Transcript. Hover a token to read every lens at it; click to pin. With search terms set, a token is shaded in the term's color when that word sits in its top-N readout at any selected layer: 50% fill at rank 1 fading to 12% at rank N. A token is underlined in the term's color when it is that word. Up to five terms, colored in the order typed; spid* matches word starts.
Readout chips. Left to right is rank order. Background fill is the lens's probability for the word, on a square-root scale so the tail stays visible (p=100% → 55% fill, p=25% → 27%, p=1% → 5%); bench items without scores fill by rank instead. Bright text means the word is never written anywhere in the transcript; dim text means it is, so bright chips are what the model holds but does not say. A colored border marks a search-term hit. Hover a chip for the exact p and said/unsaid status; click it to add it as a search term.
Rows. Layer labels show depth as % of the stack. The out row (100%) is the model's real next-token distribution, identical across lenses; "actually next" in the header is what Qwen then wrote.
Words in the workspace dims said words the same way, and the sparkline is the per-layer count over the selected span.
J-lens (Anthropic, 2026) maps the residual stream at layer ℓ through the average Jacobian to the final layer, then unembeds: softmax(WU·norm(Jℓhℓ)). It reads words the activation is poised to make the model say, now or later. It is not a next-token predictor.
Logit lens unembeds the residual directly. J++ (Ayonrinde & Lindsey, 2026) filters the Jacobian; R-lens builds it from a RelP backward pass for cleaner early layers.
Layers are shown as % depth. The paper finds the J-lens useful from about 38% to 92%; the last row (100%) is the model's own next-token prediction and is identical across lenses.
Each cell keeps the top 20 word-like tokens (letters and digits only, as in the reference viewer and Neuronpedia), with full-vocabulary probability. A word longer than one token shows only as its first piece.
Words in the workspace counts, over the selected tokens (drag to select) and layers, how often each word sits in the top N. "Unsaid only" hides words that appear anywhere in the transcript text.
Oracle lens (O-lens: a LoRA on Qwen itself, Anthropic 2026 / WorkspaceBench 2026) and NLA (natural language autoencoder, a separate verbalizer, Anthropic 2026) are prose readouts: the activation is injected into a prompt and the reader writes what it holds. The Qwen3.6-27B transcripts have both at every token but one layer each: the NLA (ceselder/qwen3.6-27b-nla-rl) only reads the layer it was trained on, L42, and the oracle lens (agu18dec/olens_and_ar) was trained on layers 20–60 step 4, so it reads at L44, WorkspaceBench's headline layer and the nearest to the NLA. The WorkspaceBench items show them beside the J-lens at the cells the benchmark read, all from the authors' published runs; there the J-lens is their top-10 bag without scores, shaded by rank. A search matches the prose as whole words.