Don't let the LLM do the math: deterministic discount proration for receipt OCR
A vision model reads the receipt fine, then quietly loses a cent splitting the discount. Here's why I moved the arithmetic out of the model into a small Python function whose shares always sum to the amount paid.