A four-tonne piece of a SpaceX rocket spent eighteen months on a collision course with the Moon, starting at roughly the moment everyone stopped thinking about it. On August 5 it arrived.
The interesting part is not the collision. It is that the collision was scheduled eighteen months in advance by a decision nobody remembers making.
Yes. A spent SpaceX Falcon 9 upper stage crashed into the Moon on August 5, 2026, at about 06:34 UTC — 2:34 a.m. EDT — near the crater Einstein on the western limb. South Korea's Danuri orbiter confirmed the lunar impact, imaging the site across eight passes that began roughly thirty minutes before contact. The Korea Aerospace Research Institute reported that "changes in the terrain around the impact site and traces of ejecta spread were confirmed." At Lowell Observatory in Arizona, Carl Schmidt's team measured a sodium and lithium gas plume tens of kilometres across that lasted five to ten minutes.
What follows is not a space story. It is what the crash says about every system that outlives the attention paid to it.
The object has no name, only a catalogue designation: 2025-010D. It was the fourth tracked piece of hardware from the tenth orbital launch of 2025, on January 15, which carried Firefly Aerospace's Blue Ghost lander and ispace's Resilience toward the Moon. Blue Ghost landed successfully that March. Resilience lost contact about ninety seconds before touchdown and crashed in June.
Falcon 9 upper stages are normally deorbited shortly after launch and burned up in the atmosphere. That did not happen here. Pushing two landers to their deployment point consumed nearly all of the stage's propellant, so the disposal burn was no longer available. In the words of a study led by Benjamin Fernando of Los Alamos National Laboratory, the stage "was abandoned in a moon-crossing high-Earth orbit."
It then spent more than eighteen months in a lopsided 26-day orbit, swinging between roughly 220,000 and 510,000 km from Earth — repeatedly crossing the Moon's path, missing it every time, until it didn't. Astronomer Bill Gray of Project Pluto put the impact speed at 2.43 km/s and the released energy at about 14.5 billion joules, roughly three tons of TNT.
Nobody chose to crash a rocket into the Moon. The choice that mattered was made much earlier, and it looked like a fuel budget.
Decommissioning was the line item that got consumed to make the primary mission work. There was no meeting about it, no risk register entry, no dissenting engineer. The mission was harder than the margin allowed, the margin came out of end-of-life, and the end-of-life plan quietly ceased to exist.
This is the most common way software systems become permanent. The migration runs long, so the cutover eats the quarter, so the decommissioning of the old system slips to "next sprint." The old system keeps serving a trickle of traffic. Two years later it is load-bearing and nobody on the team has read its code.
The failure mode is not negligence. It is that decommissioning is the only part of a project whose deadline nothing enforces.
SpaceX did not simply walk away. Julianna Scheiman, the company's director of NASA science and Dragon programs, said a different maneuver was performed "to ensure that the second stage itself is safe per the appropriate rules and regulations."
That maneuver satisfied the rules. It did not put the object anywhere durable. As Scheiman put it, "a mixture of solar activity and gravity forces have put it on a path towards the moon."
A control that passes review at go-live is a statement about go-live. It is not a statement about the system's state eighteen months later, and treating the two as equivalent is how model risk accumulates. Sign-off is a snapshot; systems are a trajectory.
The force that moved it was sunlight. Solar radiation pressure — photons striking the tumbling stage — is, in Gray's description, "an extremely gentle force, but over months, it can really build up." It is also not fully predictable, because a tumbling object catches and reflects light differently as it turns.
No single push mattered. Eighteen months of them decided where four-and-a-bit tonnes of metal ended up.
That is drift, exactly as it appears in a scoring pipeline: no individual week's population shift is worth an alert, and the cumulative eighteen-month displacement is worth a regulatory finding. If your validation cadence is slower than your drift accumulation, you are not monitoring the model. You are just visiting it occasionally.
Here is the detail we found most uncomfortable. This object was tracked mainly by asteroid surveys and amateur astronomers, because the systems built for space debris cannot see that far out. Radar excels in low Earth orbit — Gray notes it has tracked gloves and tool bags lost by astronauts — but at lunar distance the return signal is about 25.6 billion times fainter. Gray computes orbits for high-altitude objects specifically because they are the ones "the military doesn't track."
The surveys that did see it were not trying to. Their job is planetary defence, and as Gray puts it, "time spent observing junk is time not spent finding rocks."
Every engineering organisation has this arrangement and calls it something else. The deprecated service has no dashboard and no on-call rotation, and the only reason anyone knows it is degrading is that one engineer keeps noticing it in unrelated logs. That is not observability. That is a volunteer, and volunteers leave.
The recovery here is genuinely impressive, and it was made of data. By late February the archive held 1,053 observations, contributed by observatories in Mississippi, Utah, Beijing and England. Gray's software flagged the lunar impact in September 2025 — eleven months ahead.
The forecast then tightened as observations accumulated. A June revision moved the predicted impact nine minutes earlier and 170 km to the northwest. A July revision moved it by a couple of seconds and about ten kilometres. A final update landed on August 1.
That precision had a payoff: Danuri began imaging roughly thirty minutes before the collision and made eight passes over the site, capturing genuine before-and-after coverage. You only point an instrument at the right patch of ground at the right minute if someone gave you a forecast. And you only get a forecast for a system somebody is still measuring.
Worth noting, in the spirit of not overclaiming: as of this writing there is no published measurement of the new crater. Gray predicted a single crater around 17 m across. Much of the news coverage cites up to 27 m. Both are pre-impact calculations, and they do not agree with each other.
Whether it is one crater or two is also unsettled. The 2022 Chang'e 5-T1 upper stage surprised everyone by making a double crater. Gray expects a single one here because this stage carried no comparable mass at the top — a reasonable inference, not a result.
Numbers in circulation are not measurements. That distinction is the whole job in model validation, and it does not stop being true because a number is repeated confidently by everyone.
Translate the failure modes and they get uncomfortably familiar.
| What happened to 2025-010D |
The equivalent in a lending stack |
| Disposal fuel spent on the primary mission |
Decommissioning descoped to protect a launch date |
| Compliant maneuver, then no owner |
Control signed off at go-live, never re-verified |
| Solar radiation pressure over 18 months |
Population drift between validation cycles |
| Invisible to radar at lunar distance |
No telemetry on the deprecated service |
| Tracked by volunteers with other jobs |
One engineer who happens to know it exists |
Concretely, the things worth auditing this quarter:
- Retired scorecards that still score. A superseded model version left wired into a fallback path, quietly deciding a slice of applications.
- Orphaned document stores. Buckets of customer bank statements from a pilot that ended. Nobody owns them, retention policy never applied, and they are pure liability.
- Extraction pipelines with one remaining caller. The old parser kept alive for a single lender integration, unmaintained and unmonitored.
- Credentials outliving their purpose. API keys for a data provider you stopped paying, still valid, still in a config somewhere.
- Connectors for formats that changed. A bank redesigned its statement layout; the connector now fails silently on a small percentage and nobody set an alarm.
None of these will page you. That is precisely the problem: they fail in ways that produce no signal until the failure is someone else's finding.
The fix, in orbital mechanics, is cheap — if you decide early. Put the upper stage on a trajectory that leaves the Earth-Moon system entirely and ends up orbiting the Sun. The Falcon 9 stage from the EscaPADE launch in November 2025 was disposed of this way. Chinese lunar launches have moved in the same direction.
But it only works as a design-time decision. Once the propellant is gone, no option exists at any price. Gray is also honest about the limit: the trick "only works for a small fraction of junk," namely missions already headed beyond Earth orbit. Some classes of system genuinely cannot be cleanly retired, and the time to find out which class you are in is before you build.
The question Scheiman raised is the right one to steal: "What is the best future disposal path for high-energy missions that are in the sun-Earth-moon system?" The software version is smaller and duller and just as unanswered. What is the retirement path for this model, this pipeline, this data store? Who owns it after the launch party? What signal tells us it is still running, and who receives that signal?
Systems you cannot see are systems you cannot retire. Levmetrics' Lev Document Intelligence is built so that every extraction, score and decision leaves an auditable trace — including the ones you meant to turn off. Try it at app.levmetrics.com.
A four-tonne rocket stage is a useful teacher because it is impossible to lose track of politely. It kept an orbit, it kept a catalogue number, and eighteen months after everyone declared the mission a success it arrived somewhere nobody intended at 2.43 kilometres per second. Most abandoned systems are less dramatic and more expensive. Deployment is a milestone in a system's life. It is nowhere near the end of it.
- Mike Wall, "A SpaceX rocket will crash into the moon this week. Why is it on a collision course?", Space.com, August 3, 2026.
- Chelsea Gohd, "A SpaceX rocket just crashed into the moon. Now, the race is on to get a look at the impact site", Space.com, August 5, 2026.
- Mike Wall, "SpaceX rocket's impact crater on the moon spied by Korean lunar orbiter (photos)", Space.com, August 6, 2026.
- Bill Gray, "Upper stage impacting the moon on 2026 August 5", Project Pluto.
- Adam Kovac, "A SpaceX rocket booster is on track to hit the moon at several times the speed of sound", Scientific American, May 1, 2026.
- Leonard David, "Moon Impact of Rocket Stage Observed", Inside Outer Space, August 5, 2026.
- Benjamin Fernando et al., "preprint on the predicted lunar impact of the 2025-010D upper stage", arXiv.
A four-tonne piece of a SpaceX rocket spent eighteen months on a collision course with the Moon, starting at roughly the moment everyone stopped thinking about it. On August 5 it arrived.
The interesting part is not the collision. It is that the collision was scheduled eighteen months in advance by a decision nobody remembers making.
Yes. A spent SpaceX Falcon 9 upper stage crashed into the Moon on August 5, 2026, at about 06:34 UTC — 2:34 a.m. EDT — near the crater Einstein on the western limb. South Korea's Danuri orbiter confirmed the lunar impact, imaging the site across eight passes that began roughly thirty minutes before contact. The Korea Aerospace Research Institute reported that "changes in the terrain around the impact site and traces of ejecta spread were confirmed." At Lowell Observatory in Arizona, Carl Schmidt's team measured a sodium and lithium gas plume tens of kilometres across that lasted five to ten minutes.
What follows is not a space story. It is what the crash says about every system that outlives the attention paid to it.
The object has no name, only a catalogue designation: 2025-010D. It was the fourth tracked piece of hardware from the tenth orbital launch of 2025, on January 15, which carried Firefly Aerospace's Blue Ghost lander and ispace's Resilience toward the Moon. Blue Ghost landed successfully that March. Resilience lost contact about ninety seconds before touchdown and crashed in June.
Falcon 9 upper stages are normally deorbited shortly after launch and burned up in the atmosphere. That did not happen here. Pushing two landers to their deployment point consumed nearly all of the stage's propellant, so the disposal burn was no longer available. In the words of a study led by Benjamin Fernando of Los Alamos National Laboratory, the stage "was abandoned in a moon-crossing high-Earth orbit."
It then spent more than eighteen months in a lopsided 26-day orbit, swinging between roughly 220,000 and 510,000 km from Earth — repeatedly crossing the Moon's path, missing it every time, until it didn't. Astronomer Bill Gray of Project Pluto put the impact speed at 2.43 km/s and the released energy at about 14.5 billion joules, roughly three tons of TNT.
Nobody chose to crash a rocket into the Moon. The choice that mattered was made much earlier, and it looked like a fuel budget.
Decommissioning was the line item that got consumed to make the primary mission work. There was no meeting about it, no risk register entry, no dissenting engineer. The mission was harder than the margin allowed, the margin came out of end-of-life, and the end-of-life plan quietly ceased to exist.
This is the most common way software systems become permanent. The migration runs long, so the cutover eats the quarter, so the decommissioning of the old system slips to "next sprint." The old system keeps serving a trickle of traffic. Two years later it is load-bearing and nobody on the team has read its code.
The failure mode is not negligence. It is that decommissioning is the only part of a project whose deadline nothing enforces.
SpaceX did not simply walk away. Julianna Scheiman, the company's director of NASA science and Dragon programs, said a different maneuver was performed "to ensure that the second stage itself is safe per the appropriate rules and regulations."
That maneuver satisfied the rules. It did not put the object anywhere durable. As Scheiman put it, "a mixture of solar activity and gravity forces have put it on a path towards the moon."
A control that passes review at go-live is a statement about go-live. It is not a statement about the system's state eighteen months later, and treating the two as equivalent is how model risk accumulates. Sign-off is a snapshot; systems are a trajectory.
The force that moved it was sunlight. Solar radiation pressure — photons striking the tumbling stage — is, in Gray's description, "an extremely gentle force, but over months, it can really build up." It is also not fully predictable, because a tumbling object catches and reflects light differently as it turns.
No single push mattered. Eighteen months of them decided where four-and-a-bit tonnes of metal ended up.
That is drift, exactly as it appears in a scoring pipeline: no individual week's population shift is worth an alert, and the cumulative eighteen-month displacement is worth a regulatory finding. If your validation cadence is slower than your drift accumulation, you are not monitoring the model. You are just visiting it occasionally.
Here is the detail we found most uncomfortable. This object was tracked mainly by asteroid surveys and amateur astronomers, because the systems built for space debris cannot see that far out. Radar excels in low Earth orbit — Gray notes it has tracked gloves and tool bags lost by astronauts — but at lunar distance the return signal is about 25.6 billion times fainter. Gray computes orbits for high-altitude objects specifically because they are the ones "the military doesn't track."
The surveys that did see it were not trying to. Their job is planetary defence, and as Gray puts it, "time spent observing junk is time not spent finding rocks."
Every engineering organisation has this arrangement and calls it something else. The deprecated service has no dashboard and no on-call rotation, and the only reason anyone knows it is degrading is that one engineer keeps noticing it in unrelated logs. That is not observability. That is a volunteer, and volunteers leave.
The recovery here is genuinely impressive, and it was made of data. By late February the archive held 1,053 observations, contributed by observatories in Mississippi, Utah, Beijing and England. Gray's software flagged the lunar impact in September 2025 — eleven months ahead.
The forecast then tightened as observations accumulated. A June revision moved the predicted impact nine minutes earlier and 170 km to the northwest. A July revision moved it by a couple of seconds and about ten kilometres. A final update landed on August 1.
That precision had a payoff: Danuri began imaging roughly thirty minutes before the collision and made eight passes over the site, capturing genuine before-and-after coverage. You only point an instrument at the right patch of ground at the right minute if someone gave you a forecast. And you only get a forecast for a system somebody is still measuring.
Worth noting, in the spirit of not overclaiming: as of this writing there is no published measurement of the new crater. Gray predicted a single crater around 17 m across. Much of the news coverage cites up to 27 m. Both are pre-impact calculations, and they do not agree with each other.
Whether it is one crater or two is also unsettled. The 2022 Chang'e 5-T1 upper stage surprised everyone by making a double crater. Gray expects a single one here because this stage carried no comparable mass at the top — a reasonable inference, not a result.
Numbers in circulation are not measurements. That distinction is the whole job in model validation, and it does not stop being true because a number is repeated confidently by everyone.
Translate the failure modes and they get uncomfortably familiar.
| What happened to 2025-010D |
The equivalent in a lending stack |
| Disposal fuel spent on the primary mission |
Decommissioning descoped to protect a launch date |
| Compliant maneuver, then no owner |
Control signed off at go-live, never re-verified |
| Solar radiation pressure over 18 months |
Population drift between validation cycles |
| Invisible to radar at lunar distance |
No telemetry on the deprecated service |
| Tracked by volunteers with other jobs |
One engineer who happens to know it exists |
Concretely, the things worth auditing this quarter:
- Retired scorecards that still score. A superseded model version left wired into a fallback path, quietly deciding a slice of applications.
- Orphaned document stores. Buckets of customer bank statements from a pilot that ended. Nobody owns them, retention policy never applied, and they are pure liability.
- Extraction pipelines with one remaining caller. The old parser kept alive for a single lender integration, unmaintained and unmonitored.
- Credentials outliving their purpose. API keys for a data provider you stopped paying, still valid, still in a config somewhere.
- Connectors for formats that changed. A bank redesigned its statement layout; the connector now fails silently on a small percentage and nobody set an alarm.
None of these will page you. That is precisely the problem: they fail in ways that produce no signal until the failure is someone else's finding.
The fix, in orbital mechanics, is cheap — if you decide early. Put the upper stage on a trajectory that leaves the Earth-Moon system entirely and ends up orbiting the Sun. The Falcon 9 stage from the EscaPADE launch in November 2025 was disposed of this way. Chinese lunar launches have moved in the same direction.
But it only works as a design-time decision. Once the propellant is gone, no option exists at any price. Gray is also honest about the limit: the trick "only works for a small fraction of junk," namely missions already headed beyond Earth orbit. Some classes of system genuinely cannot be cleanly retired, and the time to find out which class you are in is before you build.
The question Scheiman raised is the right one to steal: "What is the best future disposal path for high-energy missions that are in the sun-Earth-moon system?" The software version is smaller and duller and just as unanswered. What is the retirement path for this model, this pipeline, this data store? Who owns it after the launch party? What signal tells us it is still running, and who receives that signal?
Systems you cannot see are systems you cannot retire. Levmetrics' Lev Document Intelligence is built so that every extraction, score and decision leaves an auditable trace — including the ones you meant to turn off. Try it at app.levmetrics.com.
A four-tonne rocket stage is a useful teacher because it is impossible to lose track of politely. It kept an orbit, it kept a catalogue number, and eighteen months after everyone declared the mission a success it arrived somewhere nobody intended at 2.43 kilometres per second. Most abandoned systems are less dramatic and more expensive. Deployment is a milestone in a system's life. It is nowhere near the end of it.
- Mike Wall, "A SpaceX rocket will crash into the moon this week. Why is it on a collision course?", Space.com, August 3, 2026.
- Chelsea Gohd, "A SpaceX rocket just crashed into the moon. Now, the race is on to get a look at the impact site", Space.com, August 5, 2026.
- Mike Wall, "SpaceX rocket's impact crater on the moon spied by Korean lunar orbiter (photos)", Space.com, August 6, 2026.
- Bill Gray, "Upper stage impacting the moon on 2026 August 5", Project Pluto.
- Adam Kovac, "A SpaceX rocket booster is on track to hit the moon at several times the speed of sound", Scientific American, May 1, 2026.
- Leonard David, "Moon Impact of Rocket Stage Observed", Inside Outer Space, August 5, 2026.
- Benjamin Fernando et al., "preprint on the predicted lunar impact of the 2025-010D upper stage", arXiv.
A four-tonne piece of a SpaceX rocket spent eighteen months on a collision course with the Moon, starting at roughly the moment everyone stopped thinking about it. On August 5 it arrived.
The interesting part is not the collision. It is that the collision was scheduled eighteen months in advance by a decision nobody remembers making.
Yes. A spent SpaceX Falcon 9 upper stage crashed into the Moon on August 5, 2026, at about 06:34 UTC — 2:34 a.m. EDT — near the crater Einstein on the western limb. South Korea's Danuri orbiter confirmed the lunar impact, imaging the site across eight passes that began roughly thirty minutes before contact. The Korea Aerospace Research Institute reported that "changes in the terrain around the impact site and traces of ejecta spread were confirmed." At Lowell Observatory in Arizona, Carl Schmidt's team measured a sodium and lithium gas plume tens of kilometres across that lasted five to ten minutes.
What follows is not a space story. It is what the crash says about every system that outlives the attention paid to it.
The object has no name, only a catalogue designation: 2025-010D. It was the fourth tracked piece of hardware from the tenth orbital launch of 2025, on January 15, which carried Firefly Aerospace's Blue Ghost lander and ispace's Resilience toward the Moon. Blue Ghost landed successfully that March. Resilience lost contact about ninety seconds before touchdown and crashed in June.
Falcon 9 upper stages are normally deorbited shortly after launch and burned up in the atmosphere. That did not happen here. Pushing two landers to their deployment point consumed nearly all of the stage's propellant, so the disposal burn was no longer available. In the words of a study led by Benjamin Fernando of Los Alamos National Laboratory, the stage "was abandoned in a moon-crossing high-Earth orbit."
It then spent more than eighteen months in a lopsided 26-day orbit, swinging between roughly 220,000 and 510,000 km from Earth — repeatedly crossing the Moon's path, missing it every time, until it didn't. Astronomer Bill Gray of Project Pluto put the impact speed at 2.43 km/s and the released energy at about 14.5 billion joules, roughly three tons of TNT.
Nobody chose to crash a rocket into the Moon. The choice that mattered was made much earlier, and it looked like a fuel budget.
Decommissioning was the line item that got consumed to make the primary mission work. There was no meeting about it, no risk register entry, no dissenting engineer. The mission was harder than the margin allowed, the margin came out of end-of-life, and the end-of-life plan quietly ceased to exist.
This is the most common way software systems become permanent. The migration runs long, so the cutover eats the quarter, so the decommissioning of the old system slips to "next sprint." The old system keeps serving a trickle of traffic. Two years later it is load-bearing and nobody on the team has read its code.
The failure mode is not negligence. It is that decommissioning is the only part of a project whose deadline nothing enforces.
SpaceX did not simply walk away. Julianna Scheiman, the company's director of NASA science and Dragon programs, said a different maneuver was performed "to ensure that the second stage itself is safe per the appropriate rules and regulations."
That maneuver satisfied the rules. It did not put the object anywhere durable. As Scheiman put it, "a mixture of solar activity and gravity forces have put it on a path towards the moon."
A control that passes review at go-live is a statement about go-live. It is not a statement about the system's state eighteen months later, and treating the two as equivalent is how model risk accumulates. Sign-off is a snapshot; systems are a trajectory.
The force that moved it was sunlight. Solar radiation pressure — photons striking the tumbling stage — is, in Gray's description, "an extremely gentle force, but over months, it can really build up." It is also not fully predictable, because a tumbling object catches and reflects light differently as it turns.
No single push mattered. Eighteen months of them decided where four-and-a-bit tonnes of metal ended up.
That is drift, exactly as it appears in a scoring pipeline: no individual week's population shift is worth an alert, and the cumulative eighteen-month displacement is worth a regulatory finding. If your validation cadence is slower than your drift accumulation, you are not monitoring the model. You are just visiting it occasionally.
Here is the detail we found most uncomfortable. This object was tracked mainly by asteroid surveys and amateur astronomers, because the systems built for space debris cannot see that far out. Radar excels in low Earth orbit — Gray notes it has tracked gloves and tool bags lost by astronauts — but at lunar distance the return signal is about 25.6 billion times fainter. Gray computes orbits for high-altitude objects specifically because they are the ones "the military doesn't track."
The surveys that did see it were not trying to. Their job is planetary defence, and as Gray puts it, "time spent observing junk is time not spent finding rocks."
Every engineering organisation has this arrangement and calls it something else. The deprecated service has no dashboard and no on-call rotation, and the only reason anyone knows it is degrading is that one engineer keeps noticing it in unrelated logs. That is not observability. That is a volunteer, and volunteers leave.
The recovery here is genuinely impressive, and it was made of data. By late February the archive held 1,053 observations, contributed by observatories in Mississippi, Utah, Beijing and England. Gray's software flagged the lunar impact in September 2025 — eleven months ahead.
The forecast then tightened as observations accumulated. A June revision moved the predicted impact nine minutes earlier and 170 km to the northwest. A July revision moved it by a couple of seconds and about ten kilometres. A final update landed on August 1.
That precision had a payoff: Danuri began imaging roughly thirty minutes before the collision and made eight passes over the site, capturing genuine before-and-after coverage. You only point an instrument at the right patch of ground at the right minute if someone gave you a forecast. And you only get a forecast for a system somebody is still measuring.
Worth noting, in the spirit of not overclaiming: as of this writing there is no published measurement of the new crater. Gray predicted a single crater around 17 m across. Much of the news coverage cites up to 27 m. Both are pre-impact calculations, and they do not agree with each other.
Whether it is one crater or two is also unsettled. The 2022 Chang'e 5-T1 upper stage surprised everyone by making a double crater. Gray expects a single one here because this stage carried no comparable mass at the top — a reasonable inference, not a result.
Numbers in circulation are not measurements. That distinction is the whole job in model validation, and it does not stop being true because a number is repeated confidently by everyone.
Translate the failure modes and they get uncomfortably familiar.
| What happened to 2025-010D |
The equivalent in a lending stack |
| Disposal fuel spent on the primary mission |
Decommissioning descoped to protect a launch date |
| Compliant maneuver, then no owner |
Control signed off at go-live, never re-verified |
| Solar radiation pressure over 18 months |
Population drift between validation cycles |
| Invisible to radar at lunar distance |
No telemetry on the deprecated service |
| Tracked by volunteers with other jobs |
One engineer who happens to know it exists |
Concretely, the things worth auditing this quarter:
- Retired scorecards that still score. A superseded model version left wired into a fallback path, quietly deciding a slice of applications.
- Orphaned document stores. Buckets of customer bank statements from a pilot that ended. Nobody owns them, retention policy never applied, and they are pure liability.
- Extraction pipelines with one remaining caller. The old parser kept alive for a single lender integration, unmaintained and unmonitored.
- Credentials outliving their purpose. API keys for a data provider you stopped paying, still valid, still in a config somewhere.
- Connectors for formats that changed. A bank redesigned its statement layout; the connector now fails silently on a small percentage and nobody set an alarm.
None of these will page you. That is precisely the problem: they fail in ways that produce no signal until the failure is someone else's finding.
The fix, in orbital mechanics, is cheap — if you decide early. Put the upper stage on a trajectory that leaves the Earth-Moon system entirely and ends up orbiting the Sun. The Falcon 9 stage from the EscaPADE launch in November 2025 was disposed of this way. Chinese lunar launches have moved in the same direction.
But it only works as a design-time decision. Once the propellant is gone, no option exists at any price. Gray is also honest about the limit: the trick "only works for a small fraction of junk," namely missions already headed beyond Earth orbit. Some classes of system genuinely cannot be cleanly retired, and the time to find out which class you are in is before you build.
The question Scheiman raised is the right one to steal: "What is the best future disposal path for high-energy missions that are in the sun-Earth-moon system?" The software version is smaller and duller and just as unanswered. What is the retirement path for this model, this pipeline, this data store? Who owns it after the launch party? What signal tells us it is still running, and who receives that signal?
Systems you cannot see are systems you cannot retire. Levmetrics' Lev Document Intelligence is built so that every extraction, score and decision leaves an auditable trace — including the ones you meant to turn off. Try it at app.levmetrics.com.
A four-tonne rocket stage is a useful teacher because it is impossible to lose track of politely. It kept an orbit, it kept a catalogue number, and eighteen months after everyone declared the mission a success it arrived somewhere nobody intended at 2.43 kilometres per second. Most abandoned systems are less dramatic and more expensive. Deployment is a milestone in a system's life. It is nowhere near the end of it.
- Mike Wall, "A SpaceX rocket will crash into the moon this week. Why is it on a collision course?", Space.com, August 3, 2026.
- Chelsea Gohd, "A SpaceX rocket just crashed into the moon. Now, the race is on to get a look at the impact site", Space.com, August 5, 2026.
- Mike Wall, "SpaceX rocket's impact crater on the moon spied by Korean lunar orbiter (photos)", Space.com, August 6, 2026.
- Bill Gray, "Upper stage impacting the moon on 2026 August 5", Project Pluto.
- Adam Kovac, "A SpaceX rocket booster is on track to hit the moon at several times the speed of sound", Scientific American, May 1, 2026.
- Leonard David, "Moon Impact of Rocket Stage Observed", Inside Outer Space, August 5, 2026.
- Benjamin Fernando et al., "preprint on the predicted lunar impact of the 2025-010D upper stage", arXiv.
A four-tonne piece of a SpaceX rocket spent eighteen months on a collision course with the Moon, starting at roughly the moment everyone stopped thinking about it. On August 5 it arrived.
The interesting part is not the collision. It is that the collision was scheduled eighteen months in advance by a decision nobody remembers making.
Yes. A spent SpaceX Falcon 9 upper stage crashed into the Moon on August 5, 2026, at about 06:34 UTC — 2:34 a.m. EDT — near the crater Einstein on the western limb. South Korea's Danuri orbiter confirmed the lunar impact, imaging the site across eight passes that began roughly thirty minutes before contact. The Korea Aerospace Research Institute reported that "changes in the terrain around the impact site and traces of ejecta spread were confirmed." At Lowell Observatory in Arizona, Carl Schmidt's team measured a sodium and lithium gas plume tens of kilometres across that lasted five to ten minutes.
What follows is not a space story. It is what the crash says about every system that outlives the attention paid to it.
The object has no name, only a catalogue designation: 2025-010D. It was the fourth tracked piece of hardware from the tenth orbital launch of 2025, on January 15, which carried Firefly Aerospace's Blue Ghost lander and ispace's Resilience toward the Moon. Blue Ghost landed successfully that March. Resilience lost contact about ninety seconds before touchdown and crashed in June.
Falcon 9 upper stages are normally deorbited shortly after launch and burned up in the atmosphere. That did not happen here. Pushing two landers to their deployment point consumed nearly all of the stage's propellant, so the disposal burn was no longer available. In the words of a study led by Benjamin Fernando of Los Alamos National Laboratory, the stage "was abandoned in a moon-crossing high-Earth orbit."
It then spent more than eighteen months in a lopsided 26-day orbit, swinging between roughly 220,000 and 510,000 km from Earth — repeatedly crossing the Moon's path, missing it every time, until it didn't. Astronomer Bill Gray of Project Pluto put the impact speed at 2.43 km/s and the released energy at about 14.5 billion joules, roughly three tons of TNT.
Nobody chose to crash a rocket into the Moon. The choice that mattered was made much earlier, and it looked like a fuel budget.
Decommissioning was the line item that got consumed to make the primary mission work. There was no meeting about it, no risk register entry, no dissenting engineer. The mission was harder than the margin allowed, the margin came out of end-of-life, and the end-of-life plan quietly ceased to exist.
This is the most common way software systems become permanent. The migration runs long, so the cutover eats the quarter, so the decommissioning of the old system slips to "next sprint." The old system keeps serving a trickle of traffic. Two years later it is load-bearing and nobody on the team has read its code.
The failure mode is not negligence. It is that decommissioning is the only part of a project whose deadline nothing enforces.
SpaceX did not simply walk away. Julianna Scheiman, the company's director of NASA science and Dragon programs, said a different maneuver was performed "to ensure that the second stage itself is safe per the appropriate rules and regulations."
That maneuver satisfied the rules. It did not put the object anywhere durable. As Scheiman put it, "a mixture of solar activity and gravity forces have put it on a path towards the moon."
A control that passes review at go-live is a statement about go-live. It is not a statement about the system's state eighteen months later, and treating the two as equivalent is how model risk accumulates. Sign-off is a snapshot; systems are a trajectory.
The force that moved it was sunlight. Solar radiation pressure — photons striking the tumbling stage — is, in Gray's description, "an extremely gentle force, but over months, it can really build up." It is also not fully predictable, because a tumbling object catches and reflects light differently as it turns.
No single push mattered. Eighteen months of them decided where four-and-a-bit tonnes of metal ended up.
That is drift, exactly as it appears in a scoring pipeline: no individual week's population shift is worth an alert, and the cumulative eighteen-month displacement is worth a regulatory finding. If your validation cadence is slower than your drift accumulation, you are not monitoring the model. You are just visiting it occasionally.
Here is the detail we found most uncomfortable. This object was tracked mainly by asteroid surveys and amateur astronomers, because the systems built for space debris cannot see that far out. Radar excels in low Earth orbit — Gray notes it has tracked gloves and tool bags lost by astronauts — but at lunar distance the return signal is about 25.6 billion times fainter. Gray computes orbits for high-altitude objects specifically because they are the ones "the military doesn't track."
The surveys that did see it were not trying to. Their job is planetary defence, and as Gray puts it, "time spent observing junk is time not spent finding rocks."
Every engineering organisation has this arrangement and calls it something else. The deprecated service has no dashboard and no on-call rotation, and the only reason anyone knows it is degrading is that one engineer keeps noticing it in unrelated logs. That is not observability. That is a volunteer, and volunteers leave.
The recovery here is genuinely impressive, and it was made of data. By late February the archive held 1,053 observations, contributed by observatories in Mississippi, Utah, Beijing and England. Gray's software flagged the lunar impact in September 2025 — eleven months ahead.
The forecast then tightened as observations accumulated. A June revision moved the predicted impact nine minutes earlier and 170 km to the northwest. A July revision moved it by a couple of seconds and about ten kilometres. A final update landed on August 1.
That precision had a payoff: Danuri began imaging roughly thirty minutes before the collision and made eight passes over the site, capturing genuine before-and-after coverage. You only point an instrument at the right patch of ground at the right minute if someone gave you a forecast. And you only get a forecast for a system somebody is still measuring.
Worth noting, in the spirit of not overclaiming: as of this writing there is no published measurement of the new crater. Gray predicted a single crater around 17 m across. Much of the news coverage cites up to 27 m. Both are pre-impact calculations, and they do not agree with each other.
Whether it is one crater or two is also unsettled. The 2022 Chang'e 5-T1 upper stage surprised everyone by making a double crater. Gray expects a single one here because this stage carried no comparable mass at the top — a reasonable inference, not a result.
Numbers in circulation are not measurements. That distinction is the whole job in model validation, and it does not stop being true because a number is repeated confidently by everyone.
Translate the failure modes and they get uncomfortably familiar.
| What happened to 2025-010D |
The equivalent in a lending stack |
| Disposal fuel spent on the primary mission |
Decommissioning descoped to protect a launch date |
| Compliant maneuver, then no owner |
Control signed off at go-live, never re-verified |
| Solar radiation pressure over 18 months |
Population drift between validation cycles |
| Invisible to radar at lunar distance |
No telemetry on the deprecated service |
| Tracked by volunteers with other jobs |
One engineer who happens to know it exists |
Concretely, the things worth auditing this quarter:
- Retired scorecards that still score. A superseded model version left wired into a fallback path, quietly deciding a slice of applications.
- Orphaned document stores. Buckets of customer bank statements from a pilot that ended. Nobody owns them, retention policy never applied, and they are pure liability.
- Extraction pipelines with one remaining caller. The old parser kept alive for a single lender integration, unmaintained and unmonitored.
- Credentials outliving their purpose. API keys for a data provider you stopped paying, still valid, still in a config somewhere.
- Connectors for formats that changed. A bank redesigned its statement layout; the connector now fails silently on a small percentage and nobody set an alarm.
None of these will page you. That is precisely the problem: they fail in ways that produce no signal until the failure is someone else's finding.
The fix, in orbital mechanics, is cheap — if you decide early. Put the upper stage on a trajectory that leaves the Earth-Moon system entirely and ends up orbiting the Sun. The Falcon 9 stage from the EscaPADE launch in November 2025 was disposed of this way. Chinese lunar launches have moved in the same direction.
But it only works as a design-time decision. Once the propellant is gone, no option exists at any price. Gray is also honest about the limit: the trick "only works for a small fraction of junk," namely missions already headed beyond Earth orbit. Some classes of system genuinely cannot be cleanly retired, and the time to find out which class you are in is before you build.
The question Scheiman raised is the right one to steal: "What is the best future disposal path for high-energy missions that are in the sun-Earth-moon system?" The software version is smaller and duller and just as unanswered. What is the retirement path for this model, this pipeline, this data store? Who owns it after the launch party? What signal tells us it is still running, and who receives that signal?
Systems you cannot see are systems you cannot retire. Levmetrics' Lev Document Intelligence is built so that every extraction, score and decision leaves an auditable trace — including the ones you meant to turn off. Try it at app.levmetrics.com.
A four-tonne rocket stage is a useful teacher because it is impossible to lose track of politely. It kept an orbit, it kept a catalogue number, and eighteen months after everyone declared the mission a success it arrived somewhere nobody intended at 2.43 kilometres per second. Most abandoned systems are less dramatic and more expensive. Deployment is a milestone in a system's life. It is nowhere near the end of it.
- Mike Wall, "A SpaceX rocket will crash into the moon this week. Why is it on a collision course?", Space.com, August 3, 2026.
- Chelsea Gohd, "A SpaceX rocket just crashed into the moon. Now, the race is on to get a look at the impact site", Space.com, August 5, 2026.
- Mike Wall, "SpaceX rocket's impact crater on the moon spied by Korean lunar orbiter (photos)", Space.com, August 6, 2026.
- Bill Gray, "Upper stage impacting the moon on 2026 August 5", Project Pluto.
- Adam Kovac, "A SpaceX rocket booster is on track to hit the moon at several times the speed of sound", Scientific American, May 1, 2026.
- Leonard David, "Moon Impact of Rocket Stage Observed", Inside Outer Space, August 5, 2026.
- Benjamin Fernando et al., "preprint on the predicted lunar impact of the 2025-010D upper stage", arXiv.
A four-tonne piece of a SpaceX rocket spent eighteen months on a collision course with the Moon, starting at roughly the moment everyone stopped thinking about it. On August 5 it arrived.
The interesting part is not the collision. It is that the collision was scheduled eighteen months in advance by a decision nobody remembers making.
Yes. A spent SpaceX Falcon 9 upper stage crashed into the Moon on August 5, 2026, at about 06:34 UTC — 2:34 a.m. EDT — near the crater Einstein on the western limb. South Korea's Danuri orbiter confirmed the lunar impact, imaging the site across eight passes that began roughly thirty minutes before contact. The Korea Aerospace Research Institute reported that "changes in the terrain around the impact site and traces of ejecta spread were confirmed." At Lowell Observatory in Arizona, Carl Schmidt's team measured a sodium and lithium gas plume tens of kilometres across that lasted five to ten minutes.
What follows is not a space story. It is what the crash says about every system that outlives the attention paid to it.
The object has no name, only a catalogue designation: 2025-010D. It was the fourth tracked piece of hardware from the tenth orbital launch of 2025, on January 15, which carried Firefly Aerospace's Blue Ghost lander and ispace's Resilience toward the Moon. Blue Ghost landed successfully that March. Resilience lost contact about ninety seconds before touchdown and crashed in June.
Falcon 9 upper stages are normally deorbited shortly after launch and burned up in the atmosphere. That did not happen here. Pushing two landers to their deployment point consumed nearly all of the stage's propellant, so the disposal burn was no longer available. In the words of a study led by Benjamin Fernando of Los Alamos National Laboratory, the stage "was abandoned in a moon-crossing high-Earth orbit."
It then spent more than eighteen months in a lopsided 26-day orbit, swinging between roughly 220,000 and 510,000 km from Earth — repeatedly crossing the Moon's path, missing it every time, until it didn't. Astronomer Bill Gray of Project Pluto put the impact speed at 2.43 km/s and the released energy at about 14.5 billion joules, roughly three tons of TNT.
Nobody chose to crash a rocket into the Moon. The choice that mattered was made much earlier, and it looked like a fuel budget.
Decommissioning was the line item that got consumed to make the primary mission work. There was no meeting about it, no risk register entry, no dissenting engineer. The mission was harder than the margin allowed, the margin came out of end-of-life, and the end-of-life plan quietly ceased to exist.
This is the most common way software systems become permanent. The migration runs long, so the cutover eats the quarter, so the decommissioning of the old system slips to "next sprint." The old system keeps serving a trickle of traffic. Two years later it is load-bearing and nobody on the team has read its code.
The failure mode is not negligence. It is that decommissioning is the only part of a project whose deadline nothing enforces.
SpaceX did not simply walk away. Julianna Scheiman, the company's director of NASA science and Dragon programs, said a different maneuver was performed "to ensure that the second stage itself is safe per the appropriate rules and regulations."
That maneuver satisfied the rules. It did not put the object anywhere durable. As Scheiman put it, "a mixture of solar activity and gravity forces have put it on a path towards the moon."
A control that passes review at go-live is a statement about go-live. It is not a statement about the system's state eighteen months later, and treating the two as equivalent is how model risk accumulates. Sign-off is a snapshot; systems are a trajectory.
The force that moved it was sunlight. Solar radiation pressure — photons striking the tumbling stage — is, in Gray's description, "an extremely gentle force, but over months, it can really build up." It is also not fully predictable, because a tumbling object catches and reflects light differently as it turns.
No single push mattered. Eighteen months of them decided where four-and-a-bit tonnes of metal ended up.
That is drift, exactly as it appears in a scoring pipeline: no individual week's population shift is worth an alert, and the cumulative eighteen-month displacement is worth a regulatory finding. If your validation cadence is slower than your drift accumulation, you are not monitoring the model. You are just visiting it occasionally.
Here is the detail we found most uncomfortable. This object was tracked mainly by asteroid surveys and amateur astronomers, because the systems built for space debris cannot see that far out. Radar excels in low Earth orbit — Gray notes it has tracked gloves and tool bags lost by astronauts — but at lunar distance the return signal is about 25.6 billion times fainter. Gray computes orbits for high-altitude objects specifically because they are the ones "the military doesn't track."
The surveys that did see it were not trying to. Their job is planetary defence, and as Gray puts it, "time spent observing junk is time not spent finding rocks."
Every engineering organisation has this arrangement and calls it something else. The deprecated service has no dashboard and no on-call rotation, and the only reason anyone knows it is degrading is that one engineer keeps noticing it in unrelated logs. That is not observability. That is a volunteer, and volunteers leave.
The recovery here is genuinely impressive, and it was made of data. By late February the archive held 1,053 observations, contributed by observatories in Mississippi, Utah, Beijing and England. Gray's software flagged the lunar impact in September 2025 — eleven months ahead.
The forecast then tightened as observations accumulated. A June revision moved the predicted impact nine minutes earlier and 170 km to the northwest. A July revision moved it by a couple of seconds and about ten kilometres. A final update landed on August 1.
That precision had a payoff: Danuri began imaging roughly thirty minutes before the collision and made eight passes over the site, capturing genuine before-and-after coverage. You only point an instrument at the right patch of ground at the right minute if someone gave you a forecast. And you only get a forecast for a system somebody is still measuring.
Worth noting, in the spirit of not overclaiming: as of this writing there is no published measurement of the new crater. Gray predicted a single crater around 17 m across. Much of the news coverage cites up to 27 m. Both are pre-impact calculations, and they do not agree with each other.
Whether it is one crater or two is also unsettled. The 2022 Chang'e 5-T1 upper stage surprised everyone by making a double crater. Gray expects a single one here because this stage carried no comparable mass at the top — a reasonable inference, not a result.
Numbers in circulation are not measurements. That distinction is the whole job in model validation, and it does not stop being true because a number is repeated confidently by everyone.
Translate the failure modes and they get uncomfortably familiar.
| What happened to 2025-010D |
The equivalent in a lending stack |
| Disposal fuel spent on the primary mission |
Decommissioning descoped to protect a launch date |
| Compliant maneuver, then no owner |
Control signed off at go-live, never re-verified |
| Solar radiation pressure over 18 months |
Population drift between validation cycles |
| Invisible to radar at lunar distance |
No telemetry on the deprecated service |
| Tracked by volunteers with other jobs |
One engineer who happens to know it exists |
Concretely, the things worth auditing this quarter:
- Retired scorecards that still score. A superseded model version left wired into a fallback path, quietly deciding a slice of applications.
- Orphaned document stores. Buckets of customer bank statements from a pilot that ended. Nobody owns them, retention policy never applied, and they are pure liability.
- Extraction pipelines with one remaining caller. The old parser kept alive for a single lender integration, unmaintained and unmonitored.
- Credentials outliving their purpose. API keys for a data provider you stopped paying, still valid, still in a config somewhere.
- Connectors for formats that changed. A bank redesigned its statement layout; the connector now fails silently on a small percentage and nobody set an alarm.
None of these will page you. That is precisely the problem: they fail in ways that produce no signal until the failure is someone else's finding.
The fix, in orbital mechanics, is cheap — if you decide early. Put the upper stage on a trajectory that leaves the Earth-Moon system entirely and ends up orbiting the Sun. The Falcon 9 stage from the EscaPADE launch in November 2025 was disposed of this way. Chinese lunar launches have moved in the same direction.
But it only works as a design-time decision. Once the propellant is gone, no option exists at any price. Gray is also honest about the limit: the trick "only works for a small fraction of junk," namely missions already headed beyond Earth orbit. Some classes of system genuinely cannot be cleanly retired, and the time to find out which class you are in is before you build.
The question Scheiman raised is the right one to steal: "What is the best future disposal path for high-energy missions that are in the sun-Earth-moon system?" The software version is smaller and duller and just as unanswered. What is the retirement path for this model, this pipeline, this data store? Who owns it after the launch party? What signal tells us it is still running, and who receives that signal?
Systems you cannot see are systems you cannot retire. Levmetrics' Lev Document Intelligence is built so that every extraction, score and decision leaves an auditable trace — including the ones you meant to turn off. Try it at app.levmetrics.com.
A four-tonne rocket stage is a useful teacher because it is impossible to lose track of politely. It kept an orbit, it kept a catalogue number, and eighteen months after everyone declared the mission a success it arrived somewhere nobody intended at 2.43 kilometres per second. Most abandoned systems are less dramatic and more expensive. Deployment is a milestone in a system's life. It is nowhere near the end of it.
- Mike Wall, "A SpaceX rocket will crash into the moon this week. Why is it on a collision course?", Space.com, August 3, 2026.
- Chelsea Gohd, "A SpaceX rocket just crashed into the moon. Now, the race is on to get a look at the impact site", Space.com, August 5, 2026.
- Mike Wall, "SpaceX rocket's impact crater on the moon spied by Korean lunar orbiter (photos)", Space.com, August 6, 2026.
- Bill Gray, "Upper stage impacting the moon on 2026 August 5", Project Pluto.
- Adam Kovac, "A SpaceX rocket booster is on track to hit the moon at several times the speed of sound", Scientific American, May 1, 2026.
- Leonard David, "Moon Impact of Rocket Stage Observed", Inside Outer Space, August 5, 2026.
- Benjamin Fernando et al., "preprint on the predicted lunar impact of the 2025-010D upper stage", arXiv.
Lifecycle
Credit scoring
Fraud