Validation and levels

Four checks, in order, and six level conventions. Together they are why a design you did not write yourself can be allowed near your speakers.

The four checks

CHECK	WHEN IT RUNS	WHAT IT LOOKS AT
Shape	When the file is read	Field names, types, ranges, what is required. An unknown field comes back with the candidates listed
Coefficients	On save, and on load	Designs the coefficients from your values and confirms them at both internal rates
Simulation	On load, only for a chain containing something nonlinear	Passes three stimuli through it and measures what comes out
Arming	Immediately before sound starts	That the simulation passed, and that the limiter and soft start are in place

Shape and coefficients must pass before a design enters the library at all, so the list never offers something that is certain to fail when picked.

A failure names the stage, the code and the numbers that were measured. It is one of the two expected answers, not an error.
Passing the shape check does not mean it will play. Shape is only shape; the verdict starts at the coefficients.
If anything fails, playback falls back to passing the signal through untouched.
The same verdict can be taken without making a sound, which is how an assistant can check its own work before you load it.

STIMULUS	WHAT IT REVEALS
Sine at the reference level, whole periods	Pass-through gain, the compensation, and the DC component
Logarithmic sweep across the audible band	How much harmonic content lands above the audible band, and the peak
Band-limited noise	Intermodulation, and any invalid numbers

None of the three contains ultrasonic content, so anything measured up there was made by your chain.

DC is judged from the sine alone. On the other two the estimation error is larger than what is being judged, so their numbers are reported but not used to refuse.
A chain containing a long convolution is warmed up before it is measured, so loading one takes longer.
Read the results in order. Invalid numbers upstream make everything downstream meaningless.

Ultrasonic content is refused because it cannot be heard.
A full-scale tone above the audible band reaches your amplifier and tweeter unchanged, and you cannot notice it the way you would notice an audible fault. What is refused is amplifying it, judged in absolute terms, so a properly windowed high-pass filter passes.
DC is refused because it pushes a driver and holds it there.
Asymmetric shaping and even-order harmonics create it. Put the DC blocker after the stage that creates it; in front of it nothing is removed.
A linear chain is not put through this.
Simulation, the limiter and the soft start apply only to chains containing a nonlinear node. Applying them to everything would mean the path closest to a straight wire is the one most interfered with.

Fixed by the specification version. They do not change under you.

#	CONVENTION
1	1.0 is 0 dBFS. There is no hidden gain anywhere
2	The analogue reference point is -20 dBFS RMS as 1.0 V RMS, at the input node of the circuit being modelled
3	Every module passes the reference level at unity, within 0.1 dB. The checks confirm it
4	Drive on a nonlinear node is compensated at its output: the tone changes, the loudness does not
5	Nothing clips inside, because the signal is floating point. The limiter exists only at the final stage
6	A/B switching matches loudness automatically, against the pass-through as 0 dB

Conventions 4 and 6 are why an A/B comparison does not become a test of which side is louder.