One real block of Qwen3-8B, opened and re-checked in your browser
Every number on this page was recorded from the real model while it read “The researcher opened the notebook”. Your browser then redoes each step that does not need the 16 GB of weights, and tests which formula reproduces Qwen's numbers. Try the wrong ones: only the operations Qwen actually uses fit.
Following the last token, ␣notebook (position 5); the other 5 positions provide keys and values.
What this does and does not show
It shows that the recorded tensors are consistent with Qwen3's documented operations, step by step, to bfloat16 precision, and that plausible alternatives (LayerNorm, no √d scaling, a different rotation pairing, a different KV grouping, ReLU) fit worse.
It does not re-derive the matrix multiplications (q_proj, k_proj, v_proj; o_proj; gate_proj, up_proj; down_proj), which need the learned weights, and it covers one token in three blocks.
Captured with scripts/latentlens/capture_glassbox.py (Qwen/Qwen3-8B-Base @ 49e3418f, bfloat16, eager attention). The same checks were run in Python at capture time and agree.
Back to the journey: the same block with toy numbers →