Wiringprint-

WiringPrint studies model identity at the level of computation rather than conversational persona. The central question is whether changing the coordinates/names of internal components changes the model, or merely changes its parameterization. We first establish the core result on pretrained GPT-2. A compensated permutation of the 3,072-neuron MLP hidden axis preserves GPT-2 behavior: baseline and aligned accuracy are both 0.3155, perplexity is 56.629 in both cases, ECE is 0.1050, and the maximum observed logit difference is only 9.16×10⁻⁵. The corresponding uncompensated permutation collapses accuracy to 0.0409 and raises perplexity to 3892.811, with maximum logit difference 53.760. We then test whether this identity principle generalizes beyond GPT-2 across real pretrained Qwen2.5-1.5B, Gemma-2-2b, and Llama-3.2-1B models. Compensated MLP permutations preserve next-token behavior within numerical precision across all three families, while uncompensated permutations and gate-only permutations degrade behavior. We additionally examine grouped-query attention, representation geometry, causal tracing, multi-seed tests, long-context behavior, continuous perturbations, and cross-model CKA.

The key implication for digital-minds research is deliberately narrower than a claim about consciousness: raw parameter coordinates are insufficient as a complete criterion for individuating a neural model.. At least some large changes in weight space correspond to the same input-output computation. Conversely, breaking the wiring relation between components can change behavior. This gives a concrete empirical distinction between parameter identity, functional identity, and representational identity, and provides a framework for asking whether “the model,” “the instance,” or “the persona” is the relevant unit of concern. This puts us under Track 5 a mechanistic way to investigate model vs. instance vs. persona identity.