nx_gen_ditchain.nx

nx_gen_ditchain.nx -- chain N sovereign DiT blocks, weights read straight from the GGUF. This is the sovereign DiT forward stack: given the layer-0 input, the timestep embedding and the rope table, it runs N transformer blocks back to back with NO oracle tensor of any kind in the loop -- every weight comes from the model file, every adaLN vector is computed, every activation is the previous block's own output. Usage: nx_gen_ditchain <model_id> <gguf_path> <n_layers> [n_workers] [reference_name] The reference defaults to `layers.<n_layers>.blk_in`, i.e. the oracle's input to the block AFTER the ones we ran -- so a 2-layer run is graded against layers.2.blk_in. WHAT THIS ANSWERS THAT nx_gen_blockrun DOES NOT blockrun proves ONE assembled block is correct. Chaining asks whether the error stays bounded as blocks compose: 36 layers is 36x the opportunity for a small bias to accumulate into a visibly different image. Per-op fidelity says nothing about that, and neither does one block. ⚠The pass band defaults to 1e-1 because the reference is the ORACLE, which carries its own activation-quantization error at every matmul of every layer it ran. Grading a MORE accurate implementation tightly against it would fail correct work. To ask "did OUR chain drift", grade against an exact reference instead. license_tier: ORIGINAL