Provenance Research
Provenance Research builds reinforcement learning environments cloned from real company data, letting AI agents train and be graded on authentic professional workflows across 13+ connected business tools. The platform sandboxes systems like Gmail, HubSpot, Mercury, and Carta, with expert-signed graders that automatically check record consistency—for example, flagging when a signed SAFE agreement's $10M cap doesn't match a Carta entry showing $8M.
- Artificial Intelligence
- AI Agents
- Financial Technology
- Developer Tools
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
AI models are typically trained and evaluated on generic benchmarks that do not reflect the complexity, stakes, or specific workflows of real professional roles. Enterprises cannot safely test or fine-tune AI agents on live systems because mistakes in production are costly, and firms lack a controlled way to measure agent performance on their actual day-to-day work.
Solution
Provenance Research rebuilds real professional work as reinforcement learning environments by creating sandboxed clones of enterprise workflows. These environments replicate the systems, data feeds, and rules agents would encounter in production—such as order management systems, risk engines, market replay data, and SEC filings—without any route to live accounts. Graders, which are expert-signed rule sets, automatically check what the agent leaves behind (e.g., the book, ledger, or record) against the firm's standards. Held-out evaluation periods ensure scores come from data after the model's cutoff, preventing memorization. As systems or rules change, the platform updates the environment so agents continuously train on today's job.
Target Audience
Primary users are AI labs seeking to train and evaluate agents on real professional work, and enterprises—such as asset managers, execution desks, and market-making firms—that need a safe way to test AI before deployment.
Features
- Sandboxed environment with simulated order management, risk engines, market replay, SEC filings, and borrow models, isolated from production systems
- Expert-signed graders that assess agent outputs against firm-specific rules and compliance requirements
- Held-out evaluation methodology that scores agents on periods after data cutoff to prevent memorization
- PM-bench benchmark simulating a discretionary long/short equity portfolio manager with a $10M book, 63 trading days, and limits like 2× gross and 25% per name
- Continuous upkeep process that updates environments when systems or rules change to keep agents relevant
- 13+ cloned business tools connected as in real life, such as Gmail, HubSpot, Drive, Calendar, DocuSign, Mercury, and Carta, enabling cross-app agent workflows
- Automated record-consistency checks that flag mismatches like a signed SAFE agreement showing a $10M cap versus a Carta entry showing $8M
- Detailed trace logs showing every tool call (e.g., 412 tool calls across 7 apps over 6 simulated days) for full agent auditing