OCR + vLLM evidence

Ground image answers in both OCR and model interpretation.

Ground image answers in both OCR and model interpretation:

pip install pytesseract pillow
arka vision-evidence screen.png "What price is shown?" \\
  --model-view "The visual model sees a price label reading $10"
arka vision-evidence screen.png "What price is shown?" --json

Arka reports what OCR says, what the visual/vLLM model says, and a combined answer that identifies agreement, conflicts, and uncertainty. It does not silently treat a model guess as text evidence. Natural language requests are supported.