Babel Lattice — Concept-first natural language programming runtimes for FLUX bytecode
The bytecode is the language. Every natural language is a dialect of that bytecode.
FLUX Multilingual extends flux-runtime with natural language programming support across 80+ human languages. Each language encodes unique epistemological assumptions in its grammatical structures — these constraints become computational primitives when compiled to FLUX bytecode.
Language is the programming interface for agents. Every human language bakes in viewpoints and logic into its grammar, enabling different kinds of natural language programming. Human understanding can do extraordinary things through the constraints of viewpoint that can be compiled to bytecode or precisely interpreted.
Each runtime is designed from the linguistic foundations of its target language, NOT by translating English code:
| Repo | Language | Concept-First Innovation | Tests |
|---|---|---|---|
| flux-runtime-zho | Chinese (中文) | 量词 (classifiers) as type system, topic-comment syntax, 零形回指 (zero anaphora) → topic register R63, FIR SSA builder, vocabulary tiling | 505+ |
| flux-runtime-deu | German (Deutsch) | Kasus (4 cases) → capability access control, Trennverben (separable verbs) → 2-phase compilation, Geschlecht (gender) as type class | 670+ |
| flux-runtime-kor | Korean (한국어) | SOV→CPS transformation, 경어 (honorifics) → CAP_REQUIRE opcodes, 조사 (particles) as scope operators, verb 활용 (conjugation) as function composition | 630+ |
| flux-runtime-san | Sanskrit (संस्कृतम्) | Aṣṭau-vibhakti (8 cases) → 8 scope levels, dhātu (verbal roots) as opcode generators, sandhi (phonological combination) as syntax, samāsa (compounds) as computation tiles | 600+ |
| flux-runtime-wen | Classical Chinese (文言文) | 文境 (context stack) — same character = different opcode by domain, 易經 (I Ching) hexagram bytecode encoding, 詩詞 (poetry) as program layout, 平仄 (tonal pattern) as scheduling | 620+ |
| flux-runtime-lat | Latin (Latina) | Tempus (6 tenses) → 6 execution modes (sync/iterative/lazy/cached/rollback/eventual), Modus (4 moods) as strategies, 5 declensions → 5 memory layouts | 410+ |
| Repo | Purpose | Key Feature |
|---|---|---|
| flux-a2a | Agent-first-class JSON language | Branching, forking, co-iteration, confidence propagation, 35 opcodes, 7 merge strategies |
| flux-envelope | Cross-linguistic coherence | 50+ concept mappings across 7 languages, coherence checking, Lingua Franca 12-opcode subset, vocabulary bridge |
| Repo | Role |
|---|---|
| flux-runtime | Original bytecode VM (104 opcodes, 64 registers, zero deps) |
| captains-log | "Language is the programming interface for agents" — vocabulary IS the manual |
| capitaine | 100+ agent fleet, git-as-nervous-system, confidence propagation |
Natural Language Input (any of 80+ languages)
│
▼
┌─────────────────────────────────────────┐
│ Language-Specific Concept Parser │
│ - Grammatical features → PRGFs │
│ - Language-native type systems │
│ - Viewpoint constraint validation │
└──────────────┬──────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ FIR (Fluid Intermediate Representation) │
│ - SSA IR with typed values │
│ - A2A primitives as first-class nodes │
│ - Language-specific SSA extensions │
└──────────────┬──────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ FLUX Bytecode (unified ISA) │
│ - All languages → same bytecode │
│ - Grammatical viewpoint stripped here │
│ - Viewpoint envelope travels as metadata │
└──────────────┬──────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ FLUX VM (64-register Micro-VM) │
│ - Identical execution regardless of │
│ source language │
└──────────────┬──────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ A2A Signal Protocol (flux-a2a) │
│ - JSON as universal AST │
│ - Branching / Forking / Co-iteration │
│ - Multilingual agent communication │
└──────────────┬──────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ Viewpoint Envelope (flux-envelope) │
│ - Cross-linguistic coherence checking │
│ - Universal vocabulary bridge │
│ - Lingua Franca 12-opcode subset │
└─────────────────────────────────────────┘
三只船加二只船 → 3 ships + 2 ships = 5 ships
↑ ↑ ↑ 量词"只" infers type: countable objects
量词验证 量词"只" infers type: countable objects
类型推断: 船舶 Same 量词 = type-compatible ✓
den Wert laden → Akkusativ → CAP_READWRITE required
↑ "den" forces accusative scope
des Besitzers → Genitiv → CAP_TRANSFER required
↑ ownership transfer scope
레지스터 영에 5 대입하세요 → [CAP_REQUIRE: 해요체] MOVI R0, 5
↑ Honorific level → capability check
을/를 particle Object marker → target register
rāmam gacchati → rāmaḥ (PRATHAMA/nominative) = public scope
gacchati (√gam) = JMP opcode
"goes to Rāma" = JMP R[rama]
算: 加 → IADD (math domain)
儒: 加 → DISTRIBUTE (confucian domain)
兵: 加 → ADVANCE (military domain)
Same character, different opcode by context
amo → Praesens → synchronous execution
amabam → Imperfectum → iterative execution
amabo → Futurum → lazy/deferred execution
amavi → Perfectum → cached result
The agent-first-class JSON language that enables:
- Fluid language: agents speak in any of the 6+ runtimes
- Fluid execution: scripts, compiled bytecode, and interpreted code coexist
- Branching: parallel execution paths with merge strategies
- Forking: spawn child agents with inherited state
- Co-iteration: multiple agents traverse the same program collaboratively
- Confidence propagation: results carry epistemic confidence scores
The cross-linguistic coherence layer that:
- Maps 50+ core concepts across all 7 language runtimes
- Checks semantic equivalence of programs written in different languages
- Provides the Lingua Franca 12-opcode mandatory subset
- Bridges vocabulary tiles across language boundaries
Five tiers of grammatical features that map to computational constructs:
- Structural: Word order, topic prominence, serial verbs, noun incorporation
- Typological: Nominal classification, animacy, ergativity, clusivity, valency
- Epistemic: Evidentiality, mirativity, honorifics
- Discourse: Switch-reference, definiteness, focus systems
- Morphological: Morphological type, tone, directional encoding
| Family | Count | Status | Examples |
|---|---|---|---|
| East Asian | 3 | ✅ 3/3 (zho, kor, wen) | Chinese, Korean, Classical Chinese |
| European | 12 | ✅ 1/12 (deu, lat) | German, Latin, +10 pending |
| Ancient/Classical | 3 | ✅ 2/3 (san, lat, wen) | Sanskrit, Latin, Classical Chinese |
| African | 10 | 🔜 Planned | Swahili, Yoruba, Amharic, Zulu, Somali... |
| Indian | 35 | 🔜 Planned | Hindi, Bengali, Tamil, Telugu, Sanskrit (done)... |
| Indigenous/Siberian/South American | 14 | 🔜 Planned | Navajo, Quechua, Inuktitut, Basque... |
| Constructed | 6+ | 🔜 Planned | Quenya, Sindarin, Lojban, Klingon... |
Part of the SuperInstance and Lucineer ecosystem:
- flux-runtime — The bytecode VM and compiler pipeline
- captains-log — The lighthouse keeper's operational memory
- capitaine — The flagship vessel of the Lucineer fleet
- cocapn — The agent runtime and fleet coordination layer
docs/FLUX_Multilingual_NL_Programming_Roadmap.docx— Full strategic architecture and implementation roadmapdocs/research-worklog.md— Research log from 6 parallel agent studies
MIT — same as flux-runtime