Each labelled cube is a model family. Ribbon colour shows which model wrote the forward pass, and the cone shows the direction of travel. Pick a pairing and press Play to send a pulse along it. qwen out and gemma back is the thick one at 0.743. Both pairings that come back through phi3:mini are thin, at roughly a third. Which model sits on the return side matters as much as whether the two models differ, so polyglottism turns out to be more than one knob.
This is a pilot: four dyads, the same thirteen stretches as MA12, one run on 2 October 2026. Read the ordering more than the decimals.
Theory tie-in: journal §14 uses these four numbers to reopen the question of where semiotic distance lives. Families differ, but the best cross-family pairing still sits well above the register and tier conditions in MA12. Open thread T31 waits on the full matrix.
Each panel opens one step on the way from the report file to the picture, with the equation it uses.
The same recovery measure as MA12, the Jaccard index of word-token sets, averaged over the same thirteen stretches. The difference is that two different model families do the two halves.
| Symbol | Stands for |
|---|---|
| f | The model that writes the forward pass. |
| g | The model that brings the text back, without seeing the original. |
| xi | Corpus stretch i. |
| T(·) | The set of word tokens in a text. |
Each ribbon is a smooth curve from the forward model's corner Pf to the return model's corner Pg, pulled toward a control point M. M is the midpoint pushed 0.7 units sideways, and up or down a little. The two ribbons between qwen and gemma (one each way) bow to opposite sides so they don't overlap.
Thickness grows in a straight line with recovery but doesn't start at zero, so even the thinnest ribbon stays visible. The 0.743 ribbon has radius 0.127 and the 0.31 ribbon 0.070. A ribbon twice as thick does not mean twice the recovery, so read the printed number.
Colour is the forward model (qwen blue, gemma green). The cone sits 86% of the way along the curve and points along it, B′(0.86), showing the direction of travel. The corners are alternate corners of a cube, chosen only to spread the three families apart, so the distance between two corners means nothing.
Play sends a white pulse along the selected pairing. Its position t advances by 0.35 per second and wraps around, \1 Turn the model spins the cube slowly so the three corners separate on screen. Nothing moves until you press a button, including when your system is set to reduce motion. The scene draws nothing while it is still.
The four scores are typed into the page from ma15_polyglottism_report.json (pilot run, 2026-10-02). The 0.743 for qwen out and gemma back, and the range of 0.31 to 0.37 for returns through phi3:mini, are also stated in journal §14. The 0.52 for gemma out and qwen back appears only in this figure's data and is still being checked against the report. The drawing code is the script at the bottom of this page, using assets/fig3d.js and Three.js r128.
qwen out and gemma back returns the most (0.743). The same two models the other way round return less (0.52). Anything that comes back through phi3:mini returns about a third. Direction matters: the same pair of models scores differently depending on which one goes first.
Lotman held that a thinking culture needs at least two generators built differently, which the site calls polyglottism. Many AI pipelines run two copies, or two sizes, of one model and call the result a conversation. In the archived MA12 run, two sizes of one model family barely lost anything on the way back, which is what Lotman would predict inside a single language.
The dissertation's archive has exactly that two-voice structure: secret memos and public releases written for different readers by different parts of the same state. This figure is the first test of whether AI pipelines can have it too. Different families do behave differently, so the answer is not a flat no. It is still a pilot with four pairings, and the full matrix has yet to be run.
Promising, and plainly underpowered. Four pairings over thirteen stretches can show that model families differ, but not by how much or why. That's a data problem, not a design problem, and it's cheap to fix: run the full matrix with every model on both the forward and the return side, add stretches, and report the spread in each cell alongside the mean. Because recovery can reward echoing, count echoed returns separately (T30).
If differently built models keep giving different returns once the matrix is full, the multi-agent work has its first real candidate for Lotman's two-generator minimum (threads T5 and T31). If they don't, that's worth knowing too: it would mean model swaps, like personas, stay inside one language.
| Pairing (forward, then reverse) | Mean recovery as drawn | Source |
|---|---|---|
| qwen2.5:3b, then gemma2:2b | 0.743 | ma15_polyglottism_report.json; journal §14 |
| gemma2:2b, then qwen2.5:3b | 0.52 | ma15_polyglottism_report.json |
| qwen2.5:3b, then phi3:mini | 0.34 | same; journal §14 gives 0.31 to 0.37 for phi3 returns |
| gemma2:2b, then phi3:mini | 0.31 | same |
Limits. Four dyads out of the possible matrix, thirteen stretches, temperature 0.0, open local models. Token Jaccard cannot tell a faithful return from a refusal to transform: in MA12, nine of thirteen cross-model reverse passes simply echoed the forward text.