How to explore: drag empty space to rotate · scroll to zoom · click a dot (short click, don’t drag) for detail in the panel - no hover picking; that keeps the camera stable. Green = licensed RU→EN; orange = contrast; purple = intersemiotic; violet dots = audio carrier.
Each panel opens one step on the way from a passage of text to a dot in space, with the equation it uses.
Each passage, Russian or English, goes through the frozen Sentence-Transformers model paraphrase-multilingual-MiniLM-L12-v2. Every word piece comes out as a vector hi, and the average of the n vectors is the passage's fingerprint e. The language-encoding appendix walks through this step by step.
The 384 numbers are reduced to three, u, with principal component analysis: subtract the average fingerprint ē and project onto the three directions W along which the passages differ most. The dot is then placed on its shell in that direction. Every dot sits at exactly its shell's radius: 1.35 for secret memos, 2.75 for Soviet outward texts and the silence records, and 3.55 for rival print.
| Symbol | Stands for |
|---|---|
| u / ‖u‖ | The direction only, which comes from the model's reading of the topic. |
| rs | The shell radius, set from the archive's register label and not from the model. |
The page and the site describe the reduction as PCA. The exact settings live in the thesis-lab builder and should be checked there before this panel is cited.
Each arc bends through a point M above the midpoint of its two dots, so it arcs over the sphere instead of cutting through it. The data holds 38 dots (18 Russian, 20 English, 3 of them audio) and 33 arcs: 13 from secret to rival texts, 9 from secret to Soviet outward texts, 7 from secret texts to silence records, and 4 between rival texts. Green arcs are licensed Russian to English relays, orange arcs are contrasts, and purple arcs are intersemiotic.
Nothing in the data moves. Turn the model spins the shells slowly; the direction of each dot comes from the model and its distance from the archive, so watching it turn shows which passages point the same way.
Positions are computed in the thesis lab and stored in the data object at the bottom of the page, which is copied here on sync. This section is kept in explain/fig-semiosphere-embedding-3d.html in the site repo and added back on every sync.
The model happily lines up a Russian secret memo and an English newspaper story when they share a topic, even though they were written for different readers and say opposite things about it. The shells only appear because the project imposed them from the archive.
The project's early headline number came from this kind of geometry and was later withdrawn (journal §1), because fingerprints track topic and miss register. Scare quotes, the switch from "symposium" to "gathering" and the shift to drought talk are what separate the memo from the embassy release, and those are what the model doesn't see. This figure shows why the dissertation reads pairs of documents side by side instead of trusting similarity scores.
It matters for AI systems too. Most retrieval tools rank by this kind of fingerprint, so without rules about audience they will hand back a secret memo when someone asks for the public statement. Lotman's boundary has to be added from outside the model.
Pessimism is earned here. Using this geometry as evidence about the boundary has already failed twice on the site. The headline alignment number was withdrawn because it did no better than its null, and the gap between shells measured by nearest neighbours reversed once topic was controlled for. With nine licensed pairs in 384 dimensions, nothing fitted to these points is likely to survive. Keep the figure as an illustration of what fingerprints see and miss, not as a measurement.
If the angle is revisited, it needs many more licensed pairs first (T11), a second frozen encoder to rule out the model as the cause (T10), and matched nulls built in from the start. A humbler use may pay off sooner: as the baseline retriever that the typed, rule-based approaches have to beat.