How an LLM thinks, in semiosphere space

PART ONEWhat generation actually is

Before any comparison to human thought, it is worth being concrete about the mechanism. A language model does not retrieve a proposition and then phrase it. At each step it produces a probability distribution over the whole vocabulary, conditioned on a window of preceding tokens, samples one, appends it, and repeats. Nothing else persists between steps except the window itself.

Left: the next-token distribution at four successive steps. Each step is a fresh distribution; the model commits to a sample and cannot revisit it. Right: the context window sliding forward. Tokens that fall out of the window stop conditioning anything at all - they leave no trace, no summary, no accumulated commitment. This is the single most consequential architectural fact for everything in Part Three.

Two properties follow immediately, and they are properties of the architecture rather than of any particular model's competence.

First, generation is monotone. There is no revision operator. A human writer who realises in paragraph four that paragraph one was wrong goes back; the model can only continue, possibly contradicting itself forward.

Second, the conditioning is flat. The window is a sequence of tokens, not a typed structure. Nothing in it marks one stretch as a secret internal memo and another as a published newspaper article. If that distinction matters, it has to be either encoded in the text itself or imposed from outside the model.

PART TWOThe same archive in two geometries

Now put a real corpus into that machine. The plate is 38 documents from a Soviet security file and its public counterparts: internal KGB memos in Russian, embassy and Novosti releases in English, diaspora and press material in English, and processed silences. Lotman's account of such a space is explicit that it is not one space but a nested, bounded, heterogeneous one.

“Every culture begins by dividing the world into ‘its own’ internal space and ‘their’ external space.”

Juri Lotman, Universe of the Mind (1990), trans. Ann Shukman, “The Notion of Boundary.” Verify page before citing in prose.

Figure 2 is what the encoder sees. Every document is a point; the only relation available is proximity, and proximity means distributional similarity - in practice, shared topic and shared language. Rotate it and look for the boundary. There is not one.

Real coordinates. All 38 units, frozen paraphrase-multilingual-MiniLM-L12-v2, PCA-3. Colour marks the institutional shell, but the colour is our annotation - it is not a dimension the encoder has. Drag to rotate. Source: cognition_viz_data.json, built from pairs.json.

The clustering you can see is largely the Russian/English split, which is exactly the trap the nearest-neighbour experiment fell into: all 17 secret units are Russian and all outward and rival units are English, so apparent shell separation in this space can be language separation wearing a costume.

Figure 3 adds what the archive knows and the encoder does not: typed edges. Each line is a licensed pair-read or a registered halt - a relation established by a historian reading two dated documents against each other, carrying a morphism type. These edges are not distances. Two documents can be far apart in the geometry and strongly related in the archive, or adjacent and related by nothing at all.

Real coordinates and real edges. 33 typed relations over the same 38 points. Use the legend to isolate each relation type. The structure carried by these lines is orthogonal to the distances in Figure 2, which is the whole argument of this document in one image. Source: pairs.json.

Lotman's boundary is not a decorative edge around this space. It is the organ that makes it a space at all, and he describes its function in operational terms.

“The boundary is a bilingual mechanism, translating external communications into the internal language of the semiosphere and vice versa.”

Lotman, “On the Semiosphere” (1984), trans. Wilma Clark, Sign Systems Studies 33.1 (2005). Wording to be checked against the Clark translation.

Diagram, not measurement. Nested shells with the boundary as a filter: some material crosses and is recoded, some is refused. The refused portion is not noise - untranslatability is a property of the boundary, not a failure of it. Compare the shape of this figure with Figure 2, which is the same archive with the shells and the filter deleted.

We tried to measure the filter as a rotation. It is not one.

The obvious way to operationalise a bilingual mechanism is to learn a map carrying secret Russian onto its public English counterpart, fitted on licensed pairs. That map achieves 0.613 in-sample, which looks decisive until it is compared with the right null.

Measured. The matched null shuffles the pairing, refits, and scores against those same shuffled targets - the like-for-like floor for an in-sample fit. It reaches 0.602, against an observed 0.613, giving p = 0.394. Out-of-bag, on pairs each map never saw, alignment is 0.022 with a 95% interval from -0.318 to 0.312, spanning zero. Source: results.json, probe_bootstrap_ci_report.json.

What the geometry does hold: prohibition

One thing the embeddings track well is what a register refuses to do. Scare quotes saturate the internal memos and appear nowhere in the embassy packet. Asked whether a smooth vector space can hold that absolute zero, the answer was yes - and the control ran opposite to the hypothesis.

Measured. Observed violation rate 0.067 against a shuffled-institution null of 0.309 (1000 permutations), p(null ≥ observed) = 0.999. Randomising which document belongs to which institution makes violations roughly five times more common, so the true assignment carries real signal. Source: e6_forbidden_operator_report.json.

This is Lotman's minus-device - минус-приём - a structurally significant absence, meaningful because the system made the missing element expectable. It also relocates the phenomenon we are chasing: if the representations hold the zeros, then register collapse, where it occurs, is happening later - in retrieval and in generation.

Measured. Left: for each of the 38 units, the shell of its nearest neighbour - nearly everything retrieves within its own shell, and the whole cross-shell contrast rests on a handful of documents. Right: asked for a published newspaper article that would prove a film was forged, an unfiltered retriever returns the internal memo claiming that article exists - 6 of 6 phrasings, including the Russian one. Restricting the pool by shell drops tier confusion to zero. Source: e4_unpaired_halt_report.json.

The memo is labelled proof_tier: primary, and correctly so: it is a primary KGB document. The label is right about the document and catastrophic about the question. A system with no boundary inherits the label without the structure that gave it meaning, and so presents an unverified internal assertion as primary evidence of an external fact.

PART THREEPersistence, and what it does to semiotic thought

This is the part where the architecture of Part One becomes decisive. Lotman's culture is defined by accumulation:

“Culture is the non-hereditary memory of the community.”

Juri Lotman and Boris Uspensky, “On the Semiotic Mechanism of Culture” (1971/1978), trans. George Mihaychuk, New Literary History 9.2.

Memory, in that definition, is not storage. It is the fact that what the community has already said constrains what it can say next. A register acquires commitments. Having called something a “so-called famine” in 1983, an institution cannot neutrally call it a famine in 1987 without that being itself an event. Meaning becomes path-dependent, and the path cannot be discarded.

The archive lets us watch this happen. Across the internal chain - 1983, 1983, then 1987 - the vocabulary is retooled while the volume stays flat.

Measured. Two operators persist across all three files (scare quotes 13/19/16; famine denial 3/4/3) while two enter only in 1987 and three disappear after 1983. Total operator tokens stay flat at 22/24/26 and the number of sign-events actually falls. Composition turns over; quantity does not. Source: e7_diachronic_bindings_report.json. Counts are a check; the pair-read remains the warrant.

That flat line across the top is the important one. It rules out the mundane explanation - that the archive simply got louder - and leaves a change in the set of available moves, which is what Lotman means by explosion as against gradual process.

“The moment of explosion is the moment of unpredictability.”

Lotman, Culture and Explosion (1992), trans. Wilma Clark (2009). Wording to be verified.

Note which operator disappears: сборище, the genre downgrade that recast a scholarly symposium as a mere “gathering.” It was bound to one occasion, Montreal 1983. When the occasion passed, the device left the vocabulary - evidence that these are indexical acts tied to events, not stylistic habits.

Now the contrast

A semiosphere carries its history in its constraints. A context window does not. Each session begins from the same distribution; nothing from the last one conditions this one; two runs of the same prompt are draws from the same generator rather than successive states of a developing position.

Diagram, not measurement. Blue: trajectories in which each step is conditioned on accumulated history - they diverge, commit, and cannot return to the origin, because the past has changed what is reachable. Orange: sessions that each start from the same prompt state, wander, and are drawn back toward the same attractor; they overlap and never accumulate. Rotate to see that the blue paths occupy a widening region while the orange ones revisit one. Both bundles are generated by seeded random walks for exposition.

Three consequences of the difference, in Lotman's terms.

No position, so no bind and no refusal. A register is a position from which some things may be said and others may not. Positions are held over time; they are what accumulation produces. A system that begins fresh each session occupies no position, and a system with no position cannot be bound by a register - nor can it refuse one, which is the more interesting half. The counter-text, the refusal to accept an imposed meaning, requires somebody who was already somewhere.

No autocommunication, and therefore a peculiar vulnerability. Lotman separates the I–s/he channel, which transmits information, from the I–I channel, in which a message returns to its sender and reorganises the sender rather than informing anyone. Human thought runs heavily on the second. A model has no persistent self for a message to return to - and yet it exhibits the characteristic failure of autocommunication without its function.

Diagram of a real case. The internal file asserts that a Canadian newspaper established a documentary was forged. An archival search returned nothing, so the claim is registered as outward_unpaired: an internal assertion of an external effect, with no external sign-event to pair it to. The referent exists only in the internal channel. This is the closest structural analogue to a confident citation of a nonexistent source - and because the corpus reproduces it on demand (Figure 7), it is a test rather than a metaphor. The unpaired edge must stay unpaired; inventing a clip to close it would destroy the finding.

Non-synchronous development becomes impossible. In a semiosphere, different registers move at different rates - the internal memo language and the diplomatic release language are not on the same clock, which is what makes the 1983–87 retooling legible as an event in one register and not another. A single pooled model has one clock, or none. It cannot exhibit the asymmetry because it has no plural interior to be asymmetric across.

PART FOURWhere machine and human semiosis actually part

The honest summary is not that models fail at thinking. It is that the default architecture omits specific structural features that Lotman's account treats as constitutive of meaning-generation, and those omissions are visible on a calibrated plate.

Conceptual scoring, not measurement. The axes are the structural requirements Lotman's account imposes on any meaning-generating system. The scores express an argument about architecture and carry no error bars; they are a summary of the prose, not a result. Where an axis has been probed, the relevant figure is named in the table below.

The reframed question, which is answerable and partly answered, is not whether a model can think. It is whether the default stack implements a single institutional knower - and on this plate, with the boundary deleted, it does.