Model developers
How does our model perform on real tasks?
Independent evaluation, model comparisons and representative failures.
We start with DailySafe to understand where AI fails in real tasks and what those failures could cost.
Research
DailySafe tests whether AI can use context from earlier turns and images to spot emerging risks and keep helping on everyday tasks. We plan to extend this approach to professional workflows such as coding, office work and customer support.
Pilot focus
Start with a specific task. Find where the system fails, check whether the failure can be reproduced and verify whether the next version fixes it. We want to test how this evidence can support decisions for three kinds of teams.
How does our model perform on real tasks?
Independent evaluation, model comparisons and representative failures.
Can the application handle this workflow reliably?
Scenario testing, pre-launch risk assessments and regression testing.
Which tasks should this system take on?
Model selection evidence, deployment risk reports and guidance on where to use the system.
Roadmap
Understanding actual losses takes evidence about where AI is used, how often it is used, which steps people review and what happens when it fails.
Published research
Start with DailySafe to measure safety and helpfulness across multi-turn tasks.
Near-term pilots
Test demand for paid assessments, tailored scenarios and repeat testing across releases, with ongoing monitoring as a next step.
Future research
Combine evaluation results, deployment conditions, human review and loss records to estimate the likelihood and cost of failures.
Long-term vision
Explore risk identification, pricing and controls with insurers, starting with enterprise AI. Personal AI coverage is a longer-term direction.
Collaborate
We're interested in hearing about the AI workflows you deploy, the decisions you need evidence for and the risks you most want to test.