ARCHON — Multilingual Engine Architecture

Architecture

How language flows through the engine

The user submits a form with a language preference. The engine classifies the initiative internally in English (for reliable JSON parsing), then generates all governance outputs in the chosen language.

User Input

Intake Form

Organisation submits AI initiative description + selects output language (e.g. "Arabic")

→

Request Payload

"output_language": "Arabic"

Passed in POST body to /generate. Stripped alongside _api_key before storage.

↓

app.py — _run()

Engine Initialisation

GovernanceEngine(output_language="Arabic")

→

engine.py — __init__

self.language = "Arabic"

Language stored on the engine instance. Used by all prose modules via _lang_directive().

↓

Module 0 — Always English

classify(fd) → JSON

Risk tier, EU AI Act class, applicable standards, governance concerns. English only — drives downstream prompts. No language directive injected.

→

Internal Result

cl = { "risk_tier": "HIGH", … }

Structured JSON — reliably parseable because it is always in English regardless of output_language.

↓

Modules 1–5 — Target Language

_lang_directive() appended

Each prose module appends the language directive to its prompt before calling the API. Framework, Risk Register, Compliance Matrix, Blueprint, ADRs & Roadmap.

→

Governance Pack Output

All prose in Arabic

Standard abbreviations (ISO, GDPR, NIST) remain in English. All narrative text, summaries, rationales, and recommendations are in Arabic (formal MSA).

Pipeline

What each module does with language

Six modules run sequentially. The first produces internal JSON. The remaining five produce user-facing prose — these are the ones that output in the selected language.

🔍

Module 0 — Initiative Classifier

Produces structured JSON: risk tier, EU AI Act classification, applicable standards, key governance concerns. This JSON feeds all other modules — it must be parseable, so it stays in English.

Internal JSON · English always

📋

Module 1 — Governance Framework

Board-ready framework document. Executive summary, principles, RACI matrix, oversight board structure, lifecycle gates, KPIs. Full prose — output in selected language.

Translated ✓

⚠️

Module 2 — Risk Register

16 initiative-specific risks with mitigation plans, likelihood/impact ratings, deep-dive cards for critical risks. All risk narratives output in selected language.

Translated ✓

✅

Module 3 — Compliance Matrix

ISO 42001, EU AI Act, NIST AI RMF, GDPR obligations mapped to controls. Gap register and 12-month remediation calendar. Standard article numbers stay in English.

Translated ✓

🏗️

Module 4 — Architecture Blueprint

ASCII architecture diagram, component inventory, data architecture, security controls, NFRs. Diagram stays in ASCII/English for clarity; all narrative sections in selected language.

Translated ✓

🗺️

Module 5 — ADRs & Roadmap

5 Architectural Decision Records and 3-phase implementation roadmap. All rationale, context, consequences, and milestone descriptions output in selected language.

Translated ✓

Implementation

The language directive

A single method — _lang_directive() — generates the instruction appended to each prose prompt. It is intelligent: returns nothing for English (zero token overhead), adds Arabic-specific MSA guidance for Arabic, and applies a generic directive for all other languages.

engine.py

def _lang_directive(self) -> str:
    # English → return nothing. Zero token overhead.
    if self.language.lower() == "english":
        return ""

    # Arabic → add MSA + RTL instruction
    arabic_note = (
        " Use formal Modern Standard Arabic (الفصحى — MSA)."
        " Ensure all narrative text is right-to-left."
    ) if self.language.lower() in ("arabic", "العربية") else ""

    return (
        f"\n\n---\n"
        f"**LANGUAGE DIRECTIVE — CRITICAL:** Generate ALL prose, headings, "
        f"narrative text, executive summaries, rationales, and explanatory "
        f"content in **{self.language}**.{arabic_note}\n"
        f"You MAY keep in English: standard abbreviations (ISO, EU, GDPR, NIST), "
        f"proper standard names, column headers, technical acronyms.\n"
        f"Do NOT mix languages within a paragraph or section.\n"
        f"---"
    )

💡

Why append to the prompt rather than modify the system prompt? The system prompt establishes ARCHON's identity and expertise for every call. The language directive needs to be a late-binding instruction that overrides any implicit English preference — placing it at the end of the user-turn prompt, just before the model generates, gives it the highest priority. It also means English users pay zero tokens for the directive.

Design Decision

Why classification always stays in English

✅

Classification outputs English JSON — always

The classifier returns a structured dict with keys like risk_tier, eu_ai_act_classification, applicable_standards. These are consumed by all five prose modules to construct their prompts. If these values were in Arabic, downstream f-strings injecting them into English-language prompt templates would break or produce incoherent results. Keeping classification in English is the stability anchor for the whole pipeline.

⚠️

The alternative was risky

If we translated classification output too — governance summary, rationale text — we could get partial Arabic content injected mid-English prompt: "Initiative Classifier says: يواجه هذا النظام مخاطر عالية" appearing inside a prompt template designed to receive English. This would confuse the model and likely degrade output quality. The current split (English JSON structure, translated prose) is the right boundary.

Language Support

Languages the engine supports today

Any language the model can write in will work — just pass the language name. These are the tier-1 and tier-2 targets based on ARCHON's market strategy.

Tier 1 — Strategic priority

Tier 2 — Major markets

Tier 3 — Available on request

🇸🇦

Arabic

العربية — MSA

GCC — Biggest opportunity

🇫🇷

French

Français

EU + Africa + Canada

🇪🇸

Spanish

Español

LATAM + EU

🇩🇪

German

Deutsch

EU — strong data regulation

🇧🇷

Portuguese

Português

Brazil + Portugal

🇨🇳

Mandarin

中文（普通话）

APAC — government scale

🇯🇵

Japanese

日本語

APAC

🇰🇷

Korean

한국어

APAC

🇳🇱

Dutch

Nederlands

EU + Benelux

🇮🇹

Italian

Italiano

EU

🇸🇬

Malay

Bahasa Melayu

SEA — Singapore, Malaysia

🇮🇳

Hindi

हिन्दी

India — large public sector

Market Strategy

What multilingual unlocks by region

Region	Language(s)	Strategic Rationale	Key Standards to add	Tier
GCC Saudi, UAE, Qatar, Kuwait	Arabic (MSA)	Massive AI budgets, acute governance gaps, strong national AI strategies (SDAIA, UAE AI Office). Arabic support is a prerequisite for government procurement.	Saudi SDAIA Framework, UAE AI Regulation, Qatar AI Strategy	1
EU France, Germany, Spain, Italy	French, German, Spanish	EU AI Act creates compliance urgency across all member states. Governments must demonstrate compliance in their own language for tender documentation.	EU AI Act (all annexes), ENISA Guidelines, national AI strategies	1
LATAM Brazil, Mexico, Colombia	Spanish, Portuguese	Growing AI regulation wave following EU model. Brazil's LGPD is the GDPR equivalent. First-mover in Spanish/Portuguese governance tools is a wide-open space.	Brazil LGPD, Mexico AI Ethics Principles, Colombia AI strategy	2
APAC Singapore, Japan, South Korea	English, Japanese, Korean	Singapore (English-first) is already accessible. Japan and Korea have dedicated national AI governance frameworks and large government budgets.	Singapore IMDA AI Framework, Japan AI Guidelines, Korea AI Act	2
Africa Nigeria, Kenya, Senegal, Morocco	French, Arabic, English	Francophone Africa is underserved, rapidly digitising public sectors, and follows French regulatory patterns. Arabic covers North Africa. Long-term high-growth market.	AU AI Policy Framework, SADC guidelines, national digital strategies	3

Think About This

Open questions for the product strategy

1

Should regulatory standards be localised too?

Right now the compliance matrix maps to global standards (ISO 42001, NIST, GDPR). For Arabic markets, should SDAIA and UAE AI Regulation be standard inclusions alongside those? This is an engine prompt change, not a language change.

2

Arabic needs RTL in the UI

The engine generates Arabic text correctly. But the results page renders it inside left-to-right HTML tabs. A dir="rtl" toggle on the results page body — triggered when the pack's output_language is Arabic — would make it fully presentable.

3

Quality validation per language

Claude's output quality varies by language. Arabic and Chinese governance terminology is strong. Less common languages may produce lower-fidelity technical content. The right V1 safety measure: a human review disclaimer calibrated by language.

4

Language selector in the form

The engine is ready. The form isn't wired yet — the intake form doesn't have a language dropdown. This is a 20-minute UI change. Just one dropdown on Step 1 or 4, posting output_language in the JSON body.

5

Language as a pricing lever?

English as the free tier. Arabic, French, Spanish as Pro tier features. This is common in SaaS localisation — the GCC government market will pay a premium for native-language governance packs that they don't have to translate themselves.

6

Translation vs. native generation

Current approach: generate natively in the target language. Alternative: generate in English, then translate. Native generation is better for context-specific technical content. Translation is faster and cheaper. The current approach is the right one for governance documents where precision matters.

Testing

How to test multilingual output

🧪

Direct API test — no form needed. Until the form has a language selector, you can test multilingual output by POSTing JSON directly to /generate with the output_language field included.

cURL / test

# Start the server
python app.py

# POST with Arabic output language
curl -X POST http://localhost:5000/generate \
  -H "Content-Type: application/json" \
  -d '{
    "org_name": "Ministry of Digital Affairs",
    "initiative_name": "AI Fraud Detection System",
    "description": "ML system that detects fraudulent benefit claims...",
    "ai_types": ["Machine Learning", "Predictive Analytics"],
    "decision_mode": "human-in-the-loop",
    "public_facing": "yes",
    "personal_data": "yes",
    "data_description": "Citizen financial and identity records",
    "affected_populations": ["low-income households"],
    "people_affected": "500,000+",
    "geo_scope": "National",
    "sector": "Social Services",
    "key_concerns": "Bias, fairness, transparency",
    "output_language": "Arabic"
  }'

# Returns {"job_id": "..."} — poll /status/{job_id}
# Results page will show governance pack in Arabic

ARCHON is nowglobally multilingual

ARCHON is now
globally multilingual