Context. The CNEOS database offers near-global coverage of fireball events based on U.S. Government sensor detections. The database contributes to investigations of meteoroid impact fluxes. However, the accuracy of these data is not reported. Aims. We aim to evaluate the reliability of CNEOS-derived ephemerides of fireball events given the absence of the underlying data. In particular, we identify conditions leading to larger or smaller orbital uncertainties and analyze the self-consistency of velocity vectors as a proxy for errors. Methods. We analyzed 18 events that have both (i) sufficient satellite information to derive orbits and (ii) ground-based observational counterparts. Specifically, we quantified the uncertainties on these "calibrated events" using the orbital similarity metric, DD. We also examined the geocentric velocity components imbalance at the population level and identified discriminants that can indicate the accuracy of a given event. Results. We identified two distinct groups in the CNEOS database. CNEOS data produces ephemeris determinations with DD < 0.1 for fireballs reported either (i) after late 2017 or (ii) with impact energies above 0.45 kt, with 74–78% of events having DD=0.03 ± 0.02 and ∼11% showing DD < 0.008. Our statistical test confirms these two parameters as the only reliable discriminants that, when combined, explain the two accuracy groups. Daylight, the z-velocity component, low altitude, duration, and latitude might also indicate errors, although the limited dataset may obscure correlations. No clear discriminants are identified for more restrictive DD cutoffs. We provide estimates of orbital uncertainties for calibrated events. The hyperbolic fireball subset in the CNEOS database appears as an outlier in the velocity imbalance test. Conclusions. Our results confirm that the fidelity of CNEOS fireball data improved significantly from 2018, likely due to the deployment of next-generation space sensors, and show a growing number of high-velocity events. Hyperbolic candidates should be interpreted with caution, as their velocities and inclinations likely reflect measurement errors. Accuracy constraints remain limited by the dataset size, as is evidenced by the lack of statistically significant dependence on duration, preventing strong conclusions from being drawn.