nx_attribution.nx

nx_attribution.nx -- per-output bottom-up contribution tracking. User correction 2026-05-16: "your feedback is top down not bottom up as well i dont know wtf the bits are doing and no system exposes that thats the point of all the nishi lang no more black boxes or secrets". THIS primitive is the bits-up answer. For linear/matmul-class operations, the substrate EMITS what every contribution was -- signed, ranked, named-input -- so callers can ASK "which input contributed most to this output?" and get an exact answer at full bit precision. No black boxes. No hidden activations. No post-hoc inference from outputs. The substrate exposes the bits. ===== Why this matters =========================================== Every existing ML stack stores tensors as opaque blocks. When output[i] = 0.7, no one knows which input element was responsible. Mechanistic-interpretability research (Olah, Conmy, Nanda 2022+; Anthropic's circuits) reverse-engineers this MANUALLY for tiny models. Production stacks ship without it. NishiLang substrate makes attribution a first-class primitive: any matmul / linear / attention op can be called with attribution recording on, and the substrate emits the (input_idx, contribution) pairs that produced each output. ===== Math ======================================================= For y = W @ x where y is [M], x is [N], W is [M, N]: y[i] = sum_{j=0..N} W[i, j] * x[j] Each (j, W[i, j] * x[j]) pair is one contribution. We store the top-K largest-by-absolute-magnitude per output element. Signed: contribution[i, k] = W[i, j*] * x[j*] where j* is the k-th rank. For attention output[i, d] = sum_j attn_weights[i, j] * V[j, d]: