Jump: How to read them Figures that move Figures to turn around Which angles are worth pursuing Where to start
Drag to turn the scene and scroll to zoom. Every figure has a Turn the model button under it that spins the scene slowly. Figures that show a process also have a Play button (Run the tests on the rooms, Play relays on L1, Play before and after on L3), where the motion stands for something: an item moving through a pipeline, or a passage going out and coming back. Nothing starts on its own if your system is set to reduce motion, but the buttons still work.
Every card ends with a line saying which parts are measured and which are drawn for the eye. Words with a dotted underline open a short definition. Click the word a second time to jump to its full entry in the journal glossary. Phrases with a gold dashed underline open the source behind the claim. Names of journal sections, figures, experiments and open threads have a thin solid underline and open a short plain-language note on what that part of the project is.
Each figure page ends with two extra sections. "How this figure is made" has fold-out panels with the equations and constants used to draw it. "In plain words" explains everything in the picture, why it matters, and how much potential the angle has.
A passage is sent out and brought back under five conditions. The nearer a loop gets to the gold anchor, the more of the original survived the round trip. The identity run nearly closes at 0.877. A change of register leaves a wide gap at 0.274.
Measured: the five recovery scores (token Jaccard, n = 13 stretches each). Drawn: the loops and shells.
Eighteen test items flow from retriever to generator. With no gate, the generator runs twice per item, 36 calls in all. Put a rule gate in front of it and 11 items stop there with the standard halt string, so the generator only runs 14 times.
Measured: call counts and gate closures, on flat retrieval only. Drawn: particle paths.
Each corner is a model family. A ribbon runs from the model that wrote the forward pass to the one that brought it back, and its thickness is how much of the passage came home. qwen out and gemma back holds at 0.743. Anything returning through phi3:mini falls to about a third. This is the site's first test of Lotman's call for generators built differently, which the site calls polyglottism.
Measured: mean recovery for four dyads over thirteen stretches. Pilot only; the full matrix is not run yet.
One column per real delegation from a coding agent to a subagent. Height is the compression ratio on a log scale. Every column clears 1, so no subagent handed back less than it was given, and the median report ran 4.38 times the length of its instruction.
Measured: all 26 ratios, read from the same array as the 2D beeswarm. Drawn: the play sequence.
Left dome: the parent answering alone, with the lowest unsupported-claim rate in the run (0.2167). Right dome: a subagent given no sources at all. It added more new vocabulary than any other arm, and 14 of its 17 answers used a key term found nowhere in the corpus. The purple slab is the typed tag that let some of that invention into the parent's answer under a real archive id, which is what this site calls laundering.
Measured: the rates and their ordering (17 items). Drawn: every particle.
Lotman sorts exchanges by two questions: does the message change size, and does the receiver add anything of its own? His named cases, I-s/he, autocommunication and presemiotic binding, are the wire rooms. The solid room is the corner he never named. The blue ball is the median of our 26 delegations, which lands in that corner.
Measured: the blue ball (volume 4.4 times, content gain 0.84, n = 26). Drawn: Lotman's rooms, which carry no counts.
This is the site's spatial reading of Lotman's semiosphere. Thirteen surfaces from the pilot corpus sit on nested shells according to who they were written for, from the secret memo out to diaspora press. Sky-blue dots travel along the thematic relays we can prove. Pravda gets no arc at all, and that silence is one of the findings.
Measured: every arc is a stored pair-read with a proof tier. Drawn: the shells and the pulse.
Every document as a point, placed by its sentence embedding and flattened onto three axes with PCA. Distance from the centre is the register shell and direction is topic. Green arcs mark the Russian to English relays the archive licenses.
Measured: direction on each shell, from the frozen encoder. Set by hand: the shell radii, which follow register.
A 3D bar chart of wording habits by document. The scare-quote towers on the internal memos (48 hits across three memos) drop to zero on the embassy release in the same row. That empty cell is the counted version of what Lotman's boundary refuses to pass, which the site calls untranslatability.
Measured: every bar is a raw operator count.
Two phrase networks stacked in time: the two memos from before Montreal on the lower plane, the fourteen after it on top. Red lines join wording habits that only start turning up together after 1983. This is the site's test for Lotman's explosion. The glossary gives the working definition.
Measured: every line is a count of shared memos, listed in a table under the figure.
What L1 would look like with every shell and relay type filled in. The extra documents and arcs are fabricated so a reader can see the target shape before the archive work supplies it.
Nothing on this page is a count.
Worth building on now. The Montreal semiosphere is the core of the dissertation, because every arc rests on two documents read side by side. The operator counts behind it are large and one-sided. The halt result is the site's clearest practical lesson for anyone building with AI. The round trips show cleanly that changing the audience moves text much further than changing the model.
Promising, but short of data, so be patient with them. The explosion rests on two early memos and needs more from before 1983. The model pairings are four cells of a matrix that is cheap to finish. The rooms and the delegation stack rest on one agent's logs and need other kinds of exchange placed beside them. Each page ends with specific suggestions.
Dead ends, at least as evidence. Embedding geometry as proof of the boundary has failed twice and was withdrawn. A round-trip score on its own can't tell creation from damage. Content gain can't show a contribution when the sources are hidden. Isolating a subagent produced invention, not a second voice. A refusal sentence in a small model's prompt made it refuse everything. These are worth reporting as negative results, not repeating.
Open the delegation stack and the transmission room together. The first shows that no hand-off got shorter. The second shows where that leaves the parent and subagent in Lotman's scheme, which is in a room he never drew.
Then open the halt flow. It is the one with a practical lesson: put the refusal rule upstream of the model, because a sentence in the prompt asking the model to refuse got treated as a ban on everything.
The Lotman L1 to L3 models are the archive side of the argument. They read best in the gallery, where all four sit on one page.
Committee translation: the 2D charts are the evidence and these are the slides you show so people believe you read the 2D charts.
Elsewhere: all figures · journal §13 · §14 · §15 · reading list