# GroundTruth

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Timothy Adetunji](https://cerebralvalley.ai/u/timztech)
- **GitHub:** https://github.com/whizztimmy/groundtruth
- **Demo video:** https://www.youtube.com/watch?v=2nh3S1fUIpE
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2357

Every activity tracker classifies your time by application name. Chrome is "productive" whether you were reading API docs or three hours deep in a feed. The distinction people care about, focused work versus doomscrolling, is semantic, and no window-title metadata recovers it.

GroundTruth reads the screen instead and names the activity.

It escalates only when it has to. A process name resolves the unambiguous majority for free. When it cannot, and a browser never can, the agent captures a frame, reduces it to 1024px, measures how much of the screen changed since the last sample, and asks an open-weights model what the person is doing. Every answer carries a confidence and the evidence it rests on. Off-screen hours are stated by the user, not inferred. Once a week a reasoning model reports what it notices in the derived durations.

An open window is not an activity either. A focused editor on a machine set never to sleep used to record as confident deep work whether anybody was typing or not. Input alone cannot tell reading from an empty chair, so it is paired with screen churn, and a still screen nobody has touched is recorded as unattended rather than as work.

It also implements the metric every existing tracker uses. Hubstaff, the Upwork app and Time Doctor score a ten-minute segment by counting minutes containing any input. That fails twice over: reading documentation scores near zero and files as idle, while forty minutes on a feed scores near 100 and ranks as your best work. One keystroke in sixty seconds scores the same as sixty seconds of typing, which is why a mouse jiggler defeats it. GroundTruth computes that number honestly and puts it beside what the screen contained.

The privacy boundary is enforced in code. Classification runs on Muse Spark Open, on infrastructure the operator controls, and sending a frame to a vendor-managed endpoint raises PlacementViolation rather than succeeding. A backend with no declared placement defaults to the strictest option, so a misconfiguration costs a refused call instead of a silent leak.

That rule cost something real. SAM 3.1 would give better attention tracking and is available on the Meta Model API, but only as a managed endpoint, and frames may not go to one. It is implemented, tested, and deliberately unused. What ships instead measures screen churn locally with no model at all.

It knows what it does not know. Unambiguous screens return 0.88 confidence; contentless ones abstain at 0.35 and name the evidence they lack. Those reach the user as questions, and only when a span rests on two abstentions, because identical input varies by 0.53 on ambiguous screens and 0.06 on clear ones.

Not built, stated plainly: SAM region tracking is refused on placement grounds, audio check-ins are written but their model is served on no endpoint found, and nothing mobile ships. 483 tests across three suites.

## More from Meta Global AI Developer Hackathon

- [AisCart](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2354)
- [Career Glitch](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2355)
- [DevinWarp](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2356)
- [Proof Pilot](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2358)
- [Digital Dude](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2359)
- [Epigen](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2360)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2357. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt