "Authorization: Bearer $ASSAY_KEY" \
-F "country=NG" \
-F "purpose=credit_application" \
-F "file=@statement.pdf"{
"id": "anl_01hzx9km4",
"verdict": "tamper_indicated",
"checks": [
{
"id": "structural",
"result": "fail",
"detail": "Incremental save after export;
producer differs from issuer.",
"evidence": "meta.producer, meta.revisions"
},
{ "id": "arithmetic", "result": "pass" },
{
"id": "typographic",
"result": "fail",
"detail": "Glyph metrics differ on 2 rows
in the amount column.",
"evidence": "rows[11], rows[19]"
},
{
"id": "duplication",
"result": "fail",
"detail": "2 cloned transaction rows.",
"evidence": "rows[11], rows[19]"
}
],
"cash_flow": {
"verified_inflow_ngn": 840000,
"excluded_as_suspect_ngn": 1230000
}
}This is the planned interface, not a live endpoint. We publish it now so design partners can tell us where it is wrong before it hardens.
A score you cannot argue with is a score you cannot audit.
Why the response returns findings, not a number
Forgery detection is the wedge, not the whole product. A statement that passes is then worth analysing — and the signals that matter are not the ones a spreadsheet average shows.
We hold no compliance certifications and will not claim any. What we do instead:
We are building the structural and arithmetic checks against tier-1 Nigerian bank formats with design partners. Everything else on the format table is marked planned and will stay marked that way until it genuinely works. A forensics product that overstates its own coverage has already failed at the thing it sells.
Not yet, because we have not measured one honestly. Any number we quoted today would be from a sample we assembled ourselves, which tells you nothing. When we have results against a real portfolio with known outcomes, we will publish the methodology alongside the figure.
Open banking connects to the account directly, which removes the document from the loop entirely — where it is available and the applicant consents, it is the better answer. Most lending across these markets still runs on statements, and every applicant who will not connect an account still hands you a PDF. Assay is for that file.
Some, inevitably. That is why the response returns per-check findings and the evidence behind them rather than a single verdict — so a reviewer can see what triggered and decide. A black-box score would be easier to sell and worse to rely on.
They are processed and stored as evidence attached to the check, encrypted at rest. We do not use customer documents to train shared models. The data handling section sets out the specifics, and we will put it in writing for any design partner.
Join the list and tell us your volume and which banks you see most. We are taking design partners in sequence rather than opening self-serve access we cannot yet support properly.
07 · Access
The checks are only as good as the files they were built against. Early partners give us a corpus — including the applications you already rejected — and get direct influence over what ships.