# The Reverend

- **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:** [Raghav Sharma](https://cerebralvalley.ai/u/Yatsu)
- **GitHub:** https://github.com/Yatsuiii/stillhere
- **Demo video:** https://www.youtube.com/watch?v=HNpeNk4psBM
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1431

Before epilepsy or tumour surgery, surgeons map where language lives so they don't cut it. A child too young for the usual tasks can listen to a story instead. After a severe brain injury, a person who cannot respond hears a story while their brain is compared with healthy brains hearing the same recording. One in four people who could not follow commands at the bedside showed command-following brain activity on fMRI or EEG (Bodien et al., NEJM 2024).

Both scans need a healthy reference for that exact recording. Today that means scanning 12 to 15 healthy volunteers first, about $5-15k of scanner time per story, per language. So a story in the person's own language, or read by someone they know, usually cannot be used. AI can make that recording in seconds; the healthy reference is what's missing.

StillHere forecasts that reference before anyone is scanned. A researcher types the study they want. Muse Spark 1.3 turns it into a plan from a tested menu, or declines with the reason. StillHere builds the conditions, Muse Voice Transcribe reads every version back before anyone is scanned, and Meta FAIR's TRIBE v2 forecasts the healthy brain response, live on a laptop.

Then it shows how far to trust the forecast: against real public scans, beside cheap baselines, with each test's rule written down before its brain data was read.

- On Paranoia, a story outside TRIBE v2's training and evaluation sets (22 listeners), the forecast was worth about 3 real volunteers, and at least 1 (exploratory).
- On a held-out section of an audiobook from TRIBE v2's own evaluation sets (49 listeners), about 5, and at least 2.
- On 22 real listeners, its language readout was stronger for a story intact than with its sentences shuffled (pre-registered, p = 0.00516).
- Listeners' language areas followed the forecast more closely in the 7 minutes it ranked highest than in the 7 it ranked lowest, for 17 of 22 listeners (p = 0.0004). On a second story it did not replicate (12 of 22, p = 0.34), so we show the ranking as a lead, not a tool to trust yet.
- Misses stay on the page: on a second outside audiobook it was worth less than one volunteer at its lower bound, and the film face test failed.

Research use only. Tested only in healthy volunteers and public data. Not a diagnosis or medical advice.

## More from Meta Global AI Developer Hackathon

- [Competition Ops Agent](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1428)
- [Creative Platform](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1429)
- [honeycomb.](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1430)
- [PolymathSolo](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1432)
- [muse-fieldnotes](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1433)
- [PDF2PPt Study](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1434)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1431. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt