The faint-end radial-velocity tail of Gaia DR3: external verification of extreme values and a continuum of discrepancies in the 100 pc sample

Abstract

Abstract Close-encounter studies rely on Gaia DR3 radial velocities (RVs) to the faint limit of the RVS. Our earlier encounter atlas found that stars with G>13.5 and |RV|>150 km/s but disc-like tangential motion inflate the sub-parsec encounter rate sevenfold, and conjectured that their RVs are faint-end artefacts. We test this with a preregistered external comparison (hypotheses, decision rules and stopping conditions committed before any cross-match) of the 1,323 suspects and 27,538 non-suspect control stars from the same 100 pc six-dimensional sample against LAMOST, APOGEE, GALAH and RAVE. The control group produced the first result: the preregistered calibration gate failed because control stars disagree with external RVs at 3σ in 8–17 per cent of cases. The discrepant side is Gaia's: the external surveys agree with each other at the binary level while Gaia disagrees with each; the discrepancy fraction rises from 2 per cent at G=13.5–14 to 26–48 per cent at G=15.5–16, tracks the RVS signal-to-noise, and is uncorrelated with binarity indicators. Under a rule fixed before any suspect was judged, 119 of 144 suspects with an external RV (82.6 per cent; Wilson 95 per cent interval 75.6–88.0) are not reproduced, with a median |ΔRV| of 194 km/s; the other 25 are genuine high-velocity stars. The two nearest "encounters" are spectroscopically classified white dwarfs, outside the RVS pipeline's validated domain, revealing an unapplied quarantine in our atlas that we report as an erratum. We quantify the consequences for encounter rates and recommend a layered quarantine for DR3-based work. Bundle. This record carries the preprint (faintrv_en.pdf; submitted to MNRAS), the per-star cross-match results catalogue (crossmatch_results_v1.json/.csv), the control-group calibration table (calibration.json), the whitelist of 25 externally confirmed high-velocity stars (bit5_whitelist_v1.json), the rate-impact table (impact.json), the preregistration document with its dated amendment log (preregistration.md; preregistration commit d32e2671), the adjudication records (rulings #24–#27), the figures, and the complete repository archive wake-atlas-faintrv-v1.tar.gz (cross-match code in src/wake_vs/, quality gate scripts/gate_check_faintrv.py). Code files are additionally licensed under MIT (LICENSE in the archive); the record as a whole is CC BY 4.0. Relation. Supplement to the WAKE encounter atlas (doi:10.5281/zenodo.21966305), whose quarantine bit5 is the object of this test; the atlas erratum E-2 (unapplied white-dwarf quarantine bit0) is reported in Section 5 of the preprint. AI disclosure: the analysis pipeline, figures and drafts were produced with an AI coding assistant (Claude, model Claude Fable 5; Anthropic) under the direction and adjudication of the human author (preprint Appendix C); this internal verification is not a substitute for peer review.