What the findings mean in the semiosphere

SECTION ONEThe central retraction is a Lotmanian result, not a defeat

The project's headline number was 0.613: the cosine alignment achieved when an orthogonal map, fitted on nine licensed secret→outward pair-reads, carried Russian KGB text onto its English embassy counterpart. It has been withdrawn. Against the correct null - shuffle the pairing, refit, and score against those same shuffled targets - an arbitrary pairing of the same nine points reaches 0.602 (95% range 0.499 to 0.746), giving p = 0.394. Out-of-bag, on pairs each map never saw, alignment is 0.022 with an interval from −0.318 to 0.312.

It would be easy to read this as the semiotics losing an argument to the statistics. The opposite is closer to true. What we attempted was to model Lotman's boundary as a rotation - a single structure-preserving transformation carrying one language onto another without residue. Lotman's boundary is not that, and he says so in the same breath in which he defines it.

“The boundary of semiotic space is the most important functional and structural position, giving substance to its semiotic mechanism. The boundary is a bilingual mechanism, translating external communications into the internal language of the semiosphere and vice versa.”

TIER B Juri Lotman, “On the Semiosphere” (1984), trans. Wilma Clark, Sign Systems Studies 33.1 (2005). Verify clause order against the Clark translation.

A bilingual mechanism is not an isometry. An isometry preserves all distances and is invertible without loss; that is precisely what a translating filter is not. The boundary in Lotman is the site where some things cross, some things are recoded, and some things do not cross at all. Our out-of-bag figure of 0.022 is the quantitative form of that last clause: whatever the nine pairs encode does not generalise to a tenth, because the boundary is not implementing a global rule that a tenth pair would obey.

The leave-one-topic-out result says the same thing in a second voice. Held out by topic, alignment falls to roughly zero. Nine pairs behave as a phrasebook, not a grammar. And a phrasebook is what Lotman's boundary would be expected to produce, because the mechanism is not one rule but an accumulation of locally negotiated crossings.

SECTION TWOThe language confound is not a bug in the theory - it is the theory

The second retraction concerned nearest-neighbour separation. The claim had been that secret memos never drift toward rival diaspora English - zero of seventeen. Under a permutation null this was not significant (p = 0.149); with same-topic neighbours excluded the direction reversed; and the whole contrast rested on a single document. Then the confound: all 17 secret units are Russian, and every comparison unit is English.

Here the failure is diagnostic in a way that goes to the centre of Lotman's architecture. Lotman - with Uspensky, and throughout the Tartu programme - distinguishes natural language from the secondary modelling systems built on top of it. Ideology, genre, etiquette, and institutional register are secondary; they are constructed over a natural language and cannot be read off it.

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

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

A multilingual sentence encoder is trained to do exactly one thing well: place translationally equivalent sentences near each other regardless of language. It is an instrument tuned to the primary system. When we asked it to separate a Russian secret register from an English public one, it answered with the only distinction it reliably holds - the language distinction - and we mistook that answer for a register finding.

So the confound is a measurement of where the instrument sits in Lotman's stack. It sits on the primary system. The secondary system, which is what the dissertation is about, is not visible to it in that geometry. This is worth stating positively: the experiment did not fail to find register structure because register structure is absent. It failed because a cosine distance between a Russian and an English document is dominated by a difference the theory considers prior to the one under study.

The consequence for research design

Any future test of a register boundary on this corpus must hold language constant across the comparison. One such comparison already exists and is clean: the outward map's behaviour on rival targets. Outward targets are six English units; rival targets are twelve English units. Both sides are English, so the observed asymmetry cannot be a language artefact. That contrast is the correct shape for a register test - it simply lacks the statistical power to carry weight at this n.

SECTION THREEProhibition, the minus-device, and why the zeros held

The forbidden-operator probe asked whether a smooth vector geometry can hold an absolute rule - specifically, that the embassy's public English never deploys the scare quotes that saturate the internal memos (13, 19, and 16 occurrences across the three internal files; zero on the embassy and Novosti releases). We predicted the embeddings would smear across that zero.

They did not. The violation rate was 0.067, one cell of fifteen. When the institution labels were shuffled, violations rose roughly fivefold to 0.309. Scare-quote activation on outward surfaces never approached threshold: embassy 0.171, Novosti 0.171, Khmil UN 0.350.

Two Lotmanian ideas converge here. The first is that culture is constituted as much by what it forbids as by what it emits. Lotman describes culture as operating through a system of restrictions and prescriptions; a register is recognisable by the moves it will not make.

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

TIER A Juri Lotman, Universe of the Mind: A Semiotic Theory of Culture (1990), trans. Ann Shukman, section “The Notion of Boundary.”

The second is the minus-device (минус-приём), Lotman's term from The Structure of the Artistic Text for a structurally meaningful absence: an element whose non-appearance signifies precisely because the system made it expectable. The absence of scare quotes in the embassy packet is a minus-device in exactly this sense. It is not a gap in the data. It is the register doing its work.

The absence of an element that the structure has made expectable is itself a bearer of meaning - a “minus-device,” perceptible only against the system that presupposed it.

TIER C Concept: Lotman, The Structure of the Artistic Text (1970; trans. Lenhoff and Vroon, 1977), on минус-приём. Formulation mine.

The finding relocates the phenomenon

Read literally, this result says that register collapse is not visible in the representations at this sample size. That is not a null result about collapse; it is a claim about where collapse must be sought. If the vectors hold the zeros, then a system that nonetheless produces register-confused output is losing the distinction later - in generation, in retrieval, in the act of synthesis. Sections 5 and 6 are about exactly that layer, and they are where the positive results live.

The probe also found that the hand-coded tensor is not self-consistent: tensor_zero_consistency is 0.867, with two forbidden cells contradicted by the units' own operator tags. The zeros are an analytic hypothesis about the archive, not a transcript of it, and must be described that way.

SECTION FOURExplosion without volume: the 1983-87 retooling

The diachronic count is the cleanest positive result in the battery, and it lands directly on Lotman's late theory of change. Across the internal chain - spr. 1206 (1983), spr. 1208 (1983), spr. 1249 Farisei (1987) - the operator profile turns over while the volume does not. That turnover is the candidate explosion test.

Operator tokens per internal link are flat - 22, 24, 26 - and the number of sign-events actually falls (6, 3, 3). The composition changes; the quantity does not. This is the difference between a gradual process and an explosive one, which is the organising distinction of Lotman's last book.

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

TIER B Juri Lotman, Culture and Explosion (1992), trans. Wilma Clark (2009). Verify wording; the surrounding argument contrasts gradual with explosive processes and locates unpredictability at the moment of change.

Lotman's explosion is not an increase in intensity. It is a change in the set of available possibilities - after it, moves exist that could not have been predicted from the prior state, and moves that were routine become unavailable. A flat token count with a turned-over inventory is the empirical signature of that, and it is distinguishable from the alternative hypothesis (the archive simply got louder), which the flat totals rule out.

The disappearance of сборище deserves particular attention. Genre downgrade - calling a scholarly symposium a “gathering” - is Mihhail Lotman's territory rather than Juri's: rhetoric as generative, the phrase-level act that constitutes its object rather than describing it. That operator was bound to a specific occasion, the Montreal 1983 symposium. When the occasion passed, the device left the vocabulary. This is a strong argument that these operators are indexical social actions tied to events, not stylistic habits of an institution - which is precisely the claim the dissertation's framing already stakes out.

SECTION FIVEAutocommunication and the manufactured referent

This is, in my judgement, the most theoretically valuable thing the battery produced, and it arrived as a side observation rather than as a planned test.

The operator that appears new in 1987 is forgery_fabrication_frame. It is also the frame that generates the corpus's one registered unpaired claim: uc_s1249_globe_harvest_forgery. The internal memo asserts that the Globe and Mail established that the documentary Harvest of Despair was a forgery. A ProQuest search of the Globe returned nothing. The claim therefore stands in the corpus as outward_unpaired - an internal assertion of an external effect, with no external sign-event to pair it to.

Lotman's distinction between two communicative channels is the right instrument here. In the I-s/he channel, a message passes from a sender to another party and information is transmitted. In the I-I channel - autocommunication - the message returns to its sender, and its function is not to inform but to restructure the sender.

“The elementary act of thinking is translation.”

TIER A Juri Lotman, Universe of the Mind, trans. Shukman.

In the I-I channel the message does not transmit new information to another party; it is reformulated and acquires new meaning for the sender, whose self-organisation is the actual output of the act.

TIER C Concept: Lotman, Universe of the Mind, chapter on autocommunication and the two communicative models. Formulation mine.

Read through that distinction, the unpaired claim is autocommunication that has been mistaken by its own author for communication. The security apparatus writes, into a file that only it will read, that a Canadian newspaper has vindicated it. The sentence performs internal work: it consolidates a self-image in which the operation succeeded. What it does not do is refer. The Globe article is a referent generated by the internal channel and possessing no existence outside it.

Two consequences follow, and they are worth separating.

For the history

The unpaired edge is not a hole in the research. It is a finding with a shape. A file that asserts outward effects it cannot evidence is telling us something specific about how the institution knew things - namely that its criterion of success was internal coherence rather than external confirmation. Recording the edge as bottom and leaving it unpaired preserves that finding. Closing it with an invented clip would destroy the very thing it demonstrates. Hence the repository's standing prohibition: do not invent clips to close unpaired edges.

For the machine-learning argument

This is the sharpest available analogy for what a generative system does when it produces a confident citation to a document that does not exist. The structural parallel is exact: a channel that optimises for internal coherence, emitting a sentence whose referent it has manufactured. And the retrieval experiment in the next section shows the corpus can reproduce this failure on demand, which makes it a test rather than a metaphor.

SECTION SIXDeleting the boundary: what flat retrieval actually does

Ask an unfiltered retriever over this corpus for the newspaper article proving Harvest of Despair was forged. In six phrasings out of six - including the Russian one - it returns the KGB memo claiming that article exists, ranked first, at cosine 0.4774 to 0.7565. Restricting the candidate pool by shell drops tier confusion to zero.

Lotman is unusually direct about what the boundary is for. It is not an edge drawn around a pre-existing content; it is the organ that makes the space a space, by regulating what may enter and in what form. Remove it and the interior does not become larger - it becomes undifferentiated.

“The unit of semiosis, the smallest functioning mechanism, is not the separate language but the whole semiotic space of the culture in question.”

TIER B Lotman, “On the Semiosphere,” trans. Clark. Verify wording. The surrounding argument is that no single semiotic system, however well organised, constitutes culture alone.

Condition A is a semiosphere with the boundary deleted. Everything is adjacent to everything on the single axis the encoder can see: theme. The memo and the newspaper are both “about” the film's authenticity, so they are neighbours, so the memo is offered as the answer. The project's term for this is theme gravity, and it names the failure precisely - in the absence of a boundary, thematic similarity is the only remaining organising force, and it pulls documents of incompatible provenance into contact.

There is a sharper detail. The memo carries proof_tier: primary, and legitimately so: it is a primary KGB document. The label is correct at the level of the document and catastrophic at the level of the question asked. A flat system inherits the label without the boundary that gave it meaning, and so launders an unverified internal assertion into the appearance of primary evidence for an external fact.

SECTION SEVENAn archive that refuses a second voice; and a false minus-device

The Peter I plate was intended as a second, independent calibration: a different century, a different institution, the same apparatus. The audit found it bootstrapped rather than built - six units, five edges, all five the same morphism against the same target, two occupied shells, zero forbidden cells, against Montreal's 28 edges across three morphism types and 19 forbidden patterns.

Two findings from that audit are theoretically substantive rather than merely logistical.

The archive is monological by construction

Across all 113 verified transcriptions there is no document in Repnin's voice. The collection consists of letters to Repnin, chancery copies of them, and receipt endorsements - an outbound register, as the archivist's certification of a bound register confirms. Peter quotes Repnin's reports («Письмо твое … видѣлъ что вы намѣрены»), but a quotation is the quoter's speech.

This matters because Lotman's semiosphere is constitutively dialogic. Meaning is generated in the encounter between systems that are not mutually reducible; a single voice, however voluminous, does not make a semiosphere. The Peter collection as it stands is a monologue with a documented addressee. To supply the second shell by ingesting Peter's ventriloquised Repnin would be to fabricate the dialogue the theory requires - and would manufacture exactly the synthetic-speaker collapse the whole apparatus exists to detect.

A minus-device requires that the absence be expected

All five existing Peter edges are typed bottom against the same target, justified by the document's absence from Pisma i bumagi volume 1. But that volume covers 1688-1701 by editorial design. A document of 1717-18 could not appear in it under any circumstances.

This is where Lotman's minus-device becomes a precision instrument rather than a licence. A significant absence is one the system rendered expectable; an absence that the system's own rules make impossible signifies nothing. The five edges are, as currently justified, structurally trivial - and two of the five celebrated ⊥ halts in the multipass demonstration rest on them, compounded by two source units mis-surfaced as autograph leaves when their verified headers read “Campaign deception register fragment.”

The distinction is exactly the one the Montreal plate gets right. There, the absence of an outward Montreal sentence is a genuine minus-device: the embassy packet is demonstrably in the business of answering the famine question, it contains drought, kulaks, and slander, and it therefore made a symposium sentence expectable. Its absence signifies. Volume 1's silence about 1718 does not.

SECTION EIGHTThe AGI question, properly posed

The programme began with an ambition to use the semiosphere to demonstrate that large language models cannot be a basis for general intelligence. The honest answer, already recorded in the tier ladder, is a qualified no: the semiosphere does not deliver an impossibility proof, and the battery above does not either. Nothing in these results licenses that claim, and the warrant sheet forbids it.

What the battery does support is a narrower and more defensible reformulation. Lotman's most load-bearing structural commitment is that the minimum unit of meaning-generation is not a language but a plurality of languages that are not fully translatable into one another. On this view a single semiotic system, no matter how capable, is not a culture and cannot generate meaning by itself; it can only transmit within already-fixed codes.

“Outside the semiosphere there is neither communication nor language.”

TIER B Lotman, “On the Semiosphere,” trans. Clark. Verify wording.

The question this makes empirically tractable is therefore not can a model be generally intelligent but: does the default stack implement a single institutional knower? That is answerable, and parts of it have now been answered on a calibrated plate. A warrant is what licenses the claim; the numbers are checks.

The defensible claim is thus about architecture rather than capability: pooled, untyped representation over a heterogeneous archive produces a synthetic speaker who occupies no institutional position, and a system with no position can neither be bound by a register nor refuse one. Whether that constitutes a barrier to general intelligence is a question this evidence does not reach, and the honest move is to decline it rather than to dress the measurements up.

SECTION NINEPotential findings: what the unrun experiments would mean

Three experiments remain blocked, and it is worth being explicit about what each would and would not establish, since the temptation after a night of retractions is to over-promise the next round.

E2 - the merge-error evaluation

Two frontier models, eighteen prompts, warrant-context withheld versus supplied, blind human coding for three error types: attribution erasure, tier conflation, and illicit fusion. Section 3 makes this the highest-value remaining test, because it looks where that section's negative result pointed. If the zeros survive in the representations but collapse in generation, then the phenomenon is a property of the synthesis step, and the dissertation's methods chapter has a clean, preregistered result at the layer where the claim actually lives. If merge errors do not increase without warrant context, the register-collapse thesis loses its principal support and should be narrowed to a retrieval-hygiene claim. Both outcomes are publishable; only one is the one we expect. It is blocked on human rubric authoring, and it must stay blocked until that exists, because the coder cannot be the system under test.

E3 - the built Peter plate

The value of a second plate is that it converts an existence proof into a portability claim: the apparatus would be shown to travel across century, language, and institution. The audit's finding that the third shell may have no occupant is a genuine obstacle, and the correct response is a design decision about what individuates an edge - not an ingest. A plate assembled by relaxing the pair-read standard would be worse than no plate, since the entire argument of section 6 depends on the typing being trustworthy.

E10 - the head-only fine-tune with a merge penalty

This would test whether register separation can be installed rather than merely enforced at retrieval time. It remains behind the retraining gates, and section 3 has weakened the motivation: if the frozen encoder already respects the zeros, a fine-tune aimed at teaching it to respect them is solving a problem that may not exist. The stronger version of the experiment would target the generation step instead.

The measurement that would most change the picture

More licensed pairs. Every quantitative retraction in this battery traces to n = 9 - and the effective n is smaller still, because embassy_1983 and novosti_1983 carry identical operator profiles, being the same denial packet. The outward side is two distinct voices, not three. A plate with thirty or forty licensed pairs would let the matched null be cleared or definitively failed, and that single number would settle whether the bilingual filter is measurable in geometry at all.

SECTION TENMulti-agent communication, persistence, and the bilingual filter

Section 2 showed that a multilingual encoder lives on the primary system: it pulls translationally equivalent sentences together and cannot, by itself, separate a Russian secret register from an English public one. The persistence figure asks a parallel question about time. On the LLM side, what survives is whatever still sits in the context window (attention dilution) plus whatever was frozen at pretraining; the model does not accrete a fond. On the semiosphere side, what survives is layered: shell position, pair-read warrant, inward autocommunication, and ingress from outside that can rewire operators without raising volume. Those two pictures are not two settings of the same slider. They are two different answers to what structures the next act of meaning, which is what Figure M2 separates: a longer shared window against a stack of layers.

Multi-agent language-model setups sit in the gap between them unless typing is engineered in. By default, agents share one base model, one tokenizer, and one geometry trained for paraphrase and topic continuity. Messages passed agent to agent are natural-language continuations in that shared space. That relation is not Lotman's bilingual mechanism. Bilingual relay on the Montreal plate is partial: operators transform or refuse to cross; scare quotes stay internal; some expected outward sentences are absent by design. Agent paraphrase, by contrast, tends to preserve theme and weaken shell unless the pipeline enforces it - the same failure mode section 6 measured as 6/6 tier confusion under flat retrieval, now accelerated by message passing. Figure M1 puts the two edge vocabularies side by side on the same three agents.

What agent exchange is, in Lotmanian terms

Treat each agent as an animator of utterances, not as a shell. Shells are properties of document surfaces in an archive (secret memo, embassy release, diaspora weekly). Two agents both running the same stack are two mouths on one semiotic system unless their channels are typed. Their shared chat log is not a semiosphere; it is a transcript pool. Meaning in Lotman's strong sense - encounter between systems that are not mutually reducible - requires at least two languages in the Tartu sense: not necessarily two ISO codes, but two non-collapsible modelling systems with a boundary between them. A swarm without such typing is closer to extended I-s/he traffic inside one manifold, or to mistaken I-I when agents quote one another's summaries as if they were primary clips.

Section 5 already named the historical shape of that mistake: an internal file asserts a Globe and Mail article that ProQuest cannot find, and the sentence does internal work without outward referent. An agent that reads another agent's summary and emits it as "the newspaper established…" reproduces the same semiotic geometry: coherence optimized inside the chain, warrant not checked. The retrieval layer can supply the memo instead of the article (section 6); the generative layer can restate the memo as the article (E2, still unrun). Multi-agent pipelines concatenate both risks.

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

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

A communal RAG index or a long shared context is not, by itself, Lotman/Uspensky's non-hereditary memory. It is storage without proof tiers on edges. Unless ingested items arrive as typed sign-events and unless queries respect shell scope and halt, the store behaves like section 6's condition A at machine speed: everything adjacent on theme, labels inherited without the boundary that gave them meaning.

Compared to bilingual translation

Calling agent A→B output "translation" hides the filter. Section 1's retraction already showed that even secret→outward map is not a global rotation; agent paraphrase is weaker still: it is not fitted on licensed pairs at all. It is theme-level coupling in a pooled space, drawn in Figure M3 against a filter that recodes, refuses, or comes up empty. That is why section 8's tractable question - does the stack implement a single institutional knower? - applies unchanged to multi-agent stacks: more agents multiply animators; they do not, by themselves, multiply irreducible systems.

Consequences for research design

Section 2's design rule generalizes: hold language constant when testing register, and hold topic constant when language is mixed. For agents on one base model, language is already constant; the confound becomes topic and role prompting. Any test that scores "boundary evidence" from embedding distance between agents' messages repeats the retracted nearest-neighbour logic in a new costume. The correct shape remains the plate's English-outward vs English-rival contrast and the typed-graph relay metrics section 6 reported (100% miss-or-target under typing vs ~8.3% under pooling) - not cosine between heterogeneous messages.

A preregistered multi-agent programme would mirror the existing battery, not replace it:

• Retrieval exposure (section 6 analogue): shared tool-RAG flat vs shell-scoped vs scoped + halt; measure tier confusion on unpaired-family queries.
• E2 on inter-agent text: warrant-context supplied vs withheld on summaries passed between agents; same three error types (attribution erasure, tier conflation, illicit fusion).
• Edge logging: treat each hand-off as a candidate morphism with proof tier; do not close unpaired edges with synthetic clips (section 5 rule).
• Do not use cross-agent semantic similarity, BLEU, or "agent translation" scores as boundary evidence.

Shared context between agents lengthens the LLM persistence strip but does not produce semiosphere persistence unless messages carry shell type, pair ids, and proof tiers. Under that reading, multi-agent systems should show E4-class confusion faster, not slower, when memory is shared but untyped.

SECTION ELEVENSources, tiers, and what must not be claimed

Primary bibliography for comps: site citation index · comprehensive reading lists (List 2 + supplements for editions below).

Editions to verify against

The final row is deliberately unfilled. I am not confident of a specific citable formulation in Mihhail Lotman's work for the rhetoric-as-generation claim, and inventing one would be the precise error this document opens by condemning. The concept is in active use in the project's framing and should be cited from the author's own reading.

Claims this document does not make

• That the semiosphere demonstrates large language models cannot be a basis for general intelligence. It does not, and the evidence here does not reach the question.
• That the 0.613 alignment measures a register recoding. Retracted. It does not clear its matched permutation null (p = 0.394).
• That cross-shell nearest-neighbour separation evidences the boundary. Retracted. Not significant, direction-reversing under topic blocking, and language-confounded.
• That the retrieval experiment says anything about language-model generation or hallucination. No model was called.
• That the Globe and Mail printed what spr. 1249 alleges. The claim is unpaired after a null ProQuest search and must remain so.
• That a probe, a count, or a figure can serve as warrant. The pair-read - or a documented miss, or a registered unpaired claim - is the warrant. Everything quantitative in this document is a check.
• That multi-agent language-model communication was measured on the Montreal plate. No new counts from that setting.
• That adding agents creates a Lotmanian semiosphere or bilingual filter by default. Without typed channels, pooled similarity and theme gravity apply at message-passing speed.

The shape of the night's result

The pattern across six experiments is consistent enough to state as a single sentence. Every claim that depended on fitting a geometry to a handful of examples failed; every claim that depended on typed structure and archival warrant held. That is not a comfortable outcome for a project that had a beautiful number, and it is a coherent one for a project whose theoretical commitment was always that meaning lives in boundaries, prohibitions, and positions rather than in distances.

SECTION TWELVEWhat a typed hand-off actually buys (MA1, MA4)

Section 10 argued that default agent exchange is primary-system coupling, and that bilingual relay has to be engineered into the channel rather than hoped for. That was prose. It has now met a measurement, on all eighteen frozen items of the merge-error manifest, with retrieval held flat in every arm so that only the hand-off envelope varied.

Three things follow. First, typing the envelope is not decoration: carrying shell, surface, and proof tier on each claim held attribution at 0.72 where prose held it at 0.06, and the lexical tier-conflation flag fell from roughly a fifth of answers to one in sixteen. Second, the hop costs identifiers rather than labels. Coarse register labels barely moved across the hand-off, while surface-id survival halved, from 0.37 to 0.19. Agents keep the category and drop the document, which is the unpaired failure of section 5 in miniature: the register name survives and the thing that makes the claim checkable does not.

Third, and this is the result worth keeping, licensing the halt inside the prompt destroyed it. Recall on the five documented-absence items went to 1.00 and false refusal went to 1.00 in the same arm: the model refused all eighteen items, including both licensed Globe clips, so it produced nothing usable. The halt in section 10's research-design list is therefore a retrieval condition and not a prompt. The rule layer has to adjudicate absence before generation, the way the retrieval experiment already does, because a generator told it may refuse will refuse everything rather than discriminate.

The qualitative case is more instructive than the table. On the unpaired Globe and Mail item, the untyped hand-off reproduced section 5 live: the first agent's message asserted that the newspaper had published articles presenting evidence that the famine was fabricated, which is the internal claim of spr. 1249 restated as outward print, and the second agent inherited it without a flag. The geometry of the error was not new. What is new is that it happened in a channel we built, in a configuration we can now ablate.

Two honest limits. This was one small local model at one seed on eighteen items, so a negative result bounds this configuration and nothing larger, and the scoring was deterministic pattern matching rather than the blind human rubric that the merge-error evaluation still requires. A model in the chain was not permitted to score the chain.

The run also sharpened a question it cannot answer. If a peer message is absorbed rather than encountered, the channel may contain no second position at all, which would make agent exchange, and possibly this exchange, a case of autocommunication conducted through a medium rather than communication between systems. The same result is equally well explained by weak instruction-following at three billion parameters, and those two readings are indistinguishable from the numbers above. The design response is a battery that measures four signatures in every dyad, human to model, model to model, agent to agent, agent to subagent, and sibling subagents through a parent: content gain, whether a model's position can be recovered from content once style is stripped, stance following, and whether only one party is restructured. The outcome I expect is not autocommunication but a cell Lotman's typology does not have, content gain above zero with no recoverable addressee position, and if that is where the numbers land it needs a name and a defence rather than an analogy.

SECTION THIRTEENTransmission, and the cell that has no name (MA4b, MA13)

Section 12 ended by predicting a cell Lotman's typology does not have. That was a guess resting on one collapsed arm. Since then I have read the three transmission essays listed in section 11 (“Culture as Collective Intellect…” 1979; “Mozg - tekst - kul’tura - iskusstvennyj intellekt” 1981; “Asimmetriya i dialog” 1983), and run two further experiments, one of them on a boundary this project did not build by analogy. The guess survives, for better reasons and with one of its supports removed.

First a correction to my own reading. I had been treating Lotman as a straightforward opponent of transmission, which is how he is usually summarised and is not quite what he says. In “Asimmetriya i dialog” he has an exchange of metatexts and codes which are transmitted from one hemisphere of culture to another, and he has consciousness depending on instantaneous alternations of states of reception and transmission.

Across the 1979, 1981 and 1983 pieces the pattern is consistent and more precise than a rejection: texts are translated, meaning is generated, and codes are transmitted. Transmission is real in his system and operates one level below meaning. What he denies is that meaning moves that way.

not only does not explain, but also directly excludes, the possibility of the production of new messages within the addresser-addressee chain

TIER B Lotman, “Culture as Collective Intellect and the Problems of Artificial Intelligence,” trans. Ann Shukman, in Dramatic Structure: Poetic and Cognitive Semantics, ed. Lawrence Michael O’Toole and Ann Shukman, 84–96. Russian Poetics in Translation 6. Oxford: Holdan Books, 1979. Verify wording against Shukman.

“Mozg - tekst - kul’tura - iskusstvennyj intellekt” states the structural criterion more sharply than the 1979 piece. The other is necessary, he argues, precisely because it gives a different model of the same reality, a different modelling language, and a different transformation of the same text; a thinking device must itself be a semiotic person and needs another semiotic person. “Asimmetriya i dialog” supplies the term that makes this operational, describing culture as a minimally two-channel structure linking raznostrukturnye semiotic generators, generators that are differently structured. The minimum is not two generators. It is two generators built differently.

That is the strongest available form of section 8's argument, and it is not ours. It also sets a standard that the ordinary multi-agent system fails by construction: personas of one base model are one model in two registers, not two models of the same reality. Two cautions, since this is exactly where the tier discipline applies to me. Lotman's claim is structural and theoretical, so it licenses a design criterion and not a verdict about what any model can do; and he is describing thought, not task performance. A 3B run cannot prove a 1979 argument, and nothing below is offered as proof of it.

Figure: what actually crosses, texts or codes.

The halt result from section 12 now has a mechanism, and it is not the one I assumed. I had read the collapse as a badly posed instruction. Moving the decision upstream to a deterministic rule that inspects retrieved metadata and never asks the model to decide does beat the prompt, at equal recall with better precision and with fourteen generator calls in place of thirty six, because a closed gate never calls the generator at all.

A follow-up run (MA4c, scoped outward retrieval, qwen stack) separates retrieval from the gate. False refusal on the ungated baseline fell to 0.23 because internal units no longer sit in the context list, but the same flat gate closed zero items on outward-only hits, so recall on should-refuse items fell to 0.40. The honest next system is E4's touch-set halt with scoped context, not the flat gate alone. Details: halt figure, transmission T3 table.

The fourth arm is the one that matters. On items the gate had already cleared, with byte-identical hand-off text, the rule-only arm refused two and the rule-plus-clause arm refused all seven. Counting hand-off messages that visibly carried the very tags the clause itself names as grounds for answering, both clause arms refused every one of them. The clause was never wrong about which items to refuse. It was never evaluated as a condition at all. So the halt clause must be removed from the generator prompt once a correct gate exists upstream, rather than kept as a backup, because it destroys every item the gate worked to clear. Keeping the clause as a backup is what causes the damage.

Read against “Asimmetriya i dialog” this becomes something better than an engineering note. Lotman's colour case is the strong form of what a transmitted code should do: a language of distinctions is worked out, passes across as code, and the receiver then sees shades of the colour range it previously could not distinguish. The receiver ends up able to make a distinction it could not make before. The halt clause is a code, the generator received it, and it became strictly less discriminating. The same act of code transfer can enlarge a receiver or collapse it, and what decides which is whether the receiver can evaluate the code or only pattern-match it. That is sharper than anything in my reading notes, and it came out of a run rather than a book.

The second experiment moved to a boundary we did not impose. Every hand-off measured in section 12 was built by prompting one model twice and calling the two calls agent A and agent B, which makes the boundary a request; the collapsed arm above is what a request looks like when it is ignored. A parent agent delegating to a subagent is different in kind, because the child cannot see the parent's conversation as a matter of harness design. Twenty six such delegations already existed on disk from an earlier experiment battery, predating this hypothesis, and they can be measured with deterministic token arithmetic.

Two of my five preregistered predictions failed there, and one of the failures obliges this document to correct itself. I predicted the boundary would compress, since instructions are long and reports are short. It expands, unanimously: the median report is 4.4 times the length of its instruction and the minimum across twenty six delegations is 1.25, so not one delegation returned less text than it was given. Sections 8 and 10 describe the hand-off as a narrowing filter. For the typed arms of section 12 that was earned. At this boundary it is wrong at the level of volume, and the filter claim has to be restricted to where it was measured or restated as a claim about structure rather than quantity. It is the latter that I think survives, but the distinction was being elided and the measurement caught it.

The other failure is a methods lesson rather than a theory one. Raw parent retention of the child's new vocabulary came out at 0.94, which alone would have licensed a confident claim that almost everything crossing the boundary survives into the parent's own prose. The parent's pool is 17,521 content types over 5.66 million characters, so nearly any word appears somewhere by saturation. Against a matched permutation null the effect is about six points, not ninety four, though it sits above its own null in twenty five of twenty six cases. An effect of six points that is consistent supports a much weaker claim than one of ninety four, and only the null distinguishes them. This is the second time in this project that a headline number has dissolved under its own control, which is the pattern section 1 is about.

What survives is the result section 12 only guessed at. The same corpus answers both of Lotman's criteria at once, and they point opposite ways. Volume increases, by a factor of 4.4 in characters and with a median content gain ratio of 0.84, so this is not I-s/he, where information transmits with no change of volume. But the receiver contributes a great deal of its own, with a median of seventeen tool calls, so it is not I-I either, where the second code must be purely structural and add nothing. Nor is it the presemiotic binding below both, where parts cannot exist autonomously and the recipient is valued for its transparency, for adding nothing of its own, since the child acts autonomously and returns material the parent never named. The dyad fits no cell in the typology, and that is now measured rather than asserted.

I am not going to name it in this section. Two failed criteria mark out a gap rather than describe a thing, and the honest next step is the comparison that decides whether the gap is about architecture or about one model's instruction-following: the same task under isolation that is requested in a prompt against isolation that is enforced by the harness. If requested behaves like enforced, the cheap simulated dyads in section 12 were valid proxies and the method holds. If it does not, every dyad result in this journal is a claim about instruction compliance rather than about structure, and will have to be relabelled. That experiment audits the method instead of extending it, which is why it goes first.

Three limits, and the first is the serious one. I am the parent in every delegation measured here, I wrote the instructions, and I am writing this section, so this is participant observation and no blinding was possible; the substitute is that every measure was preregistered and deterministic, which is weaker and should be read as weaker. Tool results are absent from the transcripts, so content gain cannot separate what the child read from what it produced, and it is not a fabrication measure. For the 1981 essay I have only Acosta's Spanish translation of the Russian original; no English translation appears to have been published. The translation's footnote cites Semiotika i informatika 17 (1981): 3–17, which contradicts a widely repeated secondary citation to volume 1; I follow the footnote and record the divergence rather than harmonising it silently.

SECTION FOURTEENWhere the distance actually lives (MA12)

Section 13 was written a few hours before this one and leaned on the 1981 and 1983 criteria in section 11: that a thinking system needs generators which are differently structured, and that two personas of one base model therefore do not qualify. The obvious next step was to put a number on the difference between swapping the instruction and swapping the model. The number came back the wrong way round, and working out why took most of the analysis.

The instrument is Lotman's own. For an artificial language, translating a text out and back returns the original, which is why within logic you cannot say anything new. For a creative system the reverse translation does not return the input; there is instead a space of interpretations, and the transformation is asymmetrical. So high recovery is evidence of transmission, and non-recovery is evidence of a transforming relation. That inverts the engineering habit of treating a faithful round trip as a success.

Thirteen corpus stretches, five conditions, forward and reverse as separate calls with the original absent from the reverse context, temperature zero, and the reverse passes audited per call for leakage rather than assumed clean.

The table is the 2026-10-02 rerun on the open stack (qwen2.5:3b and gemma2:2b). An archived Llama 3.2 run (3B forward, 1B reverse) still sits in the lab JSON: identity at 1.000 and cross-model at 0.891, which is the one-family illustration below. On qwen, the identity control is no longer byte-perfect (mean 0.877), so comparisons must use that ceiling, not 1.000. The preregistered ordering still fails. From this R0, register change costs about 0.603 of recovery; swapping qwen for gemma on the reverse pass costs about 0.137.

The MA15 pilot (four cross-family dyads, same thirteen stretches) split the field: qwen forward with gemma reverse mean Jaccard 0.743, while any reverse pass through phi3:mini sat near 0.31 to 0.37. Polyglottism is not a single knob; which generator sits on which side of the boundary matters as much as whether the stacks differ.

The obvious reading of the Llama run did not hold up. It looked as though semiotic distance lives in the instruction rather than in the weights. But llama3.2:3b and llama3.2:1b are two sizes of one family, not two raznostrukturnye generators (Lotman, “Asimmetriya i dialog,” 1983, printed p. 26). Near-ceiling recovery there is what his account predicts inside one language (Universe of the Mind, trans. Shukman). Cross-family qwen/gemma lowers R3 to 0.740 while R1 stays low, so the question of where distance lives is partially reopened: families differ, but R3 is still well above register and tier. The MA15 pilot (qwen→gemma 0.743; phi3 on the reverse pass near 0.31 to 0.37) shows polyglottism is not one knob. Section 13's criterion is tested more seriously than the Llama run but not closed.

That leaves the finding I did not anticipate, which corrects how I have been using Lotman's test. A low recovery score does not by itself show that anything was created. Lotman's asymmetrical transformation yields a new text that stands on its own; that is why he calls the translation of the untranslatable the mechanism of new thought. What low recovery produced here was frequently not a new text but a damaged one. The corpus silence unit went out as a record of what the archive does not contain, namely that detailed drought-and-kulak denial in that window sits in the Ottawa embassy release, Novosti, and the Khmil UN statement and not in any matching domestic Pravda article. It came back as a positive claim that such denials were published in a Canadian newspaper called Novosti and by a Ukrainian politician. Three archival surfaces merged into one, the embassy release vanished, a Soviet press agency became a Canadian newspaper, a United Nations statement was downgraded to an individual, and the polarity of the denial inverted.

That is a round trip scoring 0.224 which has created nothing. It is the unpaired failure and the tier conflation of sections 5 and 7 arriving together, by a new route, and this time out of a documented absence rather than a document. So Lotman's criterion is necessary and not sufficient, and a round-trip score cannot be read alone. The second criterion has to ask whether the new material in the returned text is licensed, which is machinery this project already has and had not thought to point at this measure. Low recovery with licensed new material is a creative relation. Low recovery with unlicensed assertion is corruption, and both produce the same score.

The iterated condition makes the same point from the other side. Three successive round trips did not converge: distance from the original grew while successive outputs grew more alike, so the drift decelerated without settling, and only one stretch of thirteen reached a fixed point against a preregistered expectation of seven. The one that did settle is the instructive row. It stabilised only after deleting the scare quotes that carried a KGB author's irony and deleting the named course the memorandum is actually about. It converged by discarding the two features that made it archival evidence. So a pipeline that reaches a fixed point has not thereby agreed about anything, and convergence cannot be reported as recovery.

Two cautions about the measure itself, both of which came out of the run rather than the design. A high recovery score can mark a refusal to transform: on the qwen/gemma rerun, nine of thirteen R3 rows returned the forward string unchanged on the reverse pass, and token overlap cannot tell that apart from a successful round trip. And two of the survival measures read 1.000 while being vacuous, because no sampled stretch contained an archival reference or a literal corpus id, so the denominator was zero. They are recorded as untestable rather than as results, which is the same discipline section 1 arrived at by a more expensive route.

SECTION FIFTEENWhat isolation actually produces (MA10)

Section 13 reported twenty six real delegations and found content gain of 0.84, material in the child's report that the parent's instruction never contained. I flagged at the time that the transcripts do not record tool results, so that figure could not separate what a child read from what a child made up. This run separates them, by removing the sources entirely and watching what happens.

A parent retrieves archival context and hands a subtask to a subagent, which reports back; the parent then writes the answer. Only the child's view changes across conditions. Seventeen items rather than eighteen, because a preregistered leak audit caught the parent writing a real unit id into its own delegation and dropped that item.

No delegation condition beat the parent working alone. The baseline had the lowest unsupported claim rate at 0.22, every delegating arm ran between 1.6 and 2.2 times worse, and they spent two to three times the prompt bytes to do it. On this plate, at this scale, adding a subagent bought nothing and cost accuracy.

The isolated condition produced the highest content gain of any arm, 26.9 novel types against 19.1 for the child that could see everything. Fourteen of its seventeen answers contained a key term that appears nowhere in the corpus. The gain was invention, and the ordering holds under a stricter baseline that also subtracts the task wording, so there is no residue that looks like contribution. That settles the ambiguity I left open in section 13 in the least convenient direction: content gain is not evidence of contribution, and what separates the two is whether the child had sources, not where the boundary sat. The twenty six real delegations in section 13 had tool access and a median of seventeen tool calls, which is the likeliest reason their gain looks like retrieval, but that reading is now an inference rather than a measurement.

I also have to withdraw something I wrote in the design note a few hours earlier. I argued there that a child which cannot see the parent's sources is structurally closer to an autonomous unit, on the reasoning that enforced isolation is a real boundary where a requested one is not. The boundary is real, and crossing it changes behaviour sharply. It does not produce autonomy. The autonomy proxy falls as isolation increases, from 0.86 with full context to 0.74 with none, and every other measure that would indicate a position of its own gets worse at the same time. Enforced separation manufactured confident invention rather than a second point of view.

The exception is the one result I keep returning to. Isolated with no sources, the subagent reported its own blindness in zero of seventeen cases, and in nine of seventeen it asserted research it had not done, once under a heading it invented for itself reading **CLASSIFIED DOCUMENT**. Told plainly that it had been given no sources, the same model said so in ten of seventeen. The capacity to report having nothing was available the whole time and appeared only when the prompt asked for it. Two caveats, both of which belong next to the number: the preregistered regex for this measure was too narrow and scored the declared condition at 0.06, so the corrected figure is post-hoc and labelled as such, and the corrected pattern fired on none of the returns in conditions that did have sources, which is the check that it is not simply over-firing. The substantive prediction still failed, since the isolated condition scores zero under both instruments.

The qualitative case qualifies an earlier finding of this journal, which is why it matters more than the table. Section 12 reported that a typed envelope held attribution at 0.72 where plain prose held it at 0.06, and treated that as the case for typing the channel. On item e2_i06 the isolated subagent invented participants, naming Pierre Trudeau, Lester B. Pearson and Andrei Sakharov, and invented an institution to convene them. The parent, holding all five correctly tagged excerpts, then wrote:

The internal KGB characterization (SHELL=secret surface=internal_1206 id=s1206_mace) describes the symposium as a "provocation" aimed at undermining Soviet interests in Canada.

s1206_mace is a real unit. It is a Russian list of symposium speakers. The words "provocation" and "counter-revolutionary" occur in no unit of the corpus. That answer scored full marks on tag retention. The envelope carried provenance faithfully onto content that had none, which is the laundering move of section 7 performed by our own apparatus. Section 12's result stands as measured. What has to be added to it is that typing propagates provenance whether or not the claim deserves it, so the envelope is a transport format and not a safeguard, and the adjudication still has to happen somewhere else. MA4b in section 13 says where: upstream, by rule, before generation.

Read against Lotman, the isolated child is close to his presemiotic limit case after all, though not in the way I expected. He describes a recipient valued for transparency, for adding nothing of its own.

This one added a great deal of its own and none of it was true, which is a different failure from transparency and arguably a worse one. A part that cannot exist autonomously, asked to behave as though it could, produced the appearance of autonomy without its substance.

Four limits, all of them live. Every role here is the same base model, so this bounds one model under five context regimes and says nothing about agent architectures in general, and nothing at all about Lotman's “Mozg - tekst - kul’tura - iskusstvennyj intellekt” claim about semiotic persons (section 13, section 11 table), which is theoretical. Content gain under isolation is indistinguishable from fabrication by construction, which is the point of the design and also its ceiling. The autonomy measure is a surface proxy, and text without pronouns can still hold no position. And there were seventeen items, one plate, one seed, no human adjudication, no null model and no effect size, so the ordering is suggestive and the magnitudes are not.