Maneuver Ledger · public orbit data · 3 October 2026 to 17 October 2026
How many of Starlink's collision-avoidance burns can public orbit data see?
SpaceX told the FCC its satellites made 207,152 collision-avoidance manoeuvres from 1 December 2025 to 31 May 2026: about 42 per satellite per year. We looked for them in 14.5 days of public element sets. In the cleanest test, at most 8% of them can be 16 cm/s or larger, the size public data reliably sees.
The one-minute check
42.1avoidance manoeuvres per satellite per year, filed by SpaceX (65,137 Gen 1 plus 142,015 Gen 2, over an average of 9,866 Starlink satellites in orbit)
2.8at most, per satellite per year: burns of 16 cm/s or more on satellites holding 530 to 600 km (conservative 95% bound from 6 steps of any kind in 4.8 satellite-years)
8%of even the lower filed rate (Gen 1, about 36 per satellite per year) could be burns that large
So, if the filing-period rate still held, at least 92% of Starlink avoidance manoeuvres change the orbit by less than 16 cm/s along track, or leave its size unchanged. Public element sets therefore cannot rebuild the 207,152. We did not reproduce it; this page shows why it cannot be reproduced this way.
Detection limit in that band: the 175 m threshold is about 9.5 cm/s at its median altitude of 570 km; a burn of 16 cm/s added to a real recorded history is found 98% of the time; 10 cm/s, 65%. False alarms on passive debris: 0.6 per object per year (95% upper 2.1).
How the bound is built: the upper end of the two-sided 95% Poisson interval for the step count (one-sided, that end has 97.5% coverage), divided by the lower end of the 95% interval for recall (also one-sided 97.5%). Two bounds combined this way are conservative: by Bonferroni, joint coverage is at least about 95%.
What the steps in public data are
Across 9,479 satellites holding altitude between 400 and 600 km (113 satellite-years observed), steps above 175 m of semi-major axis (about 9.7 cm/s at 470 km) happen 25.2 times per satellite per year as drag make-up boosts (a decay, then a boost back) and 8.7 times as other small single changes (95% interval 8.2 to 9.3). Both are far rarer in the highest band, where drag is weakest. Close approaches are not: DODGE's daily screen of the same satellites finds more passes under 1 km in the highest band than in the busiest one. So most visible steps follow drag, not conjunctions.
| Altitude band | Satellites | Drag make-up per sat-year | Other small steps per sat-year | DODGE passes under 1 km per sat-year |
|---|---|---|---|---|
| 400 to 460 km | 182 | 2.9 | 7.5 | 15.9 |
| 460 to 475 km | 5,717 | 31.4 | 9.3 | 16.4 |
| 475 to 490 km | 3,238 | 15.2 | 8.7 | 14.3 |
| 490 to 530 km | 44 | 9.1 | 3.6 | 30.4 |
| 530 to 600 km | 298 | 0.6 | 0.6 | 21.7 |
Steps above the calibrated threshold, from public GP sets. "Other small steps" is the avoidance-like class: an upper bound on avoidance burns above the limit, since phasing moves look the same. Passes are screening grade (DODGE), not collision probabilities.
Detection limit
Each point: 570 to 585 burns injected into randomly chosen real on-station histories, or synthetic histories with fit noise of 12.9 m and a bad-set share of 0.3%, both measured on the real data. Synthetic precision at 16 cm/s: 100.0%.
Rate against threshold
The threshold is the smallest step on the grid at which the 95% upper bound of false alarms on passive debris near Starlink altitudes is at most 5% of the filed rate. The rule is fixed in code; the threshold follows from the debris data alone. Zero false alarms are drawn on the bottom edge of the log scale.
Three public histories
STARLINK-35057: most drag make-up boosts
Mean semi-major axis from each public GP set, relative to the first; median altitude 462 km. Dashed lines: detected steps.
STARLINK-34435: biggest single bad set, correctly ignored
Mean semi-major axis from each public GP set, relative to the first; median altitude 463 km. Dashed lines: detected steps.
STARLINK-38435: orbit raising after its launch on 2026-10-13
Mean semi-major axis from each public GP set, relative to the first; median altitude 250 km. Dashed lines: detected steps.
Outside judges
| Check | Result |
|---|---|
| Synthetic orbits, burns of 16 cm/s, fit noise and bad sets as measured | precision 100.0%, recall 99.7%; 1 false alarms in 400 burn-free histories |
| Passive debris and rocket bodies, 400 to 600 km (they cannot manoeuvre) | 2 false steps in 1,264 object-days |
| Held out: satellites launched during the window (catalogue dates, never shown to the detector) | 27 labelled raising, 0 on-station |
| Held out: satellites launched over a year earlier | 37 raising, 7,295 on-station, 703 descending |
| Held out: satellites the FCC report lists as removed from operation, still in orbit | 189 descending, 62 holding altitude, of 251 |
| Public GP against SupGP (fits of SpaceX's own ephemerides) on the same steps; pre-registered line: 70% beyond chance | 29% beyond chance (48% agree, 27% by chance in time slides): not met |
| DODGE's first detector (one threshold on consecutive sets) | 50 jumps per satellite-year: it counts one bad set as two jumps and does not separate drag make-up, so it is replaced here |
Limits, stated first
- Different months: the filing covers 1 December 2025 to 31 May 2026; the data covers 3 October 2026 to 17 October 2026. Rates are compared, not counts. The 92% figure assumes satellites in the highest band dodge about as often as the Gen 1 fleet did then; DODGE's screen shows them with more close approaches, not fewer. If they dodged only a tenth as often, up to 78% of their burns could be that large.
- Only the size of the orbit is measured. Two burns that cancel inside one gap between element sets, or burns that change only the plane or shape, are invisible.
- The classes are heuristics. Drag make-up and avoidance cannot be told apart for a single step; the bound uses every step in the band, whatever its class.
- Public GP sets are derived from US Space Surveillance Network radar and optical observations (CelesTrak's description); whether they also use operator ephemerides is not known to us.
- The highest band holds 298 satellites. The bound tightens as DODGE keeps recording every 4 hours.
What an insider could check in a minute
- Is a typical Starlink avoidance burn smaller than 16 cm/s along track?
- Are the regular boosts every few days in the lower shells drag make-up, as classed here?
- Do SpaceX's published ephemerides include planned burns? Our GP and SupGP steps agree less often than we pre-registered.
Data and sources
- Filing: SpaceX Gen 1 and Gen 2 Semi-Annual Reports, cover letter 1 July 2026. Quote: "Using this very conservative threshold, along with additional miss-distance-based criteria, SpaceX Gen 1 satellites performed 65,137 propulsive maneuvers over the reporting period, averaging approximately 36 maneuvers per satellite, per year and SpaceX Gen 2 satellites performed 142,015 propulsive maneuvers over the reporting period, averaging approximately 46 maneuvers per satellite, per year."
- Orbits: public GP sets and the Starlink Supplemental GP set from CelesTrak, recorded by DODGE because CelesTrak keeps only the latest set.
- Fleet size: CelesTrak SATCAT, Starlink payloads in orbit averaged over the filing period. The filed rates imply 3,631 plus 6,196 satellites.
- Every detected step:
out/events.csv. Every number on this page:out/summary.json. Method:docs/METHODS.md.
About this page
Maneuver Ledger is a separate research project. It reads DODGE's sealed recordings of public orbit data and never records anything itself. Code, data, method and a byte-for-byte reproduction: github.com/GautamTalksDev/maneuver-ledger (method, questions). Code Apache-2.0, derived data CC BY 4.0.
This page is generated from the ledger's results page at commit 0b1ab8b (SHA-256 of that page d6fa8ad7704a9609). Paths such as out/summary.json refer to that repository.