Bank Statements to Spread-Ready Rows

Statements are tables that refuse to behave

A 40-page bank statement is one long table chopped across page breaks: headers repeat, balances flow, footers interrupt, and every bank invents its own layout. Naive parsers return 40 fragments; analysts retype them into one sheet. This post shows the structured way through.

Step 1: Extract with a table definition

Mark the transaction set as a table in your schema — an array of objects with a table hint:

{
  "type": "object",
  "properties": {
    "transactions": {
      "type": "array",
      "description": "One entry per statement line",
      "items": {
        "type": "object",
        "required": ["date", "description", "amount"],
        "properties": {
          "date": { "type": "string", "format": "date" },
          "description": { "type": "string" },
          "amount": { "type": "number" },
          "balance": { "type": "number" }
        }
      },
      "x-fieldry": { "kind": "table" }
    }
  }
}

Pages with tables cost +2 credits on top of the base 4. Pages without tables never pay it.

Step 2: Let the stitcher join the pages

Rows are stitched across page breaks into one dataset: repeated headers are detected so they never leak into the data, running balances verify continuity, and a gap flags a missing page instead of silently dropping rows. You get JSON rows, plus CSV and Markdown export with Excel-safe encoding.

Step 3: Reconcile before you trust

Two checks run on every table:

• Totals vs rows — the closing balance must foot with the transaction arithmetic. Mismatches return status: mismatch with evidence on both sides, never a silent rewrite.
• Continuity — opening balance of page N must equal closing balance of page N−1. Breaks mean missing or duplicated pages.

Underwriters work from cited rows after that: each transaction links to page and bounding box, so sampling a file takes minutes instead of an afternoon with a highlighter.

What it costs

A 40-page statement with tables throughout: 40 × 6 = 240 credits ($0.24). Through the batch queue, half that. Compare with an hour of analyst retyping.