Independent Recovery of Vanishing Sources
Between 1949 and 1958, the Palomar Observatory photographed the entire northern sky onto glass plates. Decades later, researchers started noticing something odd: some of those stars aren’t there anymore. Not dimmed, not moved. Gone. The VASCO project has cataloged thousands of these candidates by comparing old plates against modern digital surveys. Nine sources vanished simultaneously on a single plate from April 12, 1950. A bright triple transient appeared and disappeared on a July 1952 exposure. The published catalog contains 5,399 of these vanishing-source candidates.
I wanted to see how many I could independently recover using a completely different pipeline, with no reference to existing catalogs. Just raw plate data and an automated detection chain.
The pipeline
The system downloads digitized plate cutouts from three epochs (POSS-I Red from the 1950s, POSS-I Blue from the same date, and POSS-II Red from the 1990s), then detects sources via sigma-clipped thresholding, filters by PSF morphology (FWHM range and ellipticity), computes a local affine registration correction between epochs, and flags anything present on the Red plate with no counterpart on either reference epoch. A final cross-match against Pan-STARRS DR1 rejects sources with an obvious modern counterpart.
The full-sky sweep covered 122,991 grid positions across 932 plates on four rented A100 nodes.
Benchmark validation
On a 20-case benchmark harness, the pipeline recovers 8 of 9 known transients in the April 1950 field and all 3 transients in the July 1952 field. Blue-plate controls return zero candidates. Random non-crowded control fields produce ~0.2 false positives per 10 arcmin field.
The catalog match
Cross-matching the pipeline’s 2,852,431 candidate vanished sources against the 5,399-source Solano et al. (2022) catalog, I independently recovered 3,450 sources (63.9%) with a median angular separation of 0.94 arcsec. 98.7% of matches fall within 3 arcsec. Positional accuracy is not the limiting factor.
Of the ~1,950 unrecovered Solano sources within my footprint, about half are real detections in the raw catalog but get rejected by PSF filtering (they don’t look stellar enough). The other half show no detection above the 5σ threshold at all. None of the unrecovered sources have a Pan-STARRS counterpart within 3 arcsec, which means whatever they are, they’re not explained by an obvious modern match.
Nuclear tests and vanishing stars
This is the part people find most interesting. Bruehl & Villarroel (2025) reported a statistical association between the temporal distribution of POSS-I transients and above-ground nuclear weapons tests during the 1949-1957 survey period. Their finding: an elevated transient rate within ±1 day of test dates.
My independently constructed catalog shows a qualitatively similar pattern. The [+1 day] window has a 17.9% rate versus a 13.3% baseline, a relative risk of 1.35. Pre-test windows are flat. The directional asymmetry is there.
However, in my dataset, the effect doesn’t reach statistical significance (95% CI [0.91, 2.00], Fisher p = 0.17). One limitation is that every single one of the 368 observation nights in my study window contains at least one candidate vanished source, which means the binary calendar-day indicator is really just tracking whether POSS-I happened to observe on that date. A negative binomial model using nightly candidate counts with sky coverage as an exposure term gives IRR = 1.03 (p = 0.71).
The main issue is that my full-sky sweep used preliminary detection parameters (5σ, broader PSF acceptance) rather than the calibrated 8σ settings that perform much better on benchmarks. That means the catalog is noisier than it needs to be, which could easily wash out a real signal. A cleaner reprocessing with the tighter settings, plus explicit control for the observation schedule and plate-level clustering, would be needed to properly test this.
The main robust result of this study is therefore the independent catalog-level replication; the nuclear-test question remains open and will require a cleaner full-sky reprocessing plus explicit control of observational confounds.
What’s next
The 63.9% recovery rate and 0.94 arcsec positional agreement are the core result. Most of these vanishing sources aren’t catalog artifacts or astrometric failures. They’re on the plates, they pass morphological filtering, and they don’t have modern counterparts. What they are is still an open question.
The temporal question deserves a proper retest with a cleaner catalog. The calibrated 8σ settings materially improve benchmark performance and should produce a substantially smaller, less noisy dataset — the right starting point for a definitive look at the nuclear-test association.