Verified Small-Model Document Extraction for the Financial Close

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We describe the document-reading model behind Money Maxer, an agent finance team that closes a simulated month for a mid-market software company. The system's premise is that no agent is trusted: every journal entry is re-performed by a small deterministic kernel before it can post. Inside that design we ask how small, cheap and local the model that reads finance documents can be. We LoRA-fine-tune open-weight models from 0.6B to 20B parameters on 956 synthetic documents and evaluate on NorthwindBench, a benchmark we built with a time split, held-out customers and vendors, and held-out document templates. With no schema in its prompt, the fine-tuned Qwen3-4B reaches 0.972 field-F1 and 100% schema validity on 120 test documents, up from 0.061 and 0% for the same model untouched. Four frontier Claude models given the full schema in every prompt score 0.943 to 0.958. At n=120 that gap is within noise, so our claim is parity, not superiority: a model that runs on one desktop matches frontier extraction quality at an estimated $0.15 to $0.40 per million generated tokens, against $5 to $50 for the frontier models. A fine-tuned 0.6B matches the 4B on the benchmark at three times the speed. Through the real kernel and ledger, the tuned reader settles exactly the 168 of 200 remittances that gold labels settle, with zero wrong postings. We report the unflattering results too: the 20B is the weakest of the family, the 0.6B fails on long invoices, and the benchmark is synthetic and self-built. Built at HackMIT 2026. Code and data: https://github.com/Preet37/money-maxing