A spacecraft (rover, orbiter, or a JWST-class observatory) records far more raw data than it can ever radio home — the Deep Space Network gives it only a short, scheduled contact window per orbit or per day. NASA's Perseverance already runs this triage onboard with its AEGIS targeting system, and it is the same problem future observatories face when deciding which spectra or image cutouts are worth the bits.
Each observation is scored the way an onboard anomaly detector (e.g. an autoencoder's reconstruction error) would score it — a mixture of routine noise and rare, high-value spikes:
score = 0.15 + 0.20·U (85% of packets: routine)
score = 0.60 + 0.40·U (15% of packets: anomalous / novel)
where U ~ Uniform(0,1)
When a contact window opens, the AI-triage scheduler solves an approximate knapsack: keep only packets with score ≥ threshold, sort the rest by value density, and fill the budget greedily —
density_i = score_i / size_i
sort packets by density_i, descending
add packet i while Σ size ≤ budget (skip if it would overflow)
This greedy-by-density rule is what real flight computers use in place of an exact 0/1 knapsack solve — optimal packing is combinatorially expensive, and density-greedy gets within a few percent of it in practice while running in microseconds. FIFO mode disables the scorer entirely and downlinks whichever packets arrived first, exactly as a naive store-and-forward pipeline would — watch how much science value it leaves behind, and how the backlog overflows and starts discarding unsent data once the observation rate outpaces the budget.
- Downlink budget — total Mbit the ground-station pass can carry; too low and even AI triage can't keep the backlog empty.
- Observation rate — how fast new science packets are generated onboard.
- Anomaly threshold — in AI mode, the minimum score worth spending bandwidth on; raise it to be pickier.