Munim Technologies

Research/Space

Gaia's own quality flags cannot pick out its real brown-dwarf orbits

Gaia's recommended quality cut keeps only 9 of 2,981 close low-mass spectroscopic orbits. Higher-significance orbits transit about six times as often as the rest, but even the best tier is mostly not right.

Nov 20, 2026 · 6 min read

Last week we showed that most of Gaia DR3's short-period spectroscopic orbits with brown-dwarf-like minimum masses cannot be real: of 2,180 such orbits with sensitive TESS light curves, about 408 should have shown transits and only 3 did (write-up). The obvious next question is whether Gaia's own quality columns could have warned a user. We wrote down the test before fetching a single quality value, then ran it. The answer is mostly no. Gaia's recommended "gold" cut keeps only 9 of 2,981 close low-mass orbits. Orbits with high significance do transit about six times as often relative to expectation, but even in that tier most orbits are not right.

9 of 2,981

close low-mass orbits that pass Gaia's recommended gold cut

5.8x

transit rate of significance-20+ orbits relative to significance below 10

p = 0.016

pre-registered one-sided test of that rise

12%

share of top-tier orbits that can be right (central model; 95% upper 44%)

The question

Gaia DR3 published single-lined spectroscopic (SB1) orbits with columns meant to say how much to trust it: significance (amplitude over its error), goodness of fit, efficiency, the number of good velocity epochs, and a period confidence. Gosset et al. 2024 (arXiv:2410.14372, Appendix J) recommend a gold sample: efficiency at least 0.2, at least 20 good epochs, significance at least 40, and period between 2 and 800 days. Our question: within the class where we know most orbits fail, the 2,981 brown-dwarf- and very-low-mass-class orbits at 1-13 days, do these columns separate the real orbits from the rest?

What we did

A real orbit seen close to edge-on must transit. For every orbit we computed how many transits TESS should have shown if the orbit were right with its minimum mass (E_iso), and counted the robust transit detections at Gaia's period. Their ratio, after subtracting the false alarms measured on control folds, is ρ_all: 1 means the orbits behave like real ones, 0 means none do. A second number, φ, is the share of orbits that can be right once heavier companions seen nearly face-on are allowed, under a central and a most-generous (floor) population model. The 2,981 orbits carry 531.7 expected transits; 25 were seen. We then split the sample by the gold cut (Q1), by significance tier (Q2), and by each other quality column at its median (Q3, exploratory). The rules, thresholds and predictions were committed before the quality columns were downloaded.

What we found

Q1, the gold cut: UNDERPOWERED (we predicted this at 70%). Only 9 orbits pass, and they would show 1.2 transits if all were right; they show 1. That is far too few to test anything. The cut removes 99.7% of this class, so it cannot rescue it. Its significance threshold of 40 shuts out almost every orbit with a small velocity amplitude; the four known transiting brown dwarfs and planets in our sample have significance 6.7-13.6, so the gold cut would have thrown away all four.

Transits seen per transit expected (ρ_all), by Gaia significance tier

100% would mean every orbit behaves like a real one with its minimum mass. 95% intervals: T1 1.0-5.6%, T2 0.0-6.3%, T3 4.2-41.6%. False alarms from control folds are subtracted. Source: data/qcut-v1/qcut_summary.json

Q2, significance: RISES_PRIMARY_ONLY. The primary pre-registered test, top tier against bottom tier on all hits, passed (p = 0.016, 18 hits). The robustness test on substellar-size hits alone did not (p = 0.052, 6 hits). We had given this outcome only 15%; our favourite, a rise in both tests, was wrong, as was the no-rise outcome. The rise is real enough to matter, but small: once face-on companions are allowed, about 12% of the top-tier orbits can be right under the central model, and the 95% upper bound is 44%. In the two lower tiers, where 94% of these orbits sit, the share is 1.5-4.4%.

Share of orbits that can be right (φ), by tier

tierorbitshitsφ central (95%)φ floor (95%)
T1 (< 10)1,834130.015 (0.000-0.047)0.038 (0.001-0.117)
T2 (10-20)97470.024 (0.001-0.080)0.044 (0.002-0.146)
T3 (20 or more)17350.121 (0.015-0.437)0.163 (0.020-0.590)
Central and floor are the pre-registered population models for companions seen nearly face-on; floor is the most generous. Source: data/qcut-v1/qcut_summary.json.

Q3, the other columns (exploratory): none passes the corrected threshold of α = 0.01. Goodness of fit came closest (p = 0.0104) in the expected direction: the better-fitting half transits more often. Two ingredients of the gold cut, efficiency and the number of good velocity epochs, pointed the wrong way. The half Gaia would call better transits less often. That direction was not a pre-registered test, so we report it only as a description.

Each quality column split at its median (ρ_all)

columnbetter half (fixed in advance)better halfworse halfone-sided p
goodness_of_fitlow0.0560.0090.0104
rv_amplitude_robust / 2Klow0.0470.0190.10
conf_spectro_periodhigh0.0330.0320.56
efficiencyhigh0.0230.0440.90
rv_n_good_obs_primaryhigh0.0200.0460.94
A p near 1 means the better half did worse. Source: data/qcut-v1/qcut_summary.json and findings/2026-11-20-quality-cut-results.md.

What it means

For anyone using Gaia DR3 to find brown-dwarf companions in short orbits, the catalogue's own columns are not enough. The recommended cut empties the class, and significance improves the odds without making them good. These are population statistics; no single orbit is called wrong here.

Gaia DR4, due on 2 December 2026, gives a second check: whether each DR3 orbit is found again with about twice the data. That grade (Q4) is already registered with committed code. On 20 November we added two more tests to it, fixed before DR4 exists. Q4e asks whether goodness of fit, the near-miss above, predicts which orbits DR4 confirms. Q4f asks whether DR4 survival and the transit test agree about efficiency and epoch count. We predict they will disagree (60%), because more epochs make an orbit easier to repeat whether or not it is real. If they do, DR4 survival alone cannot certify these orbits, and any per-orbit reality score must be trained on transits.

More research