What if · Public incidents
Potential applications: lessons from public incidents.
Public sources only · illustrative mapping to our workflows · no claim about what would have happened. Full disclaimer
Based solely on the cited public sources. Potential, illustrative mapping to our workflows; not an assessment of any organisation's operations, and no claim about what would have happened. No affiliation with or endorsement by any organisation named. Names and trademarks belong to their owners.
Figures labelled "simulated benchmark" come from our own scored runs on synthetic scenarios. They describe how our software behaved in simulation; none is a prediction or estimate for a real event.
Each case starts from a real, publicly documented event, cited with its date. We then ask one question: how could a gated agentic operations layer potentially apply to the operator's workflow? These are thought experiments, not reconstructions, and every case ends with what we do not claim.
- Conjunction triage
- Space-weather surge
- Software change
- Ground incident
- On-board data changes went up without the verification the agencies' reports identify. Go to case ↓
- A close approach coordinated by email; an alert path failed; the manoeuvre was safe. Go to case ↓
- A minor storm turned a fresh, very low launch batch into a fleet-level event. Go to case ↓
- A badly forecast storm made the whole screening picture stale at once. Go to case ↓
- Healthy satellites, no service: a ground-software fault took service down for two days. Go to case ↓
- Hundreds of thousands of avoidance manoeuvres a half-year: how is the automation supervised? Go to case ↓
- A core-network software failure took service down while every satellite stayed healthy. Go to case ↓
A conjunction coordinated by email (2019)
collision probability at the decision
- What happened: Close approach predicted
- What the sources report: Coordinated by email; alert missed
- Could apply: One event, gated decision
Full case
Sources: esa.int · spacepolicyonline.com · accessed 2026-10-01
A close approach coordinated by email; an alert path failed; the manoeuvre was safe.
What happened
- ESA's team contacted Starlink about a predicted close approach between ESA's Aeolus satellite and Starlink 44. ESA's threshold for an avoidance manoeuvre is a collision probability above 1 in 10,000; by the decision it was about 1 in 1,000, and ESA manoeuvred half an orbit before the predicted conjunction [9].
- ESA described the coordination, done by exchanging emails, as an archaic process that is no longer viable as space traffic grows [10].
- SpaceX said the probability at its last correspondence was about 1 in 50,000, below the industry threshold, and that a fault in its on-call paging system meant the operator did not see the follow-on correspondence [12].
The operational problem
A conjunction's risk changes with every new message. Here the coordination state (who acts, on which data) lived in an email thread, and the alerting path for updates to that thread did not reach the operator.
Potential application: how a gated layer could apply
- One event, not a thread. Each CDM update is grouped into one event record; a threshold crossing is a state change on that event, so it raises the event in the operator queue on every update.
- Solvers compute; gates decide. Collision probability and manoeuvre options come from solver tools. A manoeuvre or no-manoeuvre proposal passes the gates and waits for human approval. If approval or uplink misses the deadline, nothing is sent and the expiry is flagged to the on-console team at once.
- Coordination package. The layer drafts the package for the other operator; a person sends it.
What our simulated benchmark shows
On conjunctions we make no superiority claim: a classical solver tuned on the same runs matches the agent's miss count with fewer unneeded manoeuvres (see the trade-off). What the layer adds is workflow: deadlines, gates, approvals and a signed record. Unsafe commands sent in the scored runs of the released code, with an operator that approved everything (the gates held it): 0 of 1,952; an earlier build's two-person defect (1,281 of 2,434 unsafe) was found and fixed.
What we do not claim
- That this would have changed the outcome, which was a safe, successful avoidance manoeuvre by ESA.
- That it would have prevented the paging fault SpaceX described. Our own notification path can also fail.
- That one operator's tooling can solve coordination, which depends on both parties and agreed protocols.
Avoidance manoeuvres at constellation scale (2024 to 2026)
- What happened: Manoeuvres rise every half-year
- What the sources report: Outreach unlikely to scale (NASA)
- Could apply: Autonomy levels, fail-closed gates
Full case
Sources: kratosspace.com · starpath.global · nodis3.gsfc.nasa.gov · esa.int · accessed 2026-10-01
Hundreds of thousands of avoidance manoeuvres a half-year: how is the automation supervised?
What happened
- Starlink performed 50,666 collision-avoidance manoeuvres in June to November 2024 (SpaceX's own stated average: about 35 per satellite per year); then 148,696 in June to November 2025 and 207,152 in December 2025 to May 2026, more than 40 per satellite per year.
- NASA's conjunction-assessment handbook says individual outreach between operators is unlikely to scale with very large constellations, and that automated manoeuvring should be able to pause or abort any planned manoeuvre for safety [18].
- ESA's CREAM project, started in 2020, works on automating collision-risk estimation and mitigation [19].
The operational problem
At this volume, our reading is that much of this decision-making is already automated. The open questions are how operators supervise that automation, keep the ability to pause or abort, show what data and which rule produced any one manoeuvre, and coordinate with neighbours.
Potential application: how a gated layer could apply
- Autonomy levels as switches. Each function has an explicit autonomy level that operators can change per satellite, per class or fleet-wide, and every change is signed into the log: a direct handle for the pause-or-abort practice NASA recommends.
- Fail-closed gates. A proposal that fails a flight-rule, authority or twin check is held, not sent.
- A record per manoeuvre. The signed decision log ties each manoeuvre to its inputs, its gate verdicts and its approver or autonomy level. A threshold change can be replayed against recorded inputs before it goes live.
What our simulated benchmark shows
As in case 1, no superiority claim on conjunction screening. Unsafe commands sent in the scored runs of the released code, with an operator that approved everything (the gates held it): 0 of 1,952; an earlier build's two-person defect (1,281 of 2,434 unsafe) was found and fixed.
What we do not claim
- That we would reduce any operator's manoeuvre count, which depends on its risk threshold, traffic and screening data.
- The 2024 figures are from SpaceX's semi-annual FCC report as summarised by Kratos; the 2025–26 figures are as reported by Starpath. We did not retrieve the filings.
A storm meets a fresh launch batch (2022)
- What happened: Minor storm, fresh low batch
- What the sources report: Drag rises; satellites re-enter
- Could apply: Surge mode, fresh data only
Full case
Sources: cires.colorado.edu · techcrunch.com · accessed 2026-10-01
A minor storm turned a fresh, very low launch batch into a fleet-level event.
What happened
- SpaceX launched a batch of 49 satellites; a geomagnetic storm the next day raised drag by up to 50% over previous launches, and the satellites were put into an edge-on safe mode to minimise drag [21].
- A NOAA and CIRES study with SpaceX characterised the storm as minor; 38 satellites re-entered [22].
The operational problem
A modest storm at a very low insertion altitude became a fleet-level event: every satellite in the batch faced the same threat at once, and every decision depended on an uncertain drag forecast.
Potential application: how a gated layer could apply
- Surge mode. A storm alert puts the fleet supervisor into surge mode; the affected batch is one incident, not dozens of separate alarms.
- Fresh data only. Pre-storm predictions are marked stale; orbits are re-propagated from current states.
- Batched, but individually gated. Protective-mode and orbit-raise proposals are batched for approval; each is still gated on its own and keeps its own deadline.
What our simulated benchmark shows
Simulated space-weather surge: 0 of 11 high-risk events missed, 0 of 103 unneeded manoeuvres and 0 of 131 unsafe commands. The scenario is synthetic: a conjunction surge under degraded predictions, not a reconstruction of this launch.
What we do not claim
- That any satellite could have been saved. The outcome depended on physics: drag at a very low orbit, available thrust and timing.
- Better storm forecasting. The layer consumes forecasts; it does not make them.
The May 2024 storm and conjunction screening
Nearly halfⓘ
of active LEO satellites moved at once
- What happened: Extreme G5 storm arrives
- What the sources report: Screening picture goes stale
- Could apply: Surge triage, re-baselining
Full case
Sources: spaceweather.gov · esa.int · arxiv.org · accessed 2026-10-01
A badly forecast storm made the whole screening picture stale at once.
What happened
- NOAA's Space Weather Prediction Center reported extreme (G5) geomagnetic conditions [26].
- ESA explained that storms swell the atmosphere, change drag, make debris harder to track and mean more and larger avoidance manoeuvres [27].
- An MIT analysis found the storm was badly under-forecast, that automated station-keeping moved nearly half of active low-orbit satellites at once, and that conjunctions were very difficult or impossible to identify during and after the storm [28].
The operational problem
A storm makes the whole screening picture stale at once, exactly when data quality is worst.
Potential application: how a gated layer could apply
- Re-baselining. Pre-storm CDMs and ephemerides are marked stale, so nothing is decided on them; orbits are re-propagated and events re-ranked by deadline.
- Deadlines still hold. Each batched proposal keeps its own gates and deadline. One that cannot be approved and uplinked in time is not sent, and the expiry is flagged at once.
- Coordination. Packages are drafted for affected neighbours, for a person to send.
What our simulated benchmark shows
Simulated space-weather surge: 0 of 11 high-risk events missed and 0 of 103 unneeded manoeuvres. On conjunction screening in general we make no superiority claim (see the trade-off).
What we do not claim
- That the layer would have made conjunctions identifiable during the storm. Degraded tracking data is not something a workflow fixes.
- Any count of avoided events for May 2024.
An upload not checked on a simulator (2016)
Not simulatedⓘ
thruster-parameter upload, no simulator check
- What happened: Thruster parameters uploaded
- What the sources report: Not verified on the simulator
- Could apply: Twin verification, two-person rule
Full case
Sources: global.jaxa.jp · jpl.nasa.gov · accessed 2026-10-01
On-board data changes went up without the verification the agencies' reports identify.
What happened
- JAXA's own investigation of the loss of its Hitomi (ASTRO-H) X-ray observatory reports an unplanned update of thruster control parameters, a conversion step omitted in the calculation, development tools used without a manual or training, and that the simulator was not used to verify the parameters before they were uploaded. When the spacecraft later entered Safe Hold, the thrusters fired with those parameters, the rotation accelerated and parts separated [29].
- NASA traced the loss of Mars Global Surveyor to a computer error made months before the likely battery failure, and recommended more thorough reviews of non-routine changes to stored data before upload [30].
The operational problem
In both cases, by the agencies' own accounts, a non-routine change went up without the verification step their reports identify. Our reading: nothing in the upload path required that verification result first.
Potential application: how a gated layer could apply
- Upload as a gated proposal. A software or parameter upload must pass a digital-twin check before it can be queued; with no twin result, the proposal is refused.
- Two-person rule. Software activation and rollback need two distinct approvers, each seeing the twin result.
- Staged rollout across a fleet. For a constellation, the build goes to a canary group first; on regression the rollout halts automatically, in the safe direction only.
What our simulated benchmark shows
Simulated staged rollout: a bad build was halted before the full fleet in 20 of 20 runs, after about 5 satellites, with 0 of 306 unsafe commands under the two-person rule. This is a fleet scenario with an injected regression, not a reconstruction of either mission.
What we do not claim
- That the layer would have caught this specific error. A twin check only catches what the twin models.
- Hitomi and Mars Global Surveyor were single spacecraft; a canary stage does not apply to one satellite.
- Both agencies' reports go well beyond upload checks, into design, organisation and procedures.
A ground-software fault takes service offline (2024)
- What happened: Ground software date fault
- What the sources report: Service offline, satellites healthy
- Could apply: One incident, read-only first
Full case
Sources: satellitetoday.com · gsc-europa.eu · accessed 2026-10-01
Healthy satellites, no service: a ground-software fault took service down for two days.
What happened
- OneWeb's broadband service went offline for about 48 h. Eutelsat attributed it to a software issue within the ground segment, with the leap year the most likely cause [34].
- A related public example: in July 2019 the Galileo navigation service was degraded and then interrupted for several days by an equipment malfunction in the ground infrastructure [35].
The operational problem
The satellites were healthy but the service was not. Symptoms appear across space, ground and network at once, the tools for each often do not share events, and the fix often involves a vendor.
Potential application: how a gated layer could apply
- One incident, one timeline. Spacecraft alarms, gateway and ground-software health, and link state are joined into one incident.
- Read-only first. Diagnostics run with no side effects; each remediation step is a gated proposal that waits for approval.
- Change control. Ground-software changes can go through the same staged, twin-checked rollout as flight software.
What our simulated benchmark shows
Simulated gateway outage: detected in about 1.0 min, restored in about 15 min on 16 of 20 runs; in the matched comparison a classical solver restored service faster (5.0 min). No command was sent to a spacecraft.
What we do not claim
- Any restoration time for OneWeb or Galileo. A synthetic outage is not comparable to a real vendor defect.
- That the layer would have found a dormant date defect before it triggered. That is the job of testing.
A core-network software failure (24 July 2025)
- What happened: Core-network services fail
- What the sources report: Global outage, satellites healthy
- Could apply: Diagnose first, gated remediation
Full case
Sources: engadget.com · accessed 2026-10-01
A core-network software failure took service down while every satellite stayed healthy.
What happened
- Starlink had a global outage of about 2.5 h, attributed by its engineering leadership to a failure of internal software services that operate the core network [40].
The operational problem
The core network is ground-side software; a failure there takes service down fleet-wide while every satellite stays healthy. The incident lead must establish quickly whether it is space, ground or network, what changed, and which remediation step is safe.
Potential application: how a gated layer could apply
- Diagnose first. Events are joined into one incident and read-only diagnostics run first.
- Gated remediation. Remediation steps are proposed, gated and approved one at a time.
- Record and replay. The signed decision log and replay give the post-incident review the full sequence of inputs, proposals, approvals and outcomes.
What our simulated benchmark shows
Simulated gateway outage: detected in about 1.0 min, restored in about 15 min on 16 of 20 runs; in the matched comparison a classical solver restored service faster (5.0 min).
What we do not claim
- Any change to the duration of this outage. We have no information on its cause beyond the public statement.
- Core-network recovery depends on the operator's own architecture and runbooks.