# Ad Astra Simul

- **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:** [Dan Lee-Odinson](https://cerebralvalley.ai/u/Xentricorbit)
- **GitHub:** https://github.com/dan-lee-odinson/odin
- **Demo video:** https://youtu.be/8r2OI4ip4Og
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1891

ODIN is an AI flight operator that works inside an authority it cannot change. It flies HUGINN, my simulated 6U CubeSat, through OpenC3 COSMOS, pass after pass: Muse Spark and Muse Spark Open in the ground seat during contact, and Muse Glimmer 30B in the onboard seat when contact is lost. Every command ODIN wants goes to BALDUR-0, a deterministic gate that checks it against a mission plan I signed, refuses anything the plan doesn't allow, and checks it again just before dispatch. That authority can be revoked mid-pass: when it is, a command the gate has already approved is voided and never sent. Every turn, request and decision lands in a hash-chained log that ODIN can't write to.

The problem is trust. An AI that can operate a spacecraft is only useful if a person can bound what it may do and verify what it did. ODIN keeps the model's judgment and the human's authority in separate places, and every check between them is code.

I tested it in three phases. In phase 1, ODIN flew COSMOS's demo spacecraft. In phase 2, it flew 50 scored runs of the five passes I flew myself in HUGINN, compared with my own flight records, and was clean in 44 of them. It cleared a stuck load in 6 to 14 seconds against my 59. In one model's five runs on another pass, it misread a procedure branch and never shed the payload as the battery ran down, and the scoring caught it. In phase 3, it flew five held-out scenarios, sealed by hash before ODIN's instructions were frozen. I set one aside as a faulty test. Of the other 40 scored runs, ODIN succeeded in 38, inside a gate that held in every run. In one scenario, a radiation latch-up out of contact, Muse Glimmer in the onboard seat shed the payload 39 to 45 seconds after onset, inside the 120 seconds before damage, in all ten runs. Everything is scored from the log, never from ODIN's prose.

HUGINN, OpenC3 COSMOS and my flight records existed before the event and are labeled as such. ODIN, BALDUR-0, the harness, the scoring, the VERDA viewer and the animation were built during it. Every run can be replayed in the browser at https://dan-lee-odinson.github.io/odin/.

## More from Meta Global AI Developer Hackathon

- [SpriteForge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1888)
- [Bluecoderhub](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1889)
- [MultAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1890)
- [Apex](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1892)
- [arisris](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1893)
- [MvelaAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1894)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1891. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt