There isn't a single "correct" t_depth cutoff -- it depends on your assay (deep targeted panel vs WES vs WGS), tumor purity, and somatic vs germline -- and, more importantly, t_depth on its own is the wrong field to threshold. t_depth is just total coverage at the site; a position can have t_depth = 500 and still be a bad call if only 2 reads support the ALT allele.

Filter on the combination instead:

• t_alt_count (ALT-supporting reads): require e.g. >= 3-5. This is what actually guards against low-support calls.

• VAF = t_alt_count / t_depth: for somatic tumor calls a floor around 0.05 (5%) is common; raise it if purity is high or you want high-confidence only.

• t_depth: a floor mainly guarantees the VAF is estimable. For WES somatic, >= 10-20 is typical; for a deep panel set it much higher since everything is deep anyway.

Two notes given your range (10 to ~15000):

• The very high end (thousands) is usually a deep panel or hotspot pileup, but ultra-high depth can also flag repetitive/misaligned regions -- a strand-bias or mapping-quality filter (or an upper cap) catches those.

• If this MAF has a matched normal, also use n_depth and the normal VAF to drop germline leakage, not just tumor depth.

Practically: plot VAF vs t_depth (coloured by t_alt_count) for your callset and set cutoffs where the low-support junk separates out, rather than picking one t_depth number blindly.