Latest updates

ProofPudding scores 99.3% on FinanceBench

ProofPudding v1.0 scored 99.3% (149 out of 150) on the FinanceBench benchmark, the highest reported accuracy to date. Every answer included page-level citations back to the source SEC filings.

Fin-RATE: 97.4% on financial reasoning

ProofPudding scored 97.4% accuracy on Yale's Fin-RATE benchmark. The next-highest system, Fin-R1, scored 57.5%. GPT-5 with web search scored 43.0%.

Stopping LLM Hallucinations in Extraction

LLM extraction hallucinations are rarely fabrications — they're locally plausible errors that look like real data. Here's how verification catches what prompting can't.

Document Extraction: Build or Buy?

If you have LLM access and a PDF library, building your own extraction pipeline feels within reach. Here's what grows in scope, what silently fails, and when buying wins.

LLM vs. OCR for Document Processing

OCR and LLMs both take documents and produce text, but the choice between them isn't the hard part. Here's why verification matters more than the parsing layer.