10.14.3 Processing steps
The following four filters were applied to the set of sources described in the previous subsection:

1.
Only sources for which at least 40 $G$band FoV transit magnitudes were available (after removing outliers, cf. Section 10.2.3) were retained.

2.
Although the photometry calibration (cf. Chapter 5) is continuously being improved towards the final Gaia DR4 release, the preliminary calibration used for Gaia DR3 does not yet lead to light curves that are completely free of instrumental artefacts. In particular, for many stars the highest peak in the frequencygram of the $G$band time series is at a frequency that is near (but not necessarily exactly on) a (sub)harmonic of the spacecraft spin frequency, e.g., around 4, 8, 12, 16 d${}^{1}$, etc. In fact, we found that main frequencies around 4, 8, 12 d${}^{1}$, etc. are predominantly instrumental rather than physical. We therefore removed all sources from our sample that have a main frequency in one of the following intervals: $[0.9,1.1]$, $[3.9,4.1]$, $[7.9,8.1]$, $[11.8,12.2]$, $[15.9,16.1]$, $[19.9,20.1]$, and $[23.85,24.1]$ d${}^{1}$. In addition, we removed all sources with a main frequency smaller than 0.1 d${}^{1}$, as this is outside the expected range of pulsations frequencies.

3.
Only sources for which the highest peak in the frequencygram (computed in the Variability Characterization module of the pipeline, see Section 10.2.3) had a false alarm probability (FAP; Baluev 2008) lower than ${10}^{3}$ were kept. The fairly stringent FAP threshold also helps the cleaning of sources with a main frequency that is instrumental.

4.
Given a source with main frequency ${f}_{1}=1/{P}_{1}$, we computed the phase $\psi \equiv [({t}_{n}{t}_{0})mod{P}_{1}]/{P}_{1}$ of each point in the light curve, for a reference time ${t}_{0}$. If the corresponding phase diagram (phase vs.m̃agnitude) showed a gap in phase larger than 0.25, we removed the source from our list, as we often found that in such cases the phase curve was not convincing and the extracted frequency might be an artefact coming from the particular time sampling for that star.
After applying the filters mentioned above, we are left with 54 476 candidates.
Following the procedure outlined in Section 10.2.3, each light curve was fitted with the following model:
$${m}_{G}({t}_{n})=C+\sum _{k=1}^{K}{A}_{k}\mathrm{cos}(2\pi k{f}_{1}{t}_{n}+{\phi}_{k})$$  (10.22) 
where ${m}_{G}({t}_{n})$ is the observed $G$ magnitude at time ${t}_{n}$, $C$ is the fitted model constant term, ${f}_{1}$ is the fundamental frequency, and $K$ is the number of significant harmonics, with amplitudes ${A}_{k}$ and phases ${\phi}_{k}$. In Gaia DR3, $K=15$ but only information up to the first three harmonics (if significant) is published. The vari_ms_oscillator table contains for each source the main frequency frequency1, its FAP value fap_g_freq1, the number of significant harmonics num_harmonics, the amplitudes amplitude_g_freq1, amplitude_g_freq1_harm2, and amplitude_g_freq1_harm3 of the first three harmonics, and the corresponding phases phase_g_freq1, phase_g_freq1_harm2, and phase_g_freq1_harm3, respectively.