The floor has two axes, labelled in the scene: whether the volume changes, and how much the receiver adds. The four rooms are the four corners of that floor, the same quadrants as the 2D chart. Position inside a room is not scaled, and the rooms' heights mean nothing.
Ordinary sending, I-s/he, keeps the message the same size. Autocommunication lets it grow, but only through rhythm and repetition, with no new facts. Presemiotic binding sits beneath both: parts of one organism, where the receiver is prized for adding nothing. A subagent's report grows and fills up with new words, so it fits none of the three, and the plain transmission picture that Lotman criticises rules out new messages altogether.
Press Run the tests to watch the measured hand-offs try each of Lotman's named rooms in turn.
Theory tie-in: this is the 3D version of journal §13's correction. The claim is narrow: Lotman's two axes let you see when a pipeline does not fit his named boxes. Nobody is claiming he would have called a coding agent a semiotic person. Section 13 argues the dissertation should admit the room exists before it names it.
Each panel opens one step on the way from the report file to the picture, with the test it uses.
Lotman describes his cells in words, not numbers. These conditions are the site's reading of them, set out in journal §13 and panel D3 of the delegation page. Ordinary sending keeps the volume of information the same. Autocommunication can grow it, but only through a structural second code such as rhythm or repetition, so no new content. In presemiotic binding the recipient is valued for adding nothing.
The measured medians for MA13. ρ is report characters over instruction characters, and g is the share of the report's vocabulary absent from the instruction. ρ̃ is far above 1, so the I-s/he test fails. g̃ is far above 0, so the I-I content test fails. The receiver plainly adds a great deal, so the presemiotic case fails too. That leaves the fourth corner.
| Symbol | Stands for |
|---|---|
| ρ̃ | Median compression ratio over the 26 delegations. |
| g̃ | Median content gain ratio over the same 26. |
The floor is the 2D chart's two-by-two grid laid flat. Left to right is whether the volume of information changes; back to front is how much the receiver adds. Each room is the same box, 1.5 × 0.9 × 1.5 units, centred on its quadrant at (±1.8, ±1.8), and it follows the 2D canvas's placement. The blue ball sits inside the unnamed room by category. Its position inside the room is not scaled by ρ or g, and the rooms' heights mean nothing.
The 2D page's guide text calls I-s/he the "receiver adds little" corner, while its canvas places it in the top row. This figure follows the canvas until that is settled.
Run the tests moves the blue ball into each of Lotman's three named rooms in turn. Each move takes 1.3 seconds along a small arc, the room lights up red while the caption says which test fails and why, and after 2.4 seconds the ball moves on. It ends in the unnamed room. The stops are categories, not measured positions; the argument they act out is panel D3 on the delegation page. Turn the model spins the floor slowly. Nothing moves until you press a button. The two medians and n come from ma13_real_subagent_report.json (aggregate.compression_ratio.median, aggregate.content_gain_ratio.median, corpus.n_delegations). The drawing code is the script at the bottom of this page, using assets/fig3d.js and Three.js r128.
An AI agent handing work to a subagent fits none of Lotman's named kinds. The report comes back about four times longer, and most of its words were not in the request.
The dissertation uses Lotman's types to read Soviet institutions. Secret KGB memos are partly the institution talking to itself; embassy releases are ordinary sending aimed outward. To use the same tools on AI pipelines, you first need to know which room an AI hand-off belongs in, and it belongs in none of them. Either the typology needs a new room, or AI delegation is a new kind of exchange. The journal's position is to admit the room exists before naming it. That keeps the dissertation from forcing machines into categories built for people and institutions.
Useful as a frame, and honest about its limits. Placing one median in a quadrant is a categorical argument, not a measured distance, and it rests on the site's own reading of Lotman's cells. It gets much stronger once other kinds of exchange are placed on the same floor and scored the same way: human-to-model sessions (MA7), sibling subagents talking through a parent (MA11 and MA16), and the archive's own memo-to-release pairs. Then the empty room is compared with something other than itself.
It would also help to measure the presemiotic and I-I corners directly rather than by description. If several different exchanges land in the unnamed room, the dissertation has a case for naming it. If only AI delegation does, the finding is narrower but still useful.
| Value as drawn | Report key | Source file |
|---|---|---|
| volume about 4.4 times at the median | aggregate.compression_ratio.median = 4.3843 | ma13_real_subagent_report.json |
| content gain about 0.84 at the median | aggregate.content_gain_ratio.median = 0.8387 | same |
| n = 26 real delegations | corpus.n_delegations = 26 | same |
Typology cells from Lotman, Universe of the Mind: A Semiotic Theory of Culture, trans. Ann Shukman (Bloomington: Indiana UP, 1990), autocommunication chapter. MA13 is retrospective token arithmetic over one agent's corpus with the child's intermediate prose redacted. Its content gain is a vocabulary measure and says nothing about quality.