Full-wave EM solvers (FEM, FDTD, MoM) are powerful, but most wrong results come from setup choices rather than solver bugs. My checklist before trusting any result:

Geometry
- Are materials correct, including frequency-dependent dielectric models? Laminate datasheet Dk values are often measured at one frequency and with one method; design values can differ.
- Is copper thickness and surface roughness modelled? Roughness significantly increases conductor loss at higher frequencies.
- Did you simplify away something important (solder mask, plating, via barrels, connector launches)?

Ports
- Are port sizes appropriate? A wave port that is too small cuts off the field; too large and it may support higher-order modes or include nearby structures.
- Is the port impedance normalization correct? Check the port's computed characteristic impedance against your expectation.
- De-embed port extensions to the actual reference plane.

Boundaries
- Radiation boundaries/PMLs should be far enough from radiating structures (a common guideline is about a quarter wavelength or more at the lowest frequency, but check your solver's documentation).
- For closed structures, make sure the enclosure does not create cavity resonances that are not present in reality, or deliberately include them if they are.

Mesh and convergence
- Run a convergence study. In adaptive FEM, check that delta-S has converged and that results do not change when you tighten criteria.
- Make sure the adaptive mesh frequency is appropriate; for broadband sweeps adapt at the highest frequency or use multiple frequencies.

Physics sanity checks
- Passivity: |S21|^2 + |S11|^2 must not exceed 1 for a lossless two-port; check for non-passive data before exporting to a circuit simulator.
- Reciprocity: S21 should equal S12 for passive reciprocal structures.
- Compare a simple sub-structure (a straight line) against a closed-form calculator.

Finally, validate one build against measurement. After that, you know how far to trust the model for the rest of the design.