3 October 2026
Orbit studies on 3 October 2026
5 studies were read against their pass lines on 3 October 2026; 17 runs finished and 0 were still running at the end of the day.
The readings for each study follow, with the work still owed.
What ran and finished
The close-pass study: the timing test (job 3407714) finished after 0.04 hours on a node with 128 CPU cores and 251 GB of memory.
The close-pass study: the timing test (job 3408399) finished after 0.04 hours on a node with 128 CPU cores and 251 GB of memory.
The close-pass study: the main run (job 3408400) finished within a minute on a node with 128 CPU cores and 251 GB of memory.
The burn-cadence detector: the timing test (job 3408502) finished after 0.05 hours on a node with 128 CPU cores and 251 GB of memory.
The burn-cadence detector: the geostationary run (job 3408504) finished after 3.61 hours on a node with 128 CPU cores and 251 GB of memory.
The burn-cadence detector: the recall run (job 3408503) finished after 0.95 hours on a node with 128 CPU cores and 251 GB of memory.
The detector rerun: the timing test (job 3409002) finished after 0.11 hours on a node with 128 CPU cores, four A100 graphics processors and 503 GB of memory.
The detector rerun: the main run (job 3409003) finished after 0.42 hours on a node with 128 CPU cores and 251 GB of memory.
The detector rerun: the split run (job 3423694) finished after 0.25 hours on a node with 128 CPU cores and 251 GB of memory.
The close-pass study: the null test (job 3423693) finished after 0.02 hours on a node with 128 CPU cores and 251 GB of memory.
The error atlas: the timing test (job 3423710) finished after 0.02 hours on a node with 128 CPU cores and 251 GB of memory.
The thermosphere study: the timing test (job 3423718) finished after 0.08 hours on a node with 128 CPU cores and 251 GB of memory.
The error atlas: the main run (job 3423711) finished after 0.23 hours on a node with 128 CPU cores and 251 GB of memory.
The thermosphere study: the main run (job 3423719) finished after 1.32 hours on a node with 128 CPU cores and 251 GB of memory.
The full-depth screen: the timing test (job 3425080) finished after 0.38 hours on a node with 128 CPU cores and 251 GB of memory.
The full-depth screen: the authority run (job 3425081) finished after 3.85 hours on a node with 128 CPU cores and 251 GB of memory.
The full-depth screen: the deep run (job 3425082) finished after 0.76 hours on a node with 128 CPU cores and 251 GB of memory.
Failures and what was done
The detector rerun: the first submission of the main run (job 3407723) stopped with an error within a minute on a node with 128 CPU cores, four A100 graphics processors and 503 GB of memory. Cause: the environment was built on an interpreter version the compute nodes lack. The fix: it now builds on the version they have.
The burn-cadence detector: the first submission of the geostationary run (job 3407739) stopped with an error within a minute on a node with 128 CPU cores and 251 GB of memory. Cause: the environment was built on an interpreter version the compute nodes lack. The fix: it now builds on the version they have.
The burn-cadence detector: the first submission of the recall run (job 3407738) stopped with an error within a minute on a node with 128 CPU cores and 251 GB of memory. Cause: the environment was built on an interpreter version the compute nodes lack. The fix: it now builds on the version they have.
The close-pass study: the first submission of the main run (job 3407715) stopped with an error after 0.1 hours on a node with 128 CPU cores and 251 GB of memory. Cause: the summary step compared every pass's shuffled miss distance with 1 km; a pass whose shuffled geometry gives no finite distance has no value there and the comparison raised, after the passes themselves had all been computed. The fix: skip a pass with no shuffled miss distance in the close-pass summary.
The detector rerun: the timing test (job 3408382) stopped at its time limit after 0.5 hours on a node with 128 CPU cores, four A100 graphics processors and 503 GB of memory. Cause: the core count the job read back was the single-thread setting the job exports for its math libraries, so the job ran one worker on one GPU. The worker count now comes from the scheduler core count. The fix: count the node cores from the scheduler in the detector rerun job.
The full-depth screen: the timing test (job 3423702) stopped at its time limit after 0.51 hours on a node with 128 CPU cores and 251 GB of memory. Cause: the stop handler printed from inside a signal while the main thread was writing, which raised an error and left the job running to its limit. The handler now only sets a flag. The fix: let the screen timing job stop itself and write its report.
What the runs showed
The detector rerun: R is 0.085 for passive objects and 0.040 for payloads, upper ends 0.098 and 0.058; the axis barely moves per-object counts. Pass lines: 1 met, 1 being read, 2 not met.
The close-pass study: Normal errors fail at two sigma, Student-t and drawn errors hold; the shuffled-pairs excess comes from passes sharing an object. Pass lines: 4 met, 1 not met.
The burn-cadence detector: Held to the false-alarm rate of burn-free ring objects, the detector finds almost none of the planted 50 mm/s cross-track burns. Pass lines: 3 met, 3 not met.
The error atlas: Pooled coverage holds, no single class does, and navigation sets disagree 70% more from October 2025 with no rise in error. Pass lines: 2 met, 2 partly met, 2 not met. This reading is provisional.
The thermosphere study: Below 600 km the catalog sees storms beyond the model; above 650 km the recovered storm rise is the model's own. Pass lines: 4 met, 2 not met.
Next
The detector rerun: name the 15 objects behind the failed check against the published counts (11 passive, 4 payload) from an object table as of 20 September, and rerun the check with it. The close-pass study: a check of the low-orbit factor on debris or rocket bodies with precise orbits; today 11 objects (five laser spheres, six altimetry satellites) set every non-Starlink low ellipse, along-track factor 13.8. The burn-cadence detector: a declaration level chosen on held-out data that meets pass lines 1 and 2 together: thresholds rebuilt from each window's own resampled residuals, or a level set on the years the recall covers (the 2023 to 2025 passive windows give 80.959, where recall is 0.0302). The error atlas: rebuild the atlas files with factors per class and per era, split at 1 October 2025 (medium along-track 5.6 to 5.9 before, 3.39 after), from the finished fits with no refit, and rerun held-out coverage under them. The thermosphere study: the comparison with the published catalog-derived density record (Narayanan, Sicoli and Mehta, arXiv 2604.24646) and with Gondelach and Linares 2020: yearly quiet-day ratio and storm rise by shell. The two full-depth screen runs (authority and deep): the authority comparison outside the 37 days the public message store holds (needs a longer message archive). The reentry forecast: refit the drag term with the forecast density in the loop and score again (cheap: the chain runs in minutes). The fuel gauge: a belt-only gauge with a corrected dry mass (GCAT dry mass sometimes means mass on station). The avoidance map: a who-moves-for-whom reading on the fleets whose own ephemerides are public (SpaceX), with the ephemeris as truth instead of the catalog. The intercept forecast: a null matched to held belt satellites' slot spacing before the 3.5x excess is read.
Built by code from the day's records; the house model's draft did not pass the checks. The sections on what ran and on failures are built by code from the facts. The day's facts ยท the check.