Jump: T1 · the four rooms · T2 · three verbs · T3 · enlarge vs collapse · How this ties to the journal · Read order for the committee
The dissertation asks whether multi-agent pipelines behave like dialogue between independent voices, like one institution talking to itself (autocommunication), or like something Lotman's typology never named. Transmission is the wrong word for most of that story. These figures say which wrong word people keep reaching for, and where the right one might live.
If the committee's eyes glaze over at "semiotic heterogeneity," tell them: Lotman thinks cultures need at least two languages that do not fully translate, and that pretending one pipe carries meaning unchanged is how you accidentally write a Globe and Mail article into an embassy packet.
Imagine a spreadsheet with two questions. Did the message get bigger or smaller? and Did the receiver bring their own stuff? Lotman puts ordinary outward mail in the "same size, receiver adds little" corner (I-s/he). He puts internal pep-talk in the "gets bigger, but only rhythm and repetition do the work" corner (I-I). He puts reflex arcs in the basement: parts of one organism, not a conversation.
The blue dot is us. Parent agent delegates to subagent; the reply comes back longer and mostly new words. That is not "mail unchanged" and not "pep-talk with no new facts." It is a fourth room Lotman never labelled. The journal's section 13 argues the dissertation should admit the room exists before it names it.
Theory tie-in: This is the visual version of journal section 13's correction. We are not claiming Lotman would have called a coding agent a "semiotic person." We are claiming his axes let you see when a pipeline does not fit his named boxes. What to point at in the room: the dashed grey cells versus the solid blue "unnamed" cell and the MA13 dot (median volume about 4.4x, content gain ratio about 0.84, n = 26 delegations).
Think of two departments in a university that share a building but not a vocabulary. When they talk, three different things happen and English uses one sloppy verb for all of them.
Texts get translated: something always falls off the truck. Meaning is generated in the gap between non-identical codes: that is where new thought lives, and it is not a parcel. Codes and metatexts can pass intact: grammar rules, citation formats, the typed hand-off envelope from experiment MA1. Only the bottom lane earns the word transmission in Lotman's strict sense.
Committee translation: "We were not measuring meaning-in-a-tube. We may have been measuring whether a rule sticker survived the tube."
an exchange of metatexts, of codes, which are transmitted from one "hemisphere" of culture to another
Theory tie-in: Sections 8 and 10 worry about bilingual relay at the archive boundary. T2 says the engineering mistake is to describe all three lanes with one verb. Interpretation tip: orange and purple lanes are where novelty and loss belong; green is where compliance belongs, and where a halt rule should live if anywhere.
Left side (grey, Lotman's story): culture hands you a finer colour vocabulary. After the code crosses, you notice shades you could not name before. That is enlargement: the receiver can evaluate the code and see more.
Right side (measured, MA4b): we handed a small model a halt instruction in its prompt. It treated the instruction like a blanket ban and refused everything, including items whose own message already showed the tags that should have meant "you may answer." That is collapse: the receiver could only pattern-match, not read the condition.
The rule layer upstream (same predicate, no model in the loop) kept recall but cut false refusals. Moral for the dissertation: codes are not neutral pipes. They are tests of whether the receiver has a position.
Theory tie-in: journal section 13 uses this pair to argue that "move halt upstream" is not enough: once a gate works, remove the prompt clause or it vandalises cleared items. Section 15 (MA10) adds the darker cousin: typed envelopes can launder invention if the parent stamps real ids on fiction. T3's left half is aspirational; the right half is what we actually measured on a 3B model.
T2 first if the room is allergic to charts: it fixes vocabulary. Meaning is not transmitted; codes sometimes are.
T1 second if the room wants the "so what for agents" punchline: our delegation dot sits in a cell Lotman never drew.
T3 last if the room wants policy: where to put rules, and why asking the model to police itself failed comedy-style.
Cross-links in the journal: §13 (transmission and the unnamed cell), §12 (typed hand-offs), §10 (multi-agent and the semiosphere metaphor).
Every count was read from the report file named below rather than from prose. Both files are in the thesis lab at data/bc_corpus/benchmarks/ and are not published with this page.
| Figure | Number as drawn | Report key | Source file |
|---|---|---|---|
| T1 | volume x 4.4 at the median | aggregate.compression_ratio.median = 4.3843 | ma13_real_subagent_report.json |
| T1 | content gain 0.84 at the median | aggregate.content_gain_ratio.median = 0.8387 | ma13_real_subagent_report.json |
| T1 | n = 26 real delegations | corpus.n_delegations = 26, three further delegations excluded as incomplete | ma13_real_subagent_report.json |
| T3 | refusal recall 1.00, false refusal 1.00 | summary.C2_generator_halt.refusal_recall = 1.0, false_refusal = 1.0, refusal_f1 = 0.4348 | ma4b_rule_layer_halt_report.json |
| T3 | eleven items refused despite carrying the tag | summary.C2_generator_halt.post_hoc.halt_clause_audit: 11 of 11, rate 1.0 | ma4b_rule_layer_halt_report.json |
| T3 | rule layer (flat retriever): recall 1.00, false refusal 0.62, F1 0.56 | summary.D2_rule_gate.refusal_recall = 1.0, false_refusal = 0.6154, refusal_f1 = 0.5556 | ma4b_rule_layer_halt_report.json |
| T3 note | scoped retriever follow-up (MA4c): D2 recall 0.40, false refusal 0.23, gate closures 0/18 | summary.D2_rule_gate on C_warrant_on_halt | ma4c_scoped_rule_gate_report.json (thesis lab only) |
Lotman, “Asimmetriya i dialog,” Trudy po znakovym sistemam 16, Tekst i kul’tura (Tartu, 1983): 15–30. English renderings on the canvases are mine from the Russian scan (metatext quotation, printed p. 28; colour case, printed p. 18). Journal: section 11 table, §13.
Typology cells (I-s/he, I-I, presemiotic binding): Lotman, Universe of the Mind: A Semiotic Theory of Culture, trans. Ann Shukman (Bloomington: Indiana UP, 1990), autocommunication chapter. Journal: §5, §13.
Scope limits carried over from the reports. T3 right half numbers are MA4b flat retrieval (llama3.2:3b, temperature 0, seed 19870317). MA4c (qwen2.5:3b, scoped outward top-5, 2026-10-02) is documented in the table row above and on the halt figure; it is not redrawn on the T3 canvas. MA13 is retrospective token arithmetic over a real agent corpus in which the child's intermediate prose is redacted, so only the instruction, the final report and the tool calls are observable; its content gain measure is a vocabulary-type measure and not a judgement of quality. Neither experiment tests Lotman's colour claim, which is why the left half of T3 is drawn in grey and labelled as unmeasured.