HelixCove

HelixCove is a dataset of "human traces", a record of how work unfolds across an organization over time, released by the startup Arbium on October 28, 2026. The first release covers three weeks at a fictional biotechnology company: 12 employees, 1,912 work sessions and 59,768 actions across Word, Outlook, Teams, Excel, SharePoint and other apps. All people, documents and events are fictional, but the activity is generated from patterns in hundreds of thousands of hours of real work.

Announcement: Introducing HelixCove · Explore it: helixcove.arbium.ai · Launch post: Arbium on X

Key points

1. Human traces: Arbium's term for computer activity stored over time. A trace links what someone sees and does to the activity around it: the conversation that prompted an action, the document it changed and the colleague whose next decision depends on it.
2. Why it matters: most workplace data records outputs. Traces show how outputs came to exist, through intermediate decisions, incomplete knowledge, corrections and coordination. Arbium's thesis is that this is the most valuable data for AI that does real work.
3. What each record holds: timestamp, application and window, event type and accessibility-tree text, with screen recordings for visual context.
4. Benchmark: facts planted in the story have known answers, and 430 questions test whether a system can recover them, from simple lookups to changes over time, multi-step and team questions, and questions with no answer in the data. Frontier models scored in the high 60s to low 70s percent using only the traces, per coverage on X.
5. How it was generated: Arbium specifies a fictional company, its staff, documents and story events, then has agents play the employees on virtual machines running real applications in one shared workspace. Eighteen measures of a working day's shape (burstiness, app switching, pauses and so on) are checked against ranges from real people.
6. Privacy: real-work observations only inform behavioral measurements such as timing and app transitions; all business content is fictional.

How to explore it

• Open the interactive workplace and benchmark at helixcove.arbium.ai. You can replay saved model answers, sort results by accuracy, response time or cost, and filter by question category.
• Read the method write-up at arbium.ai/helix-cove.

Good fits

• Researching AI agents that learn from computer-use observations
• Testing memory and retrieval systems on questions that span people, apps and time
• Studying agent context for workplace tasks that depend on colleagues and work in progress

About Arbium

Arbium builds "lived context for AI agents": its ROOTS-1 product turns the decisions and actions behind everyday work into context agents can use. It also researches on-device small language models with Carnegie Mellon's CyLab.