Governed AI for citation grade benchmark research
Specialist agents prepare evidence, rules enforce the method and experts decide every score.
A global business and human rights benchmarkProblem statement
What the client was up against
Every benchmark score must trace to a document, a rule and a named expert. Skilled researchers did this across long disclosures and spreadsheets. It was rigorous but slow, and hard to scale without adding people.
Our approach
How we approached it
- Froze sources and cited evidence to the page
- Turned the methodology into machine enforced rules
- Used deterministic scoring with risk based expert review
- Kept an immutable record of every decision
Solution we created
What we built
BenchmarkOS, a specialist agent system with configurable methodology, deterministic scoring, four eyes review, right of reply and sealed, citable outputs.
Agents in the system
The multi agent team behind it
SourcesCompany disclosures, methodology workbook, investor portal
Evidence agentFinds and cites relevant disclosures
Eligibility agentApplies scope, recency and threshold rules
Scoring engineCalculates scores the same way every time
Review workflowRecords accept, revise and reject decisions
People decideReview, approve and act
Evidence agentFinds and cites relevant disclosures
Eligibility agentApplies scope, recency and threshold rules
Scoring engineCalculates scores the same way every time
Review workflowRecords accept, revise and reject decisions
Company disclosuresmethodology workbookinvestor portal
What it was built to change
The outcome it is designed for
Pilot targets at least 30% less expert time, 98% citation correctness and 95% evidence recall, with 100% agreement with the methodology.
This is a pilot without published results yet. We show the problem and the build, not numbers we have not measured.
Client names, detailed case studies, demos and references can be shared on request.
Looking for a similar solution?
Tell us your version of this problem. We will walk you through what we built and what a first pilot would look like for you.
