A doctor’s question, asked of the sky
Fast transients fade on timescales from minutes to months: stellar flares, gamma-ray burst afterglows, kilonovae, stars torn apart by black holes. Each fades in its own way, so the right measurement at the right moment tells them apart. Miss the moment and the question closes for good.
OrbitalNet began as a live simulation in which satellites bid for observing work through a Contract Net auction scored on distance and battery. The research replaces that score with the expected information gain over competing explanations, and adds an explicit unknown, h₀, so the planner can recognise something it was never taught.
- Expected information gain
- How much a measurement is expected to reduce our uncertainty about what the event is.
- h₀
- An explicit “none of the above” hypothesis, modelled nonparametrically, so new physics is not forced into a known box.
- The constellation
- Many spacecraft with intermittent links, limited power, and the Earth in the way half the time.
Everything, in one place
python boot_os.py.
Progress
- Oct 2026Phase 1 · Existence proofOpen-world EIG on a toy problem with a withheld class. One seed; a preliminary signal only.Done
- Oct 2026Phase 2 · One planner, real orbitsAll strategies run on identical events, orbits, budgets and contact schedules for each seed.Done
- Oct 2026Phase 3 · Real-data replayArchival events replayed counterfactually. The tuning protocol was written down before any tuning run was seen.In progress
- Nov 2026Pre-registrationCommit the pre-registration, tune on validation seeds only, then run the untouched test seeds.Next
Every number in the paper and the roadmap is generated from files in research/results/. None is typed by hand, and test seeds stay untouched until the pre-registration is committed.