They've not said anything, and honestly, there's only a handful of people talking about it. There aren't enough units in hands yet really to get a proper lay of the land. Preferring the sound of the legacy Panama is not the same as saying that the Agoura Panama has accuracy problems.
So my reading in this area constitutes basically a few toilet sittings, a bunch of ChatGPT interrogation, and some ResearchGate browsing while plopping.
Now from what I know of amp modelling, you've got your basic bullshit stateless waveshapers. That's what everyone was using in the early days of 2004-2008, say. Not much more than cascaded tanh functions. The tanh is a seriously useful function, and roughly can sound quite 'tube like' - if you remember any early amp simulator, it'll be doing variations on stacked tanh functions.
Then you started to get amp sims built on Hammer-Weinstein principles, where engineers and DSP developers realised that distortion in an amp circuit probably isn't stateless; or memoryless. That's when you get frequency dependent distortion, with linear components feeding into non-linear components, and back into linear. That's when amp sims start to actually kind of respond properly to palm-muting versus open chords. This is where the famous EQ > Distortion > EQ concept comes from; in a general sense. You read it on forums all the time, or a variation of it by people who have half a clue. This would've been Amplitube 2 era.
And loads of stuff has been done on that basis, and complicated further by just chucking more shit into the signal path.
Then a clever cloggs comes along and goes, well no.. what if we actually model tube physics. Don't have the computation for it? Let's generalise it. That'll be your Axe FX II, your initial Softube plugins probably, possibly even Amplitube 3. Who knows. Hard to tell. But around that era there was a serious step up in quality for amp sims.
Then after that, circuit-level modelling came. Wave digital filters, modified nodal analysis, applying Newton–Raphson solvers to nonlinear equations to essentially recreate diodes and other components. From all of this, you can build complex triode or pentode dynamic models, build real-time dynamic power supply emulation, and proper feedback loops; which is essential for good NFB. That's your Axe FX III, in essence.
Then from about 2015 onwards, it's been machine learning approaches. Using WaveNet style neural networks to compute convolutional models that learn nonlinear systems. They really got popularized in the last decade or so.
Now it is difficult for me to state where Helix lies in all of this. But my guess would be somewhere between generalised modelling approaches, and full circuit level modelling. I'm sure some of OG Helix was properly circuit modelled. I reckon quite a lot of it was more generalised due to computing power and tech stack implications.
Sooooo.. my whole line of argument being... if you prefer OG Panama ... it is entirely possible that your ears PREFER the more inaccurate model. To you it might be better. But it doesn't really mean that the new Agoura models are less accurate. I'm willing to bet that they are in fact more accurate and that there isn't any problem whatsoever.
