10.7.4 Processing steps
The choice of the best period and geometrical model is performed according to the following procedure applied to the $G$band light curve after outlier removal (see Mowlavi et al. 2023):

•
We first retrieve the two most significant frequencies from each of the following three period search methods: Generalized Least Squares (Cumming et al. 1999; Zechmeister and Kürster 2009), String Length (Lafler and Kinman 1965; Burke et al. 1970), and Phase Dispersion Minimisation (Jurkevich 1971; Stellingwerf 1978; SchwarzenbergCzerny 1997).

•
We build a list of candidate frequencies by taking the uptosix most significant frequencies from the previous step and adding several of their multiple values, such as half or double.

•
For each of the candidate frequencies obtained in the previous step, we compute seven candidate models, one for each model type (see Section 10.7.3), to which a constant model is added.

•
We remove from the above list of candidate models the ones having obvious unrealistic components, such as overlapping narrow Gaussians. The Bayesian Information Criterion (BIC) is then computed for each of the remaining candidate models.

•
Finally, using the BIC value, we select the best frequency–model combination. If multiple frequency–model combinations exist within $$ from the best BIC value, the model type with highest predefined ranking is chosen (see Mowlavi et al. 2023, for details). At this stage, a source can be excluded from the catalogue of EBs, for example, if the constant model is selected, or if there are too few points in the eclipses. The final model type is provided in the field model_type.
The period and model parameters published in the Gaia DR3 catalogue of EBs are the ones from the best model selected with the above procedure. The great majority (94 %) of the candidates are modelled with two Gaussians, while only 4 % have only one Gaussian in their model, and the remaining 2 % is modelled with only one sine function. When applicable, additional attributes are derived and published from these parameters, such as an estimate of the durations and depths of the eclipses.
Uncertainties of the model parameters are computed based on a Jackknife procedure where a full remodelling of the $G$ light curve is performed by removing in an iterative way a single observation from the time series.