[ok maybe we can do the idea. right now we want the code to make symbols instead of words. thinking on process of recursive confidence] [1 to start you'll need a way to discern some basic improvement, such as something being more right if it has a given change, or something being wrong in a certain way 2a later i suppose we'll need a way of providing for changes to resonate off that information: for example transcribing it as an automated metric. 2b the system needs some way of discerning the attribute. whether it is relevent, or what it might be, or testing if it is there. some way to engage it repeatedly.] [so how does the system tell anything about symbols? is there any way to know the symbols are there?] [there is almost certainly a logits pattern that indicates the symbols. they're near different other words. right in the output.] [we could look in model internals too, but it's expected to be about the same since the input data is expected to be so consistent.] [we could check the input data too.] [so we could train the model on that.] [theoretically but training takes data. we've been conditioned to think about training, trying to connect with mainstream work. instead, we want an algorithmic pattern that helps, or just a little data.] [okay um - the letters probably have patterns that indicate them - we could resonate those patterns by identifying some data is this reasonable] [or is the resonance too vague?] [it could have gone other ways, it's a hybrid] [ok ummm it sounds like it would be quite helpful for it to identify the onset. this is resonance, but for it to be _confidence_ resonance we'd need to be more precise around how it would work. what the confidence would be. and we'd want multiple channels.] [ok so that actually looks like a good approach because autocorrelation and similar things are very general and can apply to a lot of data. we could try to find patterns that let it learn to predict which areas are code and which are text. the patterns would be found from feature-resonance between areas. but we want to formalise out the idea of feature resonance, to produce a confidence metric.] [ok so we would give it some kind of heuristic axiom around patterns in general, that would have some hardcoded confidence.] [now we have hard or inhibited part around the patterns inside [looks inhibited]] [let's do one part at a time. any pseudo-draft of 'feature resonance'?] [for example, you could train a model to simply predict what data comes next in the stream. [oops two branches here]] [we were going to consider feature resonance but then realised that such a model would already be trained to recognise the change, because it would have to learn it to do the prediction appropriately. there are different token frequencies in the two sets] [ok such a model would indeed produce confidence. what was the earlier idea though?]