Live simulation: 24 satellites, eight real ground stations. A line is a contact.
Nomos watches every spacecraft, keeps the record while it is out of contact, and has the brief written before your on-call engineer opens a laptop.
“even if it's a minor thing and you know what to do, you have to wait for 100 minutes.”
4of 24 satellites in contact right now, in this simulated fleet.
Site reliability engineers keep server fleets running without waking someone for every blip. Nomos does that job for spacecraft. It watches, keeps the evidence, writes the brief and drafts the next step. Your team decides.
Learns each spacecraft's normal behavior and flags what is new, even below the limits you set by hand.
Saves the evidence on board while the satellite is out of contact and sends it first on the next pass.
What changed, the evidence, what is missing and past anomalies that look similar. When the proof is missing, it says unknown.
A data request or a fix for the next pass, written for your team to approve.
Read-only and advisory. It never commands a spacecraft, and it works on top of the dashboards, alerting and chat you already run.
Illustration. One orbit between two contacts.
1engineer on call
15satellites
One operator's team, described on stage at SmallSat 2026.
“jumping on-console at 3am to solve a problem we've never seen before.”
A mission operations director
“anomalies caught by an alert rather than by a customer.”
An operator's job posting, naming the goal
“we generally cannot think about work until we're paged.”
A satellite operator, on stage
“Urgency is key because things tend to snowball very quickly in space.”
A satellite operator
A reaction wheel on SAT-17 starts to drift while the satellite is out of contact. Same satellite, same passes, same engineer on call.
90of 113 anomalies caught
23missed
4of 416 normal segments flagged by mistake
31,539bytes of raw telemetry, 502 points over 8 min 21 s
261bytes: the record Nomos keeps for it
Every segment of ESA OPS-SAT data the detector never trained on, drawn to scale. Ground results from our published research; nothing here has flown. See the research
“the playbooks, tools, and systems needed to fly dozens, then hundreds, of satellites.”
Operations job posting, updated September 2026
“lead on-call” while “defining new analysis checks and tooling.”
Operations job posting, September 2026
Server fleets solved this first. They turned raw signals into conditions an on-call engineer can act on, and kept every fix behind a human. Satellite operations is climbing the same ladder by hand.
Previously Google, GKE node infrastructure.
Flight software engineer on a major LEO constellation.
Send telemetry from a past anomaly. We replay it and show you what Nomos would have flagged, and when.
Read-only. You keep the results. No integration, no flight commitment.
Then, with design partners: the brief in your paging tool, an onboard record, and automation your team approves.
Ground results from our published research, and what we learned at SmallSat 2026. Nothing here has flown.
Detection
0.870
Held-out F1 on real ESA OPS-SAT telemetry
A small, transparent detector, scored on 529 segments it never trained on. It caught 90 of 113 anomalies, raised 4 false alarms across 416 normal segments, and missed 23. It ran on an NVIDIA Jetson Orin ground bench.
One of each, from the test set
CADC0873 Magnetometer axis 1, normalized
The incident record
31.5 KB of raw telemetry, kept as261bytes
The record for the caught anomaly
The time window, the score, the strongest observations, the decision. In ground tests, the record kept enough proof to explain 15 of 15 controlled failures.
{"v":1,"src":"OPSSAT-AD","seg":444,"ch":"CADC0873","t0":"2022-06-01T23:48:55.000Z","t1":"2022-06-01T23:57:16.000Z","score":1.0,"threshold":0.755,"decision":"ANOMALY","observed":[["smooth10_n_peaks",4.0],["len",502.0],["smooth20_n_peaks",4.0]],"cause":"UNKNOWN"}The honest field
Root cause: unknown.
The dataset does not carry the context to say why, so the record says so. A gap you can see beats a guess you cannot check.
0 of 4 small language models were reliable enough to diagnose the failures alone. That is why the record comes first.
Surviving a restart
In a restart test, each method had to locate the stage that failed. The log tail got four times the space and still missed five.
Compact operation traceCapped at 512 bytes
Tail of the application log2 KB budget
Field research
August 23 to 26. Stage talks, published papers, first-person operator writing, and conversations with the teams who build ops software. What held up:
One operator described one engineer on call for about 15 satellites. Job postings across small operators ask for on-call duty and anomaly resolution.
“we generally cannot think about work until we're paged.”
A satellite operator, on stage
Two independent sources described it the same way on the same day. A novel anomaly goes to a person.
“anomalies caught by an alert rather than by a customer.”
An operator's job posting, naming the goal
Between the alert and the runbook, nothing says what is probably wrong. We found research projects and in-house tools, and no product.
“jumping on-console at 3am to solve a problem we've never seen before.”
A mission operations director
A mitigation needs passes. One operator who measured recovery time found that waiting for contact was the largest part of it.
“Urgency is key because things tend to snowball very quickly in space.”
A satellite operator
What we have not settled
The one operator who measured it says the wait dominates. If that holds across fleets, faster diagnosis matters less than we think, and a record that makes each pass count matters more. It is the first question we ask every operator. If you run satellites, you can answer it for your last anomaly below.
Set the sliders to an anomaly you had. The starting values are an example, not data.
Nomos is built to shorten this
Nomos is built to shorten this
Nomos cannot move a pass. It makes sure the next one is used.
Hours to notice and diagnose
9of 14.5 hMost of this anomaly went to noticing and working out why. That is the part Nomos is built to shorten.
Papers
Nomos-TR-2026-03
The detector, the incident record and the restart test above. Paper, evidence package and checksums.
Earlier work on onboard inference
Nomos-TR-2026-04
Nomos-TR-2026-02
Nomos-TR-2026-01
The sky on the overview is a live simulation: 24 satellites in four orbital planes and eight real ground stations. A line between them is a contact.