# Discovery Clock

- **Event:** [Gemini 4 NYC Hackathon](https://cerebralvalley.ai/e/gemini-4-nyc)
- **When:** Sat, Nov 21 at 9:00 AM – 10:00 PM (EST)
- **Where:** New York, NY
- **Team:** [Sanat Gupta](https://cerebralvalley.ai/u/sanatgupta)
- **GitHub:** https://github.com/thesanatt/discovery-clock
- **Demo video:** https://youtu.be/imwy11JZKtc
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/17

Film a checklist job, type the steps, and Discovery Clock says which steps happened, when each started, and which were skipped, with a probability for each.

The problem: a seller films every order being packed, or a technician films a lockout procedure. When someone later asks whether a step was skipped, nobody has time to watch the footage. A single Gemini call can list the steps it saw, but it cannot say how sure it is.

How it works: four witnesses watch the same video. The Gemini Decisions API answers about 70 to 100 yes or no questions in one call ("had this step started by now?"), which gives a curve per step and a first read within a few seconds. Three Gemini 4 calls (Hydrogen twice, Neon once) write their own timelines. Bayes' rule combines them, weighting each witness by a reliability measured on held-out clips. A Gemini Managed Agent then audits start and end frames in a sandbox and marks each step confirm, dispute or cannot tell.

Proof: 518 held-out clips from DeepMind's Perception Test, split by a hash before any model ran, with two planted steps per clip that never happen. I shipped three versions and scored each once on clips it had never seen. Version 3, on 199 fresh clips: it catches planted steps at F1 0.87, level with one Gemini 4 Hydrogen call (0.87), the best of the fast single calls. What it adds is odds that hold up: of the steps it rated 0.8 or higher, 97 percent had happened; of those under 0.2, 11 percent had. On nine phone takes recorded at the event, 151 of 162 step verdicts were right.

Where it breaks: an already-open bag reads as "unzipped"; a faked step still read as done in 2 of 14 runs (version 2: 4 of 6); counts are only checked when the list makes them matter; on an 8 minute video it missed the one skipped step, and on long video its timing is 6 to 25 seconds off. The README lists every failure and seven things that did not help.

It is a review aid. A person checks every flagged step and signs.

Live demo: https://strategy-harmony-film-pot.trycloudflare.com

## More from Gemini 4 NYC Hackathon

- [Adds Up](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/14)
- [Elodie](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/15)
- [Amac and Murray](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/16)
- [Look to listen](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/18)
- [Casita Lista](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/19)
- [AREA-Q](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/20)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/17. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt
