Semiotic coordinate plain

One sentence

Semiosis is not a point in one embedding space; it is typed relay across incompatible document surfaces, and machine systems that merge registers reproduce embassy-shaped readings.

What we are not claiming

• Words → one universal vector plane → “meaning.”
• Cosine neighborhood = understanding.
• The probe replaces pair-read.
• Montreal 1983 legislates a law of culture.
• Explosion is proven (L3 is gated).
• Generative models should write “the public version” of a memo.

What we are claiming

1. Register is part of the sign-event. The same famine language does different work on spr. 1206 and on an embassy release five weeks later.
2. Operators are measurable. Scare quotes, gathering vs symposium, drought-and-kulak denial form an operator × surface tensor (L2). Gaps in the tensor are as real as peaks (48 internal scare-quote hits, 0 outward in the pilot matrix).
3. Relay is partial. Some thematic fields cross (famine, Canada, anniversary); some operators do not (symposium has no outward Montreal sentence).
4. Silence is a coordinate. Domestic Pravda/Izvestiia not printing the drought frame in the processed 1983 window is evidence about which channel carries which job, not “missing data.”
5. Frozen embeddings already encode the failure mode. A secret→outward linear map fits in-sample pairs but does not align secret text with Weekly/CIUS rivals; сборище lands near drought; scare-quoted symposium near Khmil’s “lie” sentence.

Three models (L1 / L2 / L3)

Probe in one paragraph

Frozen paraphrase-multilingual-MiniLM-L12-v2; 25 units, 9 secret→outward pairs; orthogonal Procrustes + ridge + Hewitt-Liang shuffled controls (scripts/run_semiosphere_feature_space.py). In-sample orthogonal beats identity and shuffled pairing; LOTO collapses (nine pairs = dictionary, not training set). Key negative result: outward-trained W does not send secret stretches toward rival/diaspora targets. Lab log: EXPERIMENT_LOG_semiosphere_feature_space.md.

Stakes (if extrapolated honestly)

Semiotics: Untranslatability becomes empirical (zero column), not metaphor. Dual models without a neutral sphere become geometry, not ideology talk.

Machine cognition: Theme-level similarity ≠ register-faithful reading. Single-manifold AI is structurally prone to collapse secret and public-the same mistake as reading the embassy release alone.

dated pair or documented miss  →  pair-read  →  operator/tensor/L1 claims
                                      ↑
                         probe / counts only as checks

0 · Warrant chain (read first)

This document extrapolates from dated pair-reads and pilot instruments on HDA SBU fond 16 (Operation Vozmezdie, 1976-1988), calibrated on Montreal 1983. Nothing here replaces the five-step checklist or hand verification of newspaper claims.

Primary sources + pair-read  →  historical claims (dissertation)
         ↓ checks only
Counts, L1-L3 models, probe  →  falsify bad averaging / false survey

• Pair-read, or documented miss, is the warrant.
• The semiosphere probe does not generate memo language and does not redo Khanenko-Friesen fond-internal CDA.
• Explosion (L3) is gated, not claimed from pilot counts alone.
• Do not promote probe numbers into SSHRC prose or comprehensive exam lists.

1 · Core thesis

Semiosis is not a point in one vector space. It is typed relay across document surfaces that obey different grammars on the same calendar week.

When institutional language crosses from secret report to embassy release to diaspora press, operators change (quotation marks, meeting names, famine frames) even when themes persist (Canada, anniversary, famine). Generic embeddings mostly recover theme-level gravity; they fail register-faithful relay unless structure is imposed.

One-sentence version: The gap in an operator × surface tensor is where ideology lives; machine cognition that averages registers fills that gap with embassy-shaped summaries.

2 · Why the usual “vectored plain” fails

2.1 Default ML picture

• One manifold (e.g. cosine space of embeddings).
• Neighborhood ≈ similarity ≈ meaning.
• More data smooths noise; multilingual alignment shares a language-neutral core.

2.2 What the archive falsifies

Montreal 1983 shows non-mergeable coordinates on the same event:

These are not dialect variants. They are incompatible reader contracts on the same commemorative season.

2.3 Lotman’s correction (operationalized)

From Lotman and Uspenskij (dual models; no neutral sphere) and Lotman (bilingual filters at the boundary):

• The boundary is a sum of filters, not a line between “Soviet” and “Western.”
• What cannot pass is structurally real (untranslatable), not failed translation.
• Silence (domestic newspapers not carrying the drought frame in the processed 1983 audit) is a channel assignment, not absence of evidence.

L1-L3 attempt to draw this so it can be checked against pair-reads, not sit as metaphor in chapter one.

3 · What gets mapped (object type)

3.1 Not a token

A sign-event is minimally:

• a stretch of text (sentence to memo section);
• on a surface (secret memo, autocommunication report, embassy/Novosti, UN, domestic paper, diaspora press, scholarly rival);
• on a date (or narrow window);
• with active operators (scare quote, gathering recode, drought frame, slander frame, etc.);
• at a proof tier (primary, tier-B processed, documented miss, not claimed).

Vectors attach to sign-events, not to lemmas. “Famine” is not one point; “famine”-as-scare-quoted-lie on spr. 1206 and “famine”-as-drought on embassy are different events.

3.2 Register → shell map (L1)

Novosti = embassy English in this corpus (verbatim outward packet). Treating Novosti as independent evidence double-counts the same release.

Pravda / Izvestiia / Rad ykr: separate clock; absence of drought frame in processed 1983 search is a finding (outward-only denial routing), not a gap to ignore.

4 · Three geometries (L1 / L2 / L3)

4.1 L1 · Semiosphere = topology of relay

Question: Where does each text sit, and which arcs are licensed?

• Nodes: surfaces (13 in empirical pilot).
• Edges: relay arcs with proof tier (proven thematic relay; tier-B processed; promising pending clip; not claimed).
• Forbidden / missing arcs are first-class (no Pravda arc for drought frame; embassy packet does not name Montreal).

Semiotic stake: Location in culture = relay position, not frequency.

ML stake: Knowledge graphs need edge types and blocked edges, not only similarity cliques.

Figures: L1 empirical · L1-demo (target state - fabricated)

4.2 L2 · Operator tensor = the honest “plain”

Question: Which operators fire on which surfaces?

Example operators:

• scare_quote_artificial_famine
• symposium_recode (→ gathering / сборище)
• drought_kulak_denial
• nationalist_slander_frame (Khmil UN)

Pilot skew (internal vs outward): 48 scare-quote hits on internal memos, 0 on embassy/Novosti/UN in the same window - hard non-equivalence, not sampling noise.

Figure: L2 operator tensor

4.3 L3 · Explosion = phase change (candidate)

Question: After Montreal 1983, does internal phrase co-occurrence rewire while 1955 training-manual operators persist?

Pilot visualization: 3→7 nodes, 3→19 edges before/after 1983; red edges = new co-occurrences.

Gate: Claim explosion only if (1) post-1983 combinations are new relative to baseline; (2) 1955 operator vocabulary still describes the office habit; (3) rewrite tracks calendar time, not ingestion order.

Figure: L3 explosion (candidate only)

5 · Semiosphere feature-space probe

5.1 Design

Literature chain: Schönemann 1966; Mikolov et al. 2013; Xing et al. 2015; Artetxe et al. 2016; Alain & Bengio 2017; Hewitt & Manning 2019; Hewitt & Liang 2019. (On comps supplements · Full Chicago entries · Complete reading lists.)

5.2 Results (2026-09-10 run)

5.3 Epistemic role of the probe

The probe is a microscope, not an oracle:

• Confirms: Thematic relay is linearly available in a generic encoder; register merge is not justified by in-sample fit alone.
• Does not confirm: LLM “understanding,” generative reconstruction of memos, or global Lotman law.
• Use in prose: Falsify false survey and single-manifold summaries; do not lead SSHRC or dissertation lede.

6 · Proposed coordinate system (extrapolation)

Generic embeddings supply unchosen axes. Lotman’s question requires named fibers:

6.1 Partial maps and untranslatables

Translation between shells is a partial function (not every input has an output):

Examples from Montreal:

• scare_quote_artificial_famine → ⊥ on outward shell (zero column in L2).
• symposium (rival sense) → ⊥ on embassy shell (documented miss on meeting name).
• Thematic famine field → drought frame (transformed relay, not copy).

Documented miss (embassy release does not name Montreal) is obstruction: the morphism does not exist for that operator on that date.

6.2 Sheaf intuition (informal)

• Local patches: each shell has its own speakable operators.
• Gluing maps: only where pair-read proves relay; tier labels say how strong the glue is.
• Obstruction: ⊥ and silence columns prevent global merge.

Formal write-up deferred to Phase 1 in the exploration roadmap.

7 · Implications for semiotics

• Structure and history: 1955 manual = grammar habit; 1983 = rewrite under public defeat.
• Reference and social action: Operators are action types with measurable surface skew (Ahearn / Fairclough).
• Untranslatability: Lotman’s filter = zero tensor column, falsifiable on new memos.
• Dual models: Weekly vs embassy - averaging them is a reading mistake made literal in embedding space.
• Silence: Trouillot meets geometry - domestic silence constrains what Novosti/embassy are doing.
• Graphical Lotman: L1-L3 + proof tiers = first operationalization tied to archival epistemology.

8 · Implications for machine cognition

• Single-manifold failure: merged-register models embassy-shape secret material.
• Generation is wrong default: object is non-copying relay with forbidden operators.
• Multilingual alignment: aligns lemma gravity; loses scare quotes and gathering recode.
• Monitoring: if L3 gate passes, phrase-network rewrite = phase change, not embedding drift.
• Ethics of archive AI: propagating internal-only operators into outward summaries re-enacts core self-description.

9 · Falsifiers (keep visible)

10 · Relationship to dissertation chapters

11 · Publication tracks (post-gate)

1. History - embassy, Holodomor anniversary, Canadian reception (SSHRC core).
2. Semiotics / Slavic - operational Lotman, untranslatables as tensor gaps.
3. ML / DH - probe as falsification instrument; pair-read benchmark.
4. Archival science - companion checklist UX; proof tiers on edges.

Rule: Pair-read remains the warrant. Experiments falsify bad averaging; they do not replace composition and uptake.

North star

Build a checkable semiotic coordinate system where:

• axes are chosen (shell, operator, date, proof tier), not inherited from a generic encoder;
• translation is a partial map with explicit untranslatables (⊥);
• machine outputs are register-typed and cannot merge secret with public without a logged merge penalty.

Phase 0 Lock what we already have (1-2 sessions)

Exit criterion: “Proven / tier B / miss / not claimed” synced with docs/EXPERIMENT_GAPS.md.

Phase 1 Make the plain explicit (theory + data model)

1a · Sign-event schema

{
  "id": "1206_1983-03-24_gathering",
  "surface": "secret_memo",
  "shell": 1,
  "date": "1983-03-24",
  "operators": ["scare_quote_artificial_famine", "symposium_recode"],
  "stretch_text_ref": "...",
  "proof_tier": "primary",
  "pair_ids": ["embassy_1983-04-28", "weekly_1983-04-10"]
}

Action: Draft data/bc_corpus/semiosphere/sign_event_schema.json; migrate 25 probe units from pairs.json.

1b · Partial map registry

For each licensed relay: source/target shell, operator map (preserved | transformed | forbidden), evidence tier, ⊥ cases.

Action: Extend register_relay_experiment.json or new partial_maps.json.

1c · One-page formal sketch

Objects = sign-events; morphisms = proof-typed relays; obstruction = miss. Deliver after 1a-1b.

Phase 2 Probe v2 (falsification, not scale)

Do not: Fine-tune the transformer on the fond. Do not present LOTO as “model understands KGB.”

Phase 3 Fill L1 arcs (empirical geometry)

Phase 4 L3 explosion gate

Phase 5 Pair-read benchmark (publishable artifact)

Task types:

1. Operator retention - pick outward candidate that preserves/forbids scare-quote operator.
2. Miss detection - does embassy packet name Montreal?
3. Relay direction - classify surface type.
4. False merge detection - flag register collapse in a summary.

Deliverable: data/bc_corpus/benchmarks/register_relay_v1.json + scorer script.

Phase 6 Writing tracks

Suggested session order

1. Phase 0 re-run + confirm dated_units.
2. Phase 1a sign-event schema (2-3 hours coding).
3. Phase 3 acquisition (Globe, East View) - parallel, human-bound.
4. Phase 2.3 operator linear probe when schema exists.
5. Phase 4 L3 gate after more memos.
6. Phase 5 benchmark when 20+ scored pairs.

Open questions