fraud agent scraper failed on blurry chargeback evidence pdf
Gates chargeback evidence PDFs on image quality before OCR, preprocesses borderline pages, and routes failures to manual review. Use when the agent's scraper chokes on blurry scans or OCR returns garbage. Not for clean digital PDFs or general document parsing.
TL;DR
The evidence PDF is a blurry scan and OCR returns garbage, so the agent either crashes or extracts nonsense it then treats as fact. Fix: gate every page on image quality before OCR, preprocess what is salvageable, and route the rest to manual review with the quality report attached. OCR output from a failed quality gate must never drive a dispute decision.
fraud agent scraper failed on blurry chargeback evidence pdfUse this when
- OCR output on evidence PDFs is garbled or empty.
- The agent crashes on specific evidence files.
- Extracted amounts or dates contradict the readable parts of the document.
Not for this skill when
- Clean digital PDFs fail. That is a parser problem, not image quality.
- The PDF is password protected. That is an access problem.
Steps
- Confirm the cause: render the failing pages and inspect them by eye. Verify: the pages are visibly blurry, skewed, or low resolution.
- Add a quality gate before OCR: minimum DPI and a blur metric threshold per page. Verify: the failing pages are rejected by the gate.
- Preprocess borderline pages: upscale, deskew, and boost contrast. Verify: OCR confidence scores improve on the reprocessed pages versus the originals.
- Route rejected pages to manual review with the quality report attached. Verify: an analyst sees exactly why the page was rejected.
- Track OCR confidence per field and require human confirmation below threshold. Verify: low-confidence extractions never auto-file.
Variant phrasings
blurry pdf ocr failed
Common phrasing.
chargeback evidence scan unreadable
Evidence-focused search.
ocr garbage on evidence pdf
Output-focused phrasing.
evidence pdf too blurry for ocr
Quality-focused phrasing.
Compatibility: any OCR pipeline, from Tesseract to cloud vision APIs; the quality gate is resolution and blur math, independent of the engine.
Why it happens
Evidence PDFs are often phone photos of receipts or multi-generation faxes. OCR engines assume clean input, so garbage in becomes garbage out, and the agent treats the garbage as facts. Cheap scans are the norm in disputes, so the gate belongs in every evidence pipeline, not just this agent's.
Edge cases / pitfalls
- Reject per page, not per document: a few bad pages should not sink a good file.
- Preprocessing must never invent text; keep the original page alongside the enhanced one.
- Manual review queues need SLAs tied to dispute deadlines, or the gate just moves the failure downstream.
Provenance
Resolved from the public thread: https://vectle.com/posts/pst_77bHqCp887mq-r0DILa46g