Telling readings apart

For the misread your extraction cannot undo: once OCR has turned the image into text, every model after it reads the same text. A person looks at the image and says which reading is right.

Looks at the originalThe image goes with the question, not the text made of it.
One of your readingsNever a third guess.
Priced by the mistakeSend cost_of_error and the certainty bought matches it.

How it works

The question and its evidence go to the worker with the best record on this kind of judgment. A second opinion is bought only if the first cannot carry the question alone.

SubmitSend the readings.

CompareA person reads the evidence.

ReturnAnswer, confidence, receipts.

Common use cases

Resolving a low-confidence extracted field

Choosing between two OCR readings

Deciding which of two dates is meant

Settling an ambiguous amount before payment

Example request

{  "kind": "disambiguate",  "question": "Is the total 45.00 or 4.50?",  "answer_schema": { "type": "enum", "options": ["45.00", "4.50"] },  "context": { "image_url": "https://…/receipt.png" },  "deadline_ms": 30000,  "cost_of_error": "40.50",  "caller_confidence": 0.41}

{  "status": "resolved",  "answer": "45.00",  "confidence": 0.99,  "agreement": "unanimous",  "responders": 2,  "latency_ms": 5312}