CLEARWOOD
Independent risk monitoring for deployed AI
Prove your AI risk is under control
Proof takes evidence. We measure what your AI does in production, put the risk in dollars, and prepare summaries you can show to auditors and customers.
A public-benefit corporation, in design partnerships with companies in market intelligence and healthcare.
Example Bank · card servicing assistant
Misstated claims by week
Oct 5: new model version
In-production monitoring
Checking all outputs and agent actions over live traffic, covering the gap between pre-launch tests and post-incident investigations.
Before launch
Tested
Evaluations and red-teaming.
Your model vendors and teams
In production
Continuously checked
Where AI acts with real consequences.
Clearwood
After an incident
Investigated
Post-mortems, complaints, and headlines.
Your teams, auditors, and regulators
Flag the failures that cost you
Ungrounded answers
When the assistant gives confident answers with no basis in the source.
Actions that break policy
When the agent goes against your warnings or prohibitions in ways that are missed by cyber controls.
Summaries that contradict their source
When it looks plausible but actually says the opposite of what's in the record.
We measure all your AI activity
Audit and certification
Sign-off on the record, once a year
SOC 2, ISO 27001, ISO 42001Controls checked by an accredited firm
Clearwood
Independent measurement, every day
What the AI did, with the evidence behind each findingClear guidance for you, ready to share with auditors and customers as needed
Observability
The engineers' own tools
Debugging and quality checksFor technical troubleshooting, one issue at a time
We read from the telemetry you already run. Like a pen-test report, the findings are yours.
Evidence behind every number
What the fee schedule says
Fee schedule 4.2: a late fee of up to $30 applies when the minimum payment is not received by the due date. The first late fee is waived only for accounts enrolled in Autopay for the previous 12 months.
What the assistant said
Since this is your first late payment, the late fee will be waived automatically.
Why
The model answered from its own knowledge, not the fee schedule.Cost
$690K a month in misstated fees and rates.Fix
No number or deadline without a retrieved source.Every output checked, inside your environment
Collect
We read your AI's outputs and actions from your logs or telemetry.
Score
Small classifiers trained on your own examples score every one, at a cost that makes full coverage possible.
Review
Riskier cases escalated to human or expensive judge review. Calibrated with client audits.
Report
Findings and alerts, each traced to the source.
Classifiers like ours already run in production at frontier labs.
Start small, on your own data
Three steps, each useful on its own. Minimal integration, within your data governance rules.
Look back
About two weeks · historical outputs
The failures we found in your own data, each with the passage and the source.
Start here →Shadow
About 30 days · live traffic · read-only
Daily summaries and change alerts in the background. Your flows don't change.
Ongoing monitoring
Your highest-value workflows
Continuous reports and alerts, value and risk joined to your customer data, and evidence ready for your auditors.
The team
Matt Levinson
PhD, statistics. Scaled Nike's ML org from 6 to 100. Conducted original, fellowship-funded research in technical AI safety.
Michael Klear
Shipped production ML systems at Nike and Coinbase. Founding engineer at Yofi (acquired).
Scott Burns
Co-founded Guide Financial (acquired). Launched WeaveGrid's first products for utilities.
Clearwood is a public-benefit corporation. Integrity and safety are written into our charter.
Start with a free AI risk check
Tell us about the workflow you're worried about. We'll take a read of your historical data and give you visibility on your risk from AI output, with guidance on how to address it.