Reproducible code, prompts, and data for the four-paper series "Foundations of the Convergent Semantic Architecture." CLCA recovers a shared semantic architecture for truth-related meaning by decompiling how sixteen unrelated languages compose it, and derives from that architecture a structural account of reasoning error (the Universal Fallacy Architecture) and of truth itself.
Papers: Paper I is available at https://doi.org/10.5281/zenodo.21709636 and Paper II at https://doi.org/10.5281/zenodo.21998398 (Papers III-IV forthcoming). This repository is the companion artifact: everything needed to inspect, run, and build on the findings.
Two papers, four companion documents, and a data repository is a lot to meet at once. You don't need it all — pick the door that matches what you came for:
- Just curious? Start with the Universal Fallacy Map. You already know what post hoc and ad hominem are. The Map takes a hundred-plus classical fallacies and shows them collapsing into five structural error types; no new vocabulary is needed to see the pattern.
- Philosopher or argumentation theorist? Paper II is the flagship: it derives the five error types from a constraint architecture and predicts the fallacy catalogue instead of curating it. Keep the Map open beside it; the Terminology Correspondence maps this vocabulary onto standard disciplinary terms.
- Linguist, or checking the evidence? Paper I is the foundation: the sixteen-language protocol and what emerged from it. This repository is its companion — prompts, code, and the full runs.
- Engineer? The oAOML ledger applies the same five types to eighteen canonical engineering failures (Therac-25, Ariane 5, Mars Climate Orbiter), analyzed at the level of the individual judgments that failed.
- Want the machinery itself? The AOML v2.2 spec is the full constraint architecture: axes, operators, validation methods, and the legality matrix.
The one-sentence version of the whole program: truth claims come in a small number of kinds, each with its own ways of being checked; modulating a claim (sourcing it, hedging it, likening it) never by itself establishes the claim (M[A] ↛ A); and most reasoning errors, human or machine, are one of five structural ways of breaking those rules.
- The protocol (
src/,prompts/,configs/) — a reproducible, LLM-assisted pipeline that elicits structured, language-specific data and synthesizes it cross-linguistically. The prompts contain no axis names and no theoretical vocabulary; the architecture is meant to emerge from the data, not be imposed on it. - The architecture (
aoml/,Aoml_v2.2_constraint_architecture.html) — AOML v2.2: six primary axes, four meta-axes, the operators and levels, and the structural rules. Machine-readable matrices plus a human-readable reference. - The data (
data*/) — the full outputs behind the papers (see below). - Reference (
documents/) — the Universal Fallacy Map (Second Edition: 102 entries mapped to five structural error types), the oAOML engineering failure analyses, and a terminology-correspondence table. The papers themselves are on Zenodo (see Citation below).
Two independent eight-language sets that share no languages — ten genetically unrelated families plus three isolates:
- Set A: Georgian, Yoruba, Finnish, Basque, Tamil, Indonesian, Turkish, Korean
- Set B: Chinese, Hebrew, Sanskrit, Greek, English, Tagalog, Swahili, Quechua
The protocol is also run two ways (one neutral, one domain-seeded) as a mutual control, so a pattern only counts as a finding if it surfaces in both sets and both protocol variants independently.
conda env create -f environment.yml
conda activate clca
export ANTHROPIC_API_KEY=... # Anthropic backend (default)
export OPENAI_API_KEY=... # only needed for the GPT-5.4 cross-model runSingle language:
python -m src.runners.run_language --language-name "Georgian" --language-code ka --script GeorgianA full eight-language set plus the cross-language G-phase synthesis (Windows .bat drivers shown; the .sh equivalents are alongside):
run_full_pipeline.bat # Anthropic, Set A -> data/
run_full_pipeline_B.bat # Anthropic, Set B -> data/
run_full_pipeline_gpt54.bat # GPT-5.4, Set A -> data_gpt54/G-phase synthesis only:
python -m src.runners.run_global_analysis \
--languages ka yo fi eu ta id tr ko \
--backend anthropic --model-name claude-sonnet-4-5 \
--output-code global_set_ACheck completeness, and diff a fresh run against the shipped one:
python -m src.runners.verify_languages configs/global_set_A_codes.toml
python -m src.runners.compare_runs --baseline data --rerun <your-output-dir>Because LLM informants are non-deterministic, exact strings will differ run to run; the claims are about structural patterns, and data_reproducibility/ is an independent second run included precisely so you can measure run-to-run stability.
Every directory is a complete run — ship-and-diff is the point.
| Directory | What it is |
|---|---|
data/ |
Canonical Anthropic (Claude Sonnet) run, Sets A & B — the primary evidence |
data_gpt54/ |
Parallel GPT-5.4 run — cross-model robustness check |
data_reproducibility/ |
Independent second Anthropic run — within-protocol run-to-run stability |
data_union/ |
16-language union pipeline (K-probe merge → consolidated F + G) behind AOML v2.2 |
data_saturation/ |
P-phase saturation-curve runs |
data_baseline_4_6/ |
Earlier baseline comparison run |
data_opus/ |
Higher-tier (Opus) cross-tier probe |
Each <lang>/ directory holds the per-phase P / I / F outputs; each global_set_*/ holds the G1–G9 cross-language synthesis, including G7b, the step that compares the elicited data against the AOML reference architecture.
- Code: Apache 2.0 (
LICENSE_APACHE_2.0) - Data / prompts / documents: CC-BY-4.0 (
LICENSE_CC_by_4.0)
You are free to use, modify, and build upon this work (including commercially) as long as you provide appropriate credit and keep provenance of derived datasets/code. These copyright licenses do not grant patent rights (see Patent below).
The knowledge here — the architecture, the legality matrix, the fallacy taxonomy, and the data — is open under the licenses above: read it, share it, teach it, build on it with attribution. Applying it in a computational or AI system (using the axes, operators, legality matrix, or AOML tuples to tag, validate, constrain, or enforce semantic structure in software or an AI model) is patent-pending and may require a separate patent license. The patent-pending enforcement systems are not contained in this repository. See PATENTS.md. (Informational, not legal advice.)
Seeley, B. A. (2026). Cross-Linguistic Compositional Analysis (CLCA): A Reproducible Protocol for Recovering Shared Conceptual Structure from Cross-Linguistic Evidence. Paper I, Foundations of the Convergent Semantic Architecture. https://doi.org/10.5281/zenodo.21709636
Seeley, B. A. (2026). The Universal Fallacy Architecture: Structural Error Prediction from AOML Constraint Geometry. Paper II, Foundations of the Convergent Semantic Architecture. https://doi.org/10.5281/zenodo.21998398
Companion: Seeley, B. A. (2026). The Universal Fallacy Map: Classical and Contemporary Fallacies Mapped to Five Structural Error Types (Second Edition). https://doi.org/10.5281/zenodo.21880635 (all versions: https://doi.org/10.5281/zenodo.21766167)
Companion: Seeley, B. A. (2026). Terminology Correspondence: Mapping CLCA/AOML/UFA Vocabulary to Standard Disciplinary Terms. https://doi.org/10.5281/zenodo.21781768
Companion: Seeley, B. A. (2026). AOML v2.2: Semantic Constraint Architecture. Claim Structure, Structural Rules, and Validation Habitats. https://doi.org/10.5281/zenodo.21792514
Companion: Seeley, B. A. (2026). oAOML: Engineering Failure Analyses. Canonical Incidents Mapped to the Five Structural Violation Types. https://doi.org/10.5281/zenodo.21960566
The series as a whole: Seeley, B. A. (2026). Foundations of the Convergent Semantic Architecture (Papers I-IV). Papers III-IV forthcoming.