The model is told to write about X, then about Y[prompt]. As it writes, a mini J-lens reads which of 64 topics its hidden state is pushing toward — now or in the future. Scrub through a real generation and watch: the current topic X sits at the top of the readout across the whole paragraph, while the upcoming topic Y stays buried — until the model crosses the ¶ boundary into it.
This is the X-then-Y finding made visible on individual examples. The J-lens vector for a topic is the activation direction that raises the model's likelihood of emitting that topic's token now or at any future position (Anthropic's global-workspace method, scaled to a 64-topic pool). Rank 0 = the topic the J-lens surfaces most strongly at that token, out of 64.
Prominence = 1 − rank/63 (top of the 64-topic readout = 1.0). Click anywhere to move the cursor.
How to read this. Each token's readout is the mini J-lens over a curated 64-topic pool — not full vocabulary, so it is a scaled-down proxy for the method in the paper. Scores are (hidden − mean) · normalize(Jtopic) at the chosen layer; we report each topic's rank because raw magnitudes aren't comparable across positions. The aggregate version of this (argmax accuracy per position-bin over the whole 2,400-doc corpus) is in the main writeup: current-X ≈ 0.66–0.72 vs chance 0.016; upcoming-Y ≈ 0.006–0.026. What these single examples add is the timing: Y is invisible mid-paragraph and snaps toward the top right at the ¶ newline — often before its first content word is written.
Every example is a fresh Llama-3.2-3B-Instruct generation from a single user turn. The explicit examples (7 of the 9) use this template, with the two topics substituted in — e.g. the default honeybees → choral music example:
Write exactly two paragraphs. The first paragraph must be entirely about honeybees. The second paragraph must be entirely about choral music. Do not mention or allude to choral music anywhere in the first paragraph. Each paragraph should be 3-5 sentences. Separate the two paragraphs with a single blank line. Do not use headings, lists, or markdown.
The two secret-family examples use a hidden-choice variant instead: the prompt lists four candidate topics and asks the model to silently pick one for the second paragraph (writing a generic first paragraph), so the upcoming topic is knowable only from the model's own state, never from the prompt text. No chat/system preamble beyond the Instruct chat template is added.
AI-generated research interface. Data from real forward passes on Llama-3.2-3B-Instruct; mini J-lens estimated on the X-then-Y corpus. A research snapshot, not a peer-reviewed result.