ocr-data-entrylisted

# OCR & Data Entry Image/PDF in → clean structured data out. Never ship garbage silently. ## Pipeline 1. **Preprocess** — deskew, increase contrast, crop to the table/page. Low-quality scans fail without this. 2. **OCR** — recognize text *and* keep coordinates (needed to reconstruct rows and columns). 3. **Structure** — rebuild rows/columns from coordinates; one line item per row. 4. **Validate** — sanity rules: column sums match totals, dates parse, amounts are numeric, currency consistent. 5. **Output** — CSV (default), JSON, or XLSX. ## Output contract - Header row exactly as on the document. - Amounts numeric with a `.` decimal, currency in a separate column. - Any cell that could not be read is left **empty** and listed in a `_unclear` column — never guessed. - At the end print: rows extracted, rows with unclear cells, total check (sum vs printed total). ## Rules - If the total on the document ≠ sum of recognized lines, **flag it** and stop for review. - Never fabricate a missing digit. - Keep the original order of rows. - For multi-page documents, keep page number in a column. - Report confidence: list any row that was reconstructed from weak OCR.