Evaluating AI on real-world tax and accounting

Evaluating document parsing models on real-world K-1s
  1. Reducto r-1: 85.2%
  2. Extend Parse Performance 2.0: 76.0%
  3. Datalab Convert (Accurate mode): 69.9%
  4. LlamaIndex LlamaParse (Agentic Plus): 69.4%
  5. Mistral OCR 4.1: 61.7%
  6. Microsoft Azure Document Intelligence Layout: 39.8%
  1. Reductor-185.2%
  2. ExtendParse Performance 2.076.0%
  3. DatalabConvert (Accurate mode)69.9%
  4. LlamaIndexLlamaParse (Agentic Plus)69.4%
  5. MistralOCR 4.161.7%
  6. Microsoft AzureDocument Intelligence Layout39.8%
Read the entire K1Bench publication (opens in a new tab)
V.

Verona makes the world legible to frontier intelligence

We build evaluations grounded in real documents and expert judgment. Starting with tax and accounting, Verona turns complex professional work into clear signals of AI capability—helping researchers understand where models succeed, where they fall short, and what meaningful progress should look like.

What are your contact details?