Grounding verification for AI agent outputs.
AndResearch builds Okugaki — an independent verification layer that machine-checks the provenance of the values LLM agents output as journal entries and reports, starting with accounting.
okugaki — the attestation colophon of Japanese manuscript culture.
The approach
As agents automate more of the work, verifying where each value came from becomes a growing manual burden. Our approach is not to search agent output for ungrounded values — values with no source evidence — after the fact, but to stop them before the output ships. When a generated output carries an ungrounded value, Okugaki holds it, points out what is ungrounded, and asks the agent to ground it: the output passes once resolved, and is halted if not. Every observation and verdict is recorded in an audit ledger independent of the agent, and the monthly catch report is generated from that ledger.
For the same reason a verifier cannot live inside the audited system, verification belongs in an independent layer.
Research
- Paper on ledger vouching (the core mechanism) and measurements — arXiv, September 2026
- ledger-vouching (reference implementation) — September 2026
Early Partner Program
We measure the grounding rate of your AI agent outputs, free of charge, for a limited number of partners.
Learn moreFounder
Shinya Kouda
Founder & CEO, AndResearch Inc. — Sapporo, Japan
Born in Noto, Ishikawa. Joined the Japan Air Self-Defense Force straight out of junior high school; after leaving the service, studied mathematics through graduate school (differential geometry). Started as a data scientist at TDSE Inc., moved into NLP with Word2Vec, and used generative models such as T5 in production before the generative-AI boom. Independent since 2025, focused on grounding and governing LLM outputs; founded AndResearch Inc. in 2026.
Interests: grounding and governing LLM outputs; physical AI; AI for wildlife-damage mitigation.