Every decision comes with the error rate it carries, checked against what actually happened.
CelestCore measures how wrong the predictions really are, and uses that to wait safely instead of burning early.
Predictions scored against the operator's own later orbit fits, millions of times. Each decision states its rate and the evidence behind it.
Wait until the last moment a burn still clears every plausible position. If a fresher prediction clears the encounter, the burn is never flown.
Fuel, service hours and money, including the orbit a retrograde burn takes out of service. You keep your own alarms and your own threshold.
We write the rules down before the test data exists, and publish every result as it came out.
3,957 test encounters built on real Starlink prediction errors, with the true paths known. Each bar is the cost as a share of the fixed rule that protects all 19.
| Retrograde burns cost service | Every burn costs service | |
|---|---|---|
| Fixed 1 in 10,000 (protects 17 of 19) | 53% | 21% |
| Fixed 1 in 100,000 (all 19) | 100% | 100% |
| CelestCore (all 19) | 37% | 21% |
Checked every day on predictions the error law had not seen, storms included.
| Day | Measured |
|---|---|
| Oct 5 | 0.50% |
| Oct 6 | 0.44% |
| Oct 7 | 0.28% |
| Oct 8 | 0.33% |
| Oct 9 | 0.38% |
| Oct 10 (storm) | 0.99% |
| Oct 11 | 0.97% |
| Oct 12 | 0.82% |
| Oct 13 | 0.59% |
| Oct 14 | 0.48% |
| Oct 15 | 0.53% |
| Oct 16 | 0.58% |
| Oct 17 | 0.51% |
| Oct 18 | 0.59% |
| Oct 19 | 0.58% |
| Oct 20 | 0.64% |
| Oct 21 | 0.64% |
Built on real Starlink prediction errors from four operator cycles, with test objects whose true paths we know. The usual rule burns at 1 in 10,000. To protect as well as CelestCore, a fixed rule has to burn at 1 in 100,000.
On these encounters the measured error rate was 0.3%, against the 1% we state. Every result is published as it came out.
Each one is scored after the encounter, in a chained record that can't be changed afterward.
Since October 18 we have recorded 11,057 decisions across 27 operator cycles, each before its encounter, and scored 3,574 encounters after they happened. We re-ran 4,618 of those decisions from scratch and every one came out the same.
Each decision names the files it read, the track record behind its error rate and the code it ran on. Run it again and you get the same decision, bit for bit.
Rules are frozen before the data that tests them exists. Every result is reported, including the ones that disappointed.
Operator predictions are replaced every cycle. A record kept since October 3, 2026 can't be rebuilt later by anyone.
From public catalog data alone, with alarms whose false-alarm rate is measured on objects that cannot maneuver.
At a 0.5% setting, measured on 859 debris objects and rocket bodies that cannot maneuver, held out from calibration.
Scored against 13,285 burns confirmed by an operator's own published trajectories, not against a model.
Runs on the public catalog alone, across a whole fleet, every day. Every alarm is timestamped and chained, just like the decisions.
Every reference conjunction within 10−12 (absolute) of CARA's TCA-adjusted 2D collision probability; 32 within 10−15.
CelestCore withheld reassurance. NASA's higher-fidelity reference later put the risk above its level for operational attention.
162,634 ESA data messages processed, covariance rebuilt to floating-point accuracy.
Design partners bring their predictions and outcomes. CelestCore brings audited decisions, in their units.