Proxy Harness

Proxy Harness is an evaluation harness for applying an institutional investor's proxy-voting policy to SEC proxy statements with LLMs, built for Pennant (YC S26). Every number traces to the exact lines in the filing, and every wrong vote traces to the step that caused it. It ingests 2026 proxy statements (Apple, Microsoft, Alphabet, Meta, JPMorgan) into page-anchored chunks in MongoDB Atlas, with Voyage embeddings and hybrid retrieval (Atlas Search + Vector Search). A written voting policy is compiled into boolean rule trees, and votes are computed in code: a missing fact escalates to REVIEW instead of guessing. Six models via OpenRouter extract 14 facts per filing with verbatim citations. Code checks every citation, answers are graded against a 70-fact human-verified answer key, and each error is blamed on parse, retrieval, extraction or citation. The harness improves with a human in the loop. An LLM judge diagnoses failures, an agent proposes prompt, definition or search changes, a human approves each one, and a validation run keeps or reverts it. Configs, proposals, lessons and judge verdicts live in MongoDB as the harness's memory.

Team

Placement

Participant