The logit is a log-odds — a raw pre-softmax score. The logit lens reads a model's own forming answer at every layer by pushing each layer's residual through the final unembedding. This is not a toy: the real STITCH char-GPT (466,176 trained weights) runs its forward pass live in your browser, and the lens is read at each of its 4 layers. The instrument opens by proving the JS forward reproduces the reference PyTorch logits — if it can't, it fails loud and nothing else is trusted. Type below; watch the guess crystallize with depth.
LIT the forward pass is the REAL trained STITCH model recomputed live in JS (matmuls, GELU in its exact erf form — erf via an Abramowitz–Stegun ~1e-7 approximation — LayerNorm ε=1e-5, per-head causal attention, scale 1/√16); the self-check throws unless the JS final logits reproduce the reference PyTorch logits (max|Δ| < 2e-3) AND the per-layer lens tops match — so the forward pass on this page IS the reference model’s forward, proven, not asserted. (That these weights are the Shakespeare-trained STITCH is corroborated by the lens walk, not proven by this check.) The logit lens is exactly head(LayerNorm_f(residual)) at each layer.
AMBER the lineage dates + the expansion cards are cited history, not re-derived here; STITCH is a tiny 4-layer CHAR model on Shakespeare, so its "answers" are next-characters and its 4-layer crystallization is coarse — the phenomenon is real, the scale is small. "the model's guess / it believes" is shorthand for a probability the projection assigns, never a belief held by anyone.
WALL the logit lens is a biased readout (the tuned lens exists precisely to fix it — see [[the-tuning-glass]]); decodability of a token from a residual is NOT the model's causal computation; a real production model has hundreds of layers and word tokens where the effect is far sharper. No one is inside the glass reading along.