Data sources. Ephemerides, eclipse duration and depth for short-period systems from
Gaia DR3 (the
vari_eclipsing_binary table), which carries an explicit frequency error from which the uncertainty is computed. Detached systems with a sharp eclipse from
GCVS 5.1, whose periods rest on a long time baseline and are therefore more accurate than Gaia for narrow eclipses. Names and variable classifications from
AAVSO VSX.
Updated ephemerides from VarAstro. For stars marked
VarAstro in the uncertainty column, the period and the time of minimum were taken from an export of
VarAstro, maintained from a database of hundreds of thousands of observed minima. These are the freshest data available, and they include the actually measured phase of the secondary minimum instead of assuming it is 0.5. Each of these ephemerides is anchored to the latest instrumentally timed minimum in VarAstro's own O−C record, with the trend of the last eight years where there is one. Its uncertainty is measured, not assumed: predicting the latest minimum of 702 stars from the ones before it gave errors of about 4.4 minutes at best, growing by 2.55 minutes for every year since the last observed minimum. Beyond ten years, or with no instrumentally timed minimum, it is shown as unknown, with the year of the last recorded minimum, and the star is kept rather than hidden.
The timing uncertainty is also given as a number for stars whose ephemeris comes from Gaia, where the frequency error was measured and published - doubled, because checked against TESS eclipse times the published error proved about half the real one. It is recomputed for every date, as the period error times the number of cycles elapsed since the reference epoch, and so grows the further you get from it. It does not include real period changes — mass transfer, or a light-time effect from a third companion.
What is not included. Stars whose only ephemeris comes from GCVS were removed from the database entirely. No period error is published there, and calibrating against Gaia showed the real error to be 6.8 times the published precision at the median and 120 times at the 90th percentile — meaning any number shown for them would have been fiction. An example: FP Sge, whose epoch is from 1934; after 52,524 cycles the deviation reaches two hours. Every row in the table rests on an ephemeris whose error can be quantified.
To verify. Cross-reference against
VarAstro's minima predictions — the link on each star's name searches for it there directly. The tool does not account for cloud, a local horizon, or the Moon.
How the windows are chosen. Each window runs from the eclipse start minus the padding to the eclipse end plus the padding. The tool checks that the
whole window — not just the eclipse — falls inside the site's dark window and stays above the altitude threshold from start to finish. A window where only the eclipse itself qualifies is not included, because the scheduler would not be able to carry it out. The number of sites on each row is a measure of the chance the request gets scheduled: a window caught from three sites competes for three systems in parallel.