Validation

RadarSim's test suite has 421 tests. They run on every push for Python 3.9 to 3.12 on Linux, and on Windows and macOS. Most of them compare a model with an independent result rather than with its own earlier output.

pip install -e ".[gui,dev]"
pytest                              # about 40 s

What is checked

Area	Reference used by the tests
Radar equation	Skolnik 2008, example 2.1: SNR 17.26 dB to within 0.01 dB
Detection range	free-space range grows as σ^¼ to within 2 %
Swerling Pd, cases 0–4	pulse-level Monte Carlo of a square-law detector, 200 000 looks, 1 and 8 pulses
Albersheim	inverse consistency with the exact Pd
CFAR multipliers (CA, GO, SO, OS)	measured false-alarm rate on exponential noise
CFAR loss	ordering SO > OS > GO > CA and known CA values
ITU-R P.676-13	ITU validation workbook, 12 significant digits
ITU-R P.838-3	published coefficients
Sea clutter	NRL model reference values
Doppler	\(f_d = -2v_r/\lambda\), sign and magnitude
Pulse-Doppler chain	matched-filter gain τB, peak in the right range and velocity bin, MTI null at DC
Antenna pattern	half power at half the 3 dB beamwidth
Kinematics	constant velocity, constant acceleration, exact coordinated turn on a circle
Terrain	mountain peaks block the ray, flat terrain does not
Tracking	straight-line convergence, NIS consistency, confirmation through a 4.7 g turn without a break
Jamming	J/S scaling with range and bandwidth, burn-through consistency
Receiver limiter	complex-Gaussian closed form at full scale
Determinism	same seed gives identical runs; reset reproduces a run exactly
Scenarios	every shipped scenario has no unknown keys, survives export and reload unchanged, and does what its description says (tests/test_scenarios.py)
Interface	every scenario loads with its radar values intact, every scope renders, settings reach the engine, editor round trip, recording and replay, a bad file leaves the run going

tests/test_known_defects.py keeps one test for each defect fixed during the 4.0 overhaul, so none of them can come back unnoticed.

Monte Carlo against theory

The coverage curve of the basic_tracking radar for a 1 m² target, with empirical detection rates from 400 independent looks per range and their 95 % Wilson intervals:

Signal-level output

A range–Doppler map from the coherent processor: three targets, LFM pulse compression, Hamming-windowed 64-point Doppler FFT. Closing targets appear at negative radial velocity.

Reproducing the figures

python tools/make_figures.py          # docs/assets/figures
python tools/capture_screenshots.py   # docs/assets/screens