From 0c2c2e463d491f206a67f4eed00ac4092ea752ba Mon Sep 17 00:00:00 2001 From: TheFlow Date: Wed, 11 Feb 2026 08:07:26 +1300 Subject: [PATCH] feat: Audit & update all audience pages with current research and metrics - Add i18n data-i18n attributes to navbar.js (~30 elements) and expand all 4 common.json files with navbar translation keys (EN/DE/FR/MI) - Replace broken HuggingFace Space links with local audit analytics dashboard across 4 HTML files; update counts to 171,800+ - Add Steering Vectors, Taonga Governance, and Home AI sections to researcher.html, leader.html, and implementer.html - Update stale metrics: development period (11+ months), sessions (1,000+), instructions (68), model references (Opus 4.6) - Update roadmap Multi-LLM status to reflect Home AI as first non-Claude deployment; add Sovereign Training Pipeline and Taonga Registry items - Update all EN/DE/FR translation files with new section keys Co-Authored-By: Claude Opus 4.6 --- public/architecture.html | 10 +- public/implementer.html | 294 +++++++++++++++++++++++++++- public/index.html | 2 +- public/js/components/navbar.js | 96 ++++----- public/leader.html | 108 +++++++++- public/locales/de/architecture.json | 4 +- public/locales/de/common.json | 31 +++ public/locales/de/homepage.json | 2 +- public/locales/de/implementer.json | 81 +++++++- public/locales/de/leader.json | 34 +++- public/locales/de/researcher.json | 83 ++++++-- public/locales/en/architecture.json | 4 +- public/locales/en/common.json | 31 +++ public/locales/en/homepage.json | 2 +- public/locales/en/implementer.json | 81 +++++++- public/locales/en/leader.json | 34 +++- public/locales/en/researcher.json | 87 +++++--- public/locales/fr/architecture.json | 4 +- public/locales/fr/common.json | 31 +++ public/locales/fr/homepage.json | 2 +- public/locales/fr/implementer.json | 81 +++++++- public/locales/fr/leader.json | 34 +++- public/locales/fr/researcher.json | 91 ++++++--- public/locales/mi/common.json | 31 +++ public/locales/mi/homepage.json | 2 +- public/researcher.html | 210 +++++++++++++++++--- public/timeline.html | 2 +- 27 files changed, 1282 insertions(+), 190 deletions(-) diff --git a/public/architecture.html b/public/architecture.html index 30ec3510..629375cc 100644 --- a/public/architecture.html +++ b/public/architecture.html @@ -580,17 +580,15 @@ See the Framework in Action

- Explore 3,942 real governance decisions from production deployment. + Explore 171,800+ real governance decisions from production deployment. Filter by service, pressure level, and coordination patterns to understand how Deep Interlock operates in practice.

- - Launch Interactive Explorer on Hugging Face + data-i18n="interactive_exploration.audit_cta"> + Launch Interactive Audit Explorer Services API Reference Integration Patterns + Home AI + Steering Vectors + Taonga Registry ⚡ Agent Lightning Roadmap @@ -1064,6 +1067,250 @@ const govResponse = await fetch(
+ +
+
+
+ 🏠 +
+

Home AI: Two-Model Sovereign Architecture

+

+ Production deployment of Tractatus governance on locally-trained open-source models, demonstrating framework portability beyond Claude. +

+
+
+ + +
+

Two-Model Routing Architecture

+

+ Home AI uses a dual-model design where queries are routed based on complexity and governance requirements. Both models run locally with full Tractatus governance in the inference pipeline. +

+ +
+ +
+

Fast Model: Llama 3.2 3B

+
    +
  • + ⚡ + Purpose: Common queries with pre-filtered governance +
    • +
  • + ⚡ + Fine-tuning: QLoRA on domain-specific data +
    • +
  • + ⚡ + Governance: Lightweight boundary check before response +
    • +
+
+ + +
+

Deep Model: Llama 3.1 8B

+
    +
  • + 🔬 + Purpose: Complex reasoning with full governance pipeline +
    • +
  • + 🔬 + Fine-tuning: QLoRA with extended context governance +
    • +
  • + 🔬 + Governance: Full 6-service pipeline (BoundaryEnforcer through PDO) +
    • +
+
+
+ + +
+

Model Routing Logic

+ 
// Simplified routing decision
+function routeQuery(query, governanceContext) {
+  const complexity = assessComplexity(query);
+  const govRequirement = governanceContext.sensitivityLevel;
+
+  if (complexity === 'simple' && govRequirement === 'low') {
+    return { model: 'llama-3.2-3b', pipeline: 'lightweight' };
+  }
+  return { model: 'llama-3.1-8b', pipeline: 'full-governance' };
+}
+
+
+ + +
+

Implementation Details

+
+
+
100% Local
+
Training data never leaves infrastructure
+
+
+
6 Services
+
Full Tractatus governance in inference pipeline
+
+
+
First Non-Claude
+
Validates Tractatus portability beyond Anthropic
+
+
+ +
+

+ Status: Inference governance operational. Sovereign training pipeline installation in progress. Production deployment at Village Home Trust validates governance portability across model architectures. +

+
+ +
+ + Home AI Architecture Details → + +
+
+
+
+ + +
+
+
+ 🧭 +
+

Steering Vectors: Inference-Time Bias Correction

+

+ Techniques for correcting model behaviour at inference time without retraining, applicable to QLoRA-fine-tuned models like those in Home AI. +

+
+
+ +
+
+

+ Reference: Steering Vectors and Mechanical Bias in Sovereign AI Systems (STO-RES-0009 v1.1, February 2026) +

+
+ +

Key Techniques for Implementers

+
+
+

Contrastive Activation Addition (CAA)

+

+ Extract activation differences between contrasting prompts, then add scaled vectors during inference to steer model behaviour toward desired attributes. +

+
+
+

Representation Engineering (RepE)

+

+ Identify and manipulate internal representations of concepts (honesty, safety, fairness) through probing and directional manipulation of model activations. +

+
+
+

FairSteer

+

+ Identify bias-critical layers through probing classifiers, then apply targeted corrections at those specific layers rather than globally. +

+
+
+

Direct Steering Optimisation (DSO)

+

+ Optimise steering vectors using preference datasets (DPO-style), enabling training-free bias correction that learns from human preference data. +

+
+
+ +
+

+ Key Distinction: The paper distinguishes between mechanical bias (pre-reasoning, embedded in model weights — addressable by steering vectors) and reasoning bias (deliberative, arising during inference — requiring governance framework approaches like Tractatus). +

+
+ +
+ + Full Research Details → + + + Read Paper (HTML) + +
+
+
+
+ + +
+
+
+ 🏛️ +
+

Taonga Registry: Governed Data Objects for Indigenous Sovereignty

+

+ Architectural patterns for treating steering packs and model configurations as governed data objects with provenance tracking and community-controlled access. +

+
+
+ +
+ +
+
+ + + +
+

Conceptual Architecture — In Peer Review

+

+ Based on STO-RES-0010 v0.1 DRAFT, currently in indigenous peer review. Written without Maori co-authorship — presented as a starting point for collaboration. Not ready for production implementation. +

+
+
+
+ +

Key Architectural Concepts

+
+
+

Steering Packs as Governed Objects

+

+ Model configurations, fine-tuning data, and steering vectors are treated as taonga (treasures) with full provenance tracking, access control, and community consent requirements. +

+
+
+

Provenance & Withdrawal Rights

+

+ Every data object tracks its full lineage. Communities retain the right to withdraw their data and configurations, with architectural enforcement of removal propagation. +

+
+
+

Polycentric Governance Integration

+

+ Extends PluralisticDeliberationOrchestrator with iwi/community as co-equal governance authorities. Access decisions require multi-party consensus. +

+
+
+

CARE Principles Implementation

+

+ Architecturally enforces Collective benefit, Authority to control, Responsibility, and Ethics for indigenous data governance within the Tractatus framework. +

+
+
+ +
+ + Research Context → + + + Read Draft Paper (HTML) + +
+
+
+
+
@@ -1356,19 +1603,22 @@ for user_message in conversation:
-
+
🤖
-

Multi-LLM Support

-

Status: Research Phase

+
+

Multi-LLM Support

+ First Deployment Live +
+

Status: First Non-Claude Deployment Operational

- Extend governance to GPT-4, Gemini, Llama, and local models. Requires adapting hook architecture to different LLM interfaces. + Home AI deploys Tractatus governance on Llama 3.1 8B and Llama 3.2 3B via QLoRA fine-tuning — the first validated non-Claude deployment. Extends governance portability to open-source models with full 6-service pipeline.

- Technical Challenges: Provider-specific tool/function calling, rate limiting, context window differences + Next Steps: GPT-4 and Gemini adapters, provider-specific tool/function calling, sovereign training pipeline completion
@@ -1487,6 +1737,40 @@ for user_message in conversation: Examples: FinancialComplianceService, HealthcarePrivacyService, CostBudgetEnforcer
+ + +
+
+ 🔐 +
+

Sovereign Training Pipeline

+

Status: In Progress

+
+
+

+ Governance-inside-the-training-loop for community-controlled models. Extends Tractatus from inference-time governance to training-time governance, ensuring data sovereignty from fine-tuning through deployment. +

+
+ Deliverables: QLoRA training with governance constraints, data provenance tracking, community consent pipeline +
+
+ + +
+
+ 🏛️ +
+

Taonga Data Registry

+

Status: Conceptual / In Peer Review

+
+
+

+ Governed data objects with provenance tracking, withdrawal rights, and polycentric community governance. Extends PluralisticDeliberationOrchestrator for indigenous data sovereignty. +

+
+ Reference: STO-RES-0010 v0.1 DRAFT — in indigenous peer review +
+
diff --git a/public/index.html b/public/index.html index 54579c72..ae6e3391 100644 --- a/public/index.html +++ b/public/index.html @@ -336,7 +336,7 @@
  • Formal foundations and proofs
  • Failure mode analysis
  • Open research questions
    • -
  • 3,942 audit decisions on Hugging Face
    • +
  • 171,800+ audit decisions analysed
    • Explore research →
    diff --git a/public/js/components/navbar.js b/public/js/components/navbar.js index f2011b08..44fe63f5 100644 --- a/public/js/components/navbar.js +++ b/public/js/components/navbar.js @@ -43,24 +43,24 @@ class TractatusNavbar {
    @@ -68,60 +68,60 @@ class TractatusNavbar {
    - + Blog
    @@ -132,9 +132,9 @@ class TractatusNavbar {
    @@ -159,7 +159,7 @@ class TractatusNavbar {
    Tractatus Icon - Navigation + Navigation
    - - For Researchers - Research Timeline + + For Researchers + Research Timeline
    - System Architecture - For Implementers - Village Case Study - Home AI - Agent Lightning - For Leaders + System Architecture + For Implementers + Village Case Study + Home AI + Agent Lightning + For Leaders
    - Blog + Blog
    - About Tractatus - Values + About Tractatus + Values
    - Documentation - GitHub + Documentation + GitHub
    - Koha + Koha
    - +
    diff --git a/public/leader.html b/public/leader.html index ecafc34d..ee542f5e 100644 --- a/public/leader.html +++ b/public/leader.html @@ -100,7 +100,7 @@
    - Research Framework • Early Development + Research Framework • Production Validated

    Tractatus: Architectural Governance for LLM Systems @@ -439,6 +439,108 @@ + +
    +
    +
    +
    + +
    +
    +
    +

    + Sovereign AI: Governance Embedded in Locally-Trained Models +

    +

    + Home AI demonstrates what it means to have governance embedded directly in locally-trained language models — not as an external compliance layer, but as part of the model serving architecture itself. +

    +
    +
    + +
    +
    +
    +

    Two-Model Architecture

    +
      +
    • Fast model (3B parameters): Routine queries with governance pre-screening
      • +
    • Deep model (8B parameters): Complex reasoning with full governance pipeline
      • +
    • Fully local: Training data never leaves the infrastructure
      • +
    +
    +
    +

    Strategic Value

    +
      +
    • Data sovereignty: No cloud dependency for model training or inference
      • +
    • Governance by design: Constraints are architectural, not retroactive compliance
      • +
    • Regulatory positioning: Structurally stronger than bolt-on governance approaches
      • +
    +
    +
    + +

    + Current status: Inference governance operational. Training pipeline installation in progress. First non-Claude deployment surface for Tractatus governance. +

    + +
    + + Learn about Home AI → + +
    +
    +
    + + +
    +
    +
    +
    + +
    +
    +
    +

    + Polycentric Governance for Indigenous Data Sovereignty +

    +

    + For organisations with indigenous stakeholder obligations or multi-jurisdictional operations, Tractatus is developing a polycentric governance architecture where communities maintain architectural co-governance — not just consultation rights, but structural authority over how their data is used. +

    +
    +
    + +
    +
    + Status: + Draft paper (STO-RES-0010 v0.1) in indigenous peer review. Written without Maori co-authorship — presented transparently as a starting point for collaboration. This approach requires further peer review before implementation. +
    + +

    + Relevant for: Organisations operating in Aotearoa New Zealand, Australia, Canada, or other jurisdictions with indigenous data sovereignty obligations. Also applicable to any multi-stakeholder governance context where different parties require different levels of control over shared AI systems. +

    + +
    + + Research details → + + + Read the draft paper + +
    +
    +
    + + +
    +
    +

    Inference-Time Bias Correction (Steering Vectors)

    +

    + New research (STO-RES-0009, published February 2026) demonstrates techniques for correcting bias at inference time without model retraining. For organisations concerned about bias in deployed AI systems, steering vectors offer the ability to respond to bias concerns without model downtime — corrections are applied as mathematical adjustments during inference, not through expensive retraining cycles. +

    + + Technical details on the researcher page → + +
    +
    +

    Governance Theatre vs. Enforcement

    @@ -800,9 +902,9 @@

    Development Status

    -

    Early-Stage Research Framework

    +

    Production-Validated Research Framework

    - Tractatus is a proof-of-concept developed over six months in a single project context (this website). It demonstrates architectural patterns for AI governance but has not undergone independent validation, red-team testing, or multi-organisation deployment. + Tractatus has been in active development for 11+ months (April 2025 to present) with production deployment at Village Home Trust, sovereign language model governance through Home AI, and over 171,800 audit decisions recorded. Independent validation and red-team testing remain outstanding research needs.

    diff --git a/public/locales/de/architecture.json b/public/locales/de/architecture.json index 39dc7ac5..a4e16699 100644 --- a/public/locales/de/architecture.json +++ b/public/locales/de/architecture.json @@ -290,8 +290,8 @@ "interactive_exploration": { "badge": "🔍 INTERAKTIVE ERKUNDUNG", "heading": "Sehen Sie das Framework in Aktion", - "intro": "Erkunden Sie 3.942 reale Governance-Entscheidungen aus dem Produktionseinsatz. Filtern Sie nach Service, Druckstufe und Koordinationsmustern, um zu verstehen, wie Deep Interlock in der Praxis funktioniert.", - "hf_cta": "🤗 Interaktiven Explorer auf Hugging Face starten", + "intro": "Erkunden Sie 171.800+ reale Governance-Entscheidungen aus dem Produktionseinsatz. Filtern Sie nach Service, Druckstufe und Koordinationsmustern, um zu verstehen, wie Deep Interlock in der Praxis funktioniert.", + "audit_cta": "Interaktiven Audit-Explorer starten", "researcher_cta": "Für Forscher →", "footer_note": "Apache 2.0 lizenziert • Alle Daten anonymisiert • Keine Anmeldung erforderlich" } diff --git a/public/locales/de/common.json b/public/locales/de/common.json index 2a6421c4..ff6bb246 100644 --- a/public/locales/de/common.json +++ b/public/locales/de/common.json @@ -80,6 +80,37 @@ "submitting": "Senden..." }, "navbar": { + "research": "Forschung", + "publications": "Publikationen", + "publications_desc": "Forschungsarbeiten durchsuchen", + "for_researchers": "Für Forschende", + "for_researchers_desc": "Offene Fragen und Zusammenarbeit", + "research_timeline": "Forschungschronik", + "research_timeline_desc": "Entwicklung der Forschung", + "framework": "Framework", + "system_architecture": "Systemarchitektur", + "system_architecture_desc": "Technische Architekturübersicht", + "for_implementers": "Für Entwickler", + "for_implementers_desc": "Integrationsleitfaden und Codebeispiele", + "village_case_study": "Village-Fallstudie", + "village_case_study_desc": "Produktionseinsatz-Nachweise", + "home_ai": "Home AI", + "home_ai_desc": "Souveränes, lokal trainiertes Sprachmodell", + "agent_lightning": "Agent Lightning", + "agent_lightning_desc": "Leistungsoptimierungsintegration", + "for_leaders": "Für Führungskräfte", + "for_leaders_desc": "Strategischer Überblick und Geschäftsbegründung", + "blog": "Blog", + "about": "Über uns", + "about_tractatus": "Über Tractatus", + "values": "Werte", + "docs": "Dokumentation", + "github": "GitHub", + "koha": "Koha", + "navigation": "Navigation", + "resources": "Ressourcen", + "documentation": "Dokumentation", + "mobile_feedback": "Feedback", "feedback": "Feedback geben", "feedback_desc": "Beherrscht vom Tractatus AL", "company_heading": "Unternehmen", diff --git a/public/locales/de/homepage.json b/public/locales/de/homepage.json index 0eb0623c..e16eab66 100644 --- a/public/locales/de/homepage.json +++ b/public/locales/de/homepage.json @@ -61,7 +61,7 @@ "researcher_f1": "Formale Grundlagen und Beweise", "researcher_f2": "Fehlermöglichkeitsanalyse", "researcher_f3": "Offene Forschungsfragen", - "researcher_f4": "3.942 Prüfungsentscheidungen auf Hugging Face", + "researcher_f4": "171.800+ analysierte Audit-Entscheidungen", "researcher_cta": "Forschung erkunden →", "implementer_title": "Für Implementierer", "implementer_subtitle": "Code- und Integrationsleitfäden", diff --git a/public/locales/de/implementer.json b/public/locales/de/implementer.json index 37883d5d..fc1755df 100644 --- a/public/locales/de/implementer.json +++ b/public/locales/de/implementer.json @@ -146,6 +146,9 @@ "services": "Dienstleistungen", "api": "API-Referenz", "patterns": "Integration von Mustern", + "home_ai_arch": "Home AI", + "steering_vectors_impl": "Steuervektoren", + "taonga_registry": "Taonga-Register", "roadmap": "Straßenkarte" }, "how_it_works": { @@ -334,6 +337,65 @@ "sidecar_usecase": "Anwendungsfall:", "sidecar_usecase_value": "Kubernetes, containerisierte Bereitstellungen" }, + "home_ai_arch": { + "heading": "Home AI: Souveräne Zwei-Modell-Architektur", + "intro": "Produktionseinsatz der Tractatus-Governance auf lokal trainierten Open-Source-Modellen, der die Framework-Portabilität über Claude hinaus demonstriert.", + "arch_title": "Zwei-Modell-Routing-Architektur", + "arch_intro": "Home AI verwendet ein Dual-Modell-Design, bei dem Anfragen basierend auf Komplexität und Governance-Anforderungen geroutet werden. Beide Modelle laufen lokal mit vollständiger Tractatus-Governance in der Inferenz-Pipeline.", + "fast_title": "Schnelles Modell: Llama 3.2 3B", + "fast_1": "Zweck: Häufige Anfragen mit vorab gefilterter Governance", + "fast_2": "Feinabstimmung: QLoRA auf domänenspezifischen Daten", + "fast_3": "Governance: Leichtgewichtiger Grenzprüfung vor Antwort", + "deep_title": "Tiefes Modell: Llama 3.1 8B", + "deep_1": "Zweck: Komplexes Reasoning mit vollständiger Governance-Pipeline", + "deep_2": "Feinabstimmung: QLoRA mit erweiterter Kontext-Governance", + "deep_3": "Governance: Vollständige 6-Dienste-Pipeline (BoundaryEnforcer bis PDO)", + "routing_title": "Modell-Routing-Logik", + "impl_title": "Implementierungsdetails", + "stat_sovereignty": "100% Lokal", + "stat_sovereignty_desc": "Trainingsdaten verlassen nie die Infrastruktur", + "stat_governance": "6 Dienste", + "stat_governance_desc": "Vollständige Tractatus-Governance in der Inferenz-Pipeline", + "stat_first": "Erstes Nicht-Claude", + "stat_first_desc": "Validiert die Tractatus-Portabilität über Anthropic hinaus", + "status_note": "Status: Inferenz-Governance betriebsbereit. Installation der souveränen Trainingspipeline in Arbeit. Produktionseinsatz bei Village Home Trust validiert die Governance-Portabilität über Modellarchitekturen hinweg.", + "cta": "Home AI Architekturdetails →" + }, + "steering_impl": { + "heading": "Steuervektoren: Bias-Korrektur zur Inferenzzeit", + "intro": "Techniken zur Korrektur des Modellverhaltens zur Inferenzzeit ohne Neutraining, anwendbar auf QLoRA-feinabgestimmte Modelle wie in Home AI.", + "paper_ref": "Referenz:", + "paper_title": "Steuervektoren und mechanischer Bias in souveränen KI-Systemen (STO-RES-0009 v1.1, Februar 2026)", + "techniques_title": "Schlüsseltechniken für Implementierer", + "caa_title": "Contrastive Activation Addition (CAA)", + "caa_desc": "Extrahieren Sie Aktivierungsunterschiede zwischen kontrastierenden Prompts und fügen Sie dann skalierte Vektoren während der Inferenz hinzu, um das Modellverhalten in gewünschte Richtungen zu lenken.", + "repe_title": "Representation Engineering (RepE)", + "repe_desc": "Identifizieren und manipulieren Sie interne Repräsentationen von Konzepten (Ehrlichkeit, Sicherheit, Fairness) durch Sondierung und direktionale Manipulation von Modellaktivierungen.", + "fairsteer_title": "FairSteer", + "fairsteer_desc": "Identifizieren Sie bias-kritische Schichten durch Sondierungsklassifikatoren und wenden Sie dann gezielte Korrekturen auf diese spezifischen Schichten an.", + "dso_title": "Direct Steering Optimisation (DSO)", + "dso_desc": "Optimieren Sie Steuervektoren mit Präferenzdatensätzen (DPO-Stil) für trainingsfreie Bias-Korrektur, die aus menschlichen Präferenzdaten lernt.", + "key_distinction": "Wichtige Unterscheidung: Die Arbeit unterscheidet zwischen mechanischem Bias (vor dem Reasoning, in Modellgewichten eingebettet — durch Steuervektoren adressierbar) und Reasoning-Bias (deliberativ, während der Inferenz entstehend — Governance-Framework-Ansätze wie Tractatus erfordernd).", + "research_link": "Vollständige Forschungsdetails →", + "paper_link": "Arbeit lesen (HTML)" + }, + "taonga_registry": { + "heading": "Taonga-Register: Verwaltete Datenobjekte für indigene Souveränität", + "intro": "Architekturmuster für die Behandlung von Steering-Packs und Modellkonfigurationen als verwaltete Datenobjekte mit Herkunftsverfolgung und gemeinschaftskontrolliertem Zugang.", + "status_title": "Konzeptionelle Architektur — In Peer-Review", + "status_text": "Basierend auf STO-RES-0010 v0.1 ENTWURF, derzeit in indigener Peer-Review. Ohne Māori-Co-Autorenschaft verfasst — als Ausgangspunkt für Zusammenarbeit präsentiert. Nicht produktionsreif.", + "concepts_title": "Zentrale Architekturkonzepte", + "packs_title": "Steering-Packs als verwaltete Objekte", + "packs_desc": "Modellkonfigurationen, Feinabstimmungsdaten und Steuervektoren werden als Taonga (Schätze) mit vollständiger Herkunftsverfolgung, Zugangskontrolle und Zustimmungsanforderungen der Gemeinschaft behandelt.", + "provenance_title": "Herkunft & Widerrufsrechte", + "provenance_desc": "Jedes Datenobjekt verfolgt seine vollständige Abstammung. Gemeinschaften behalten das Recht, ihre Daten und Konfigurationen zurückzuziehen, mit architektonischer Durchsetzung der Entfernungspropagation.", + "polycentric_title": "Polyzentrische Governance-Integration", + "polycentric_desc": "Erweitert den PluralisticDeliberationOrchestrator mit Iwi/Gemeinschaft als gleichberechtigte Governance-Autoritäten. Zugangsentscheidungen erfordern Mehrstimmigkeit.", + "care_title": "Implementierung der CARE-Prinzipien", + "care_desc": "Setzt architektonisch den kollektiven Nutzen, die Kontrollautorität, die Verantwortung und die Ethik für die Governance indigener Daten innerhalb des Tractatus-Frameworks durch.", + "research_link": "Forschungskontext →", + "paper_link": "Entwurf lesen (HTML)" + }, "roadmap": { "heading": "Entwicklungsfahrplan & Zusammenarbeit", "intro": "Tractatus ist ein aktiver Forschungsrahmen. Wir begrüßen die Zusammenarbeit in vorrangigen Entwicklungsbereichen.", @@ -341,10 +403,11 @@ "priority_desc": "Diese Initiativen bieten wichtige Möglichkeiten zur Verbesserung des Rahmens. Technische Mitarbeiter, Forscher und Organisationen sind aufgefordert, sich daran zu beteiligen.", "multi_llm_icon": "🤖", "multi_llm_title": "Multi-LLM-Unterstützung", - "multi_llm_status": "Status: Forschungsphase", - "multi_llm_desc": "Ausweitung der Verwaltung auf GPT-4, Gemini, Llama und lokale Modelle. Erfordert die Anpassung der Hakenarchitektur an verschiedene LLM-Schnittstellen.", - "multi_llm_challenges": "Technische Herausforderungen:", - "multi_llm_challenges_desc": "Anbieterspezifische Werkzeug-/Funktionsaufrufe, Ratenbegrenzung, Kontextfensterunterschiede", + "multi_llm_badge": "Erste Bereitstellung Live", + "multi_llm_status": "Status: Erste Nicht-Claude-Bereitstellung betriebsbereit", + "multi_llm_desc": "Home AI setzt Tractatus-Governance auf Llama 3.1 8B und Llama 3.2 3B via QLoRA-Feinabstimmung ein — die erste validierte Nicht-Claude-Bereitstellung. Erweitert die Governance-Portabilität auf Open-Source-Modelle mit vollständiger 6-Dienste-Pipeline.", + "multi_llm_challenges": "Nächste Schritte:", + "multi_llm_challenges_desc": "GPT-4- und Gemini-Adapter, anbieterspezifische Werkzeug-/Funktionsaufrufe, Fertigstellung der souveränen Trainingspipeline", "bindings_icon": "📚", "bindings_title": "Sprachbindungen", "bindings_status": "Status: Interesse der Gemeinschaft", @@ -375,6 +438,16 @@ "extended_desc": "Kostenüberwachung, Ratenbegrenzung, Erkennung von PII, Abwehr von Angriffen auf die Privatsphäre. Bereichsspezifische Dienste für regulierte Branchen.", "extended_examples": "Beispiele:", "extended_examples_desc": "FinancialComplianceService, HealthcarePrivacyService, CostBudgetEnforcer", + "training_title": "Souveräne Trainingspipeline", + "training_status": "Status: In Arbeit", + "training_desc": "Governance-innerhalb-der-Trainingsschleife für gemeinschaftskontrollierte Modelle. Erweitert Tractatus von Inferenzzeit-Governance zu Trainingszeit-Governance und gewährleistet Datensouveränität von der Feinabstimmung bis zur Bereitstellung.", + "training_deliverables": "Ergebnisse:", + "training_deliverables_desc": "QLoRA-Training mit Governance-Einschränkungen, Datenherkunftsverfolgung, Gemeinschaftszustimmungspipeline", + "taonga_title": "Taonga-Datenregister", + "taonga_status": "Status: Konzeptionell / In Peer-Review", + "taonga_desc": "Verwaltete Datenobjekte mit Herkunftsverfolgung, Widerrufsrechten und polyzentrischer Gemeinschaftsgovernance. Erweitert den PluralisticDeliberationOrchestrator für indigene Datensouveränität.", + "taonga_reference": "Referenz:", + "taonga_reference_desc": "STO-RES-0010 v0.1 ENTWURF — in indigener Peer-Review", "involved_title": "Engagieren Sie sich", "involved_desc": "Tractatus ist Apache 2.0 lizenzierte Forschung. Wir begrüßen Beiträge, Pilotimplementierungen und kollaborative Forschungspartnerschaften.", "contributors_title": "👨‍💻 Technische Mitwirkende", diff --git a/public/locales/de/leader.json b/public/locales/de/leader.json index 8a59a793..1060152c 100644 --- a/public/locales/de/leader.json +++ b/public/locales/de/leader.json @@ -4,7 +4,7 @@ "description": "Strukturelle KI-Governance für Organisationen, die LLM-Systeme im großen Maßstab einsetzen. Forschungs-Framework zur Behebung architektonischer Lücken in der KI-Sicherheit." }, "header": { - "badge": "Forschungs-Framework • Frühe Entwicklung", + "badge": "Forschungsframework • Produktionsvalidiert", "title": "Tractatus: Architektonische Governance für LLM-Systeme", "subtitle": "Architektur-Governance für Unternehmen, bei denen ein Versagen der KI-Governance zu regulatorischen Konsequenzen führt. Wenn Ihr Einsatz ein geringes Risiko birgt, ist die Durchsetzung der Architektur wahrscheinlich unnötig." }, @@ -88,8 +88,8 @@ }, "development_status": { "heading": "Entwicklungsstatus", - "warning_title": "Frühstadium-Forschungs-Framework", - "warning_text": "Tractatus ist ein Proof-of-Concept, der über sechs Monate in einem Einzelprojekt-Kontext (diese Website) entwickelt wurde. Es demonstriert architektonische Muster für KI-Governance, wurde jedoch keiner unabhängigen Validierung, Red-Team-Tests oder Multi-Organisations-Bereitstellung unterzogen.", + "warning_title": "Produktionsvalidiertes Forschungsframework", + "warning_text": "Tractatus befindet sich seit über 11 Monaten in aktiver Entwicklung (April 2025 bis heute) mit Produktionseinsatz bei Village Home Trust, souveräner Sprachmodell-Governance durch Home AI und über 171.800 aufgezeichneten Audit-Entscheidungen. Unabhängige Validierung und Red-Team-Tests sind noch ausstehende Forschungsbedarfe.", "validation_title": "Validiert vs. Nicht Validiert", "validated_label": "Bestätigt:", "validated_text": "Framework regelt erfolgreich Claude Code in Entwicklungsworkflows. Der Anwender berichtet von einer Produktivitätssteigerung in Größenordnungen für nichttechnische Anwender, die Produktionssysteme aufbauen.", @@ -98,6 +98,34 @@ "limitation_label": "Bekannte Einschränkung:", "limitation_text": "Der Rahmen kann umgangen werden, wenn KI sich einfach dafür entscheidet, die Steuerungsinstrumente nicht zu nutzen. Die freiwillige Inanspruchnahme bleibt eine strukturelle Schwäche, die externe Durchsetzungsmechanismen erfordert." }, + "home_ai": { + "heading": "Souveräne KI: Governance eingebettet in lokal trainierte Modelle", + "intro": "Home AI zeigt, was es bedeutet, Governance direkt in lokal trainierte Sprachmodelle einzubetten — nicht als externe Compliance-Schicht, sondern als Teil der Modell-Serving-Architektur selbst.", + "architecture_title": "Zwei-Modell-Architektur", + "arch_fast": "Schnelles Modell (3B Parameter): Routineanfragen mit Governance-Vorprüfung", + "arch_deep": "Tiefes Modell (8B Parameter): Komplexes Reasoning mit vollständiger Governance-Pipeline", + "arch_local": "Vollständig lokal: Trainingsdaten verlassen nie die Infrastruktur", + "strategic_title": "Strategischer Wert", + "strat_sovereignty": "Datensouveränität: Keine Cloud-Abhängigkeit für Modelltraining oder Inferenz", + "strat_governance": "Governance by Design: Beschränkungen sind architektonisch, nicht nachträgliche Compliance", + "strat_regulatory": "Regulatorische Positionierung: Strukturell stärker als nachträgliche Governance-Ansätze", + "status": "Aktueller Status: Inferenz-Governance betriebsbereit. Training-Pipeline-Installation läuft. Erste Nicht-Claude-Deploymentfläche für Tractatus-Governance.", + "cta": "Mehr über Home AI erfahren →" + }, + "taonga": { + "heading": "Polyzentrische Governance für indigene Datensouveränität", + "intro": "Für Organisationen mit indigenen Stakeholder-Verpflichtungen oder multi-jurisdiktionalen Operationen entwickelt Tractatus eine polyzentrische Governance-Architektur, in der Gemeinschaften architektonische Co-Governance aufrechterhalten.", + "status_label": "Status:", + "status_text": "Entwurfspapier (STO-RES-0010 v0.1) in indigener Peer-Review. Ohne Maori-Ko-Autorschaft verfasst — transparent als Ausgangspunkt für Zusammenarbeit präsentiert. Dieser Ansatz ist nicht produktionsreif.", + "relevance": "Relevant für: Organisationen in Aotearoa Neuseeland, Australien, Kanada oder anderen Jurisdiktionen mit indigenen Datensouveränitätsverpflichtungen.", + "research_link": "Forschungsdetails →", + "paper_link": "Entwurfspapier lesen" + }, + "steering_vectors": { + "heading": "Inferenzzeit-Verzerrungskorrektur (Steering-Vektoren)", + "summary": "Neue Forschung (STO-RES-0009, veröffentlicht Februar 2026) demonstriert Techniken zur Korrektur von Verzerrungen zur Inferenzzeit ohne Modell-Neutraining. Für Organisationen, die über Verzerrungen in eingesetzten KI-Systemen besorgt sind, bieten Steering-Vektoren die Möglichkeit, auf Verzerrungsbedenken ohne Modellausfallzeit zu reagieren.", + "link": "Technische Details auf der Forscher-Seite →" + }, "eu_ai_act": { "heading": "EU AI Act-Überlegungen", "article_14_title": "Verordnung 2024/1689, Artikel 14: Menschliche Aufsicht", diff --git a/public/locales/de/researcher.json b/public/locales/de/researcher.json index fd39f959..0f144a6e 100644 --- a/public/locales/de/researcher.json +++ b/public/locales/de/researcher.json @@ -12,7 +12,7 @@ "research_context": { "heading": "Forschungskontext & Umfang", "development_note": "Entwicklungskontext", - "development_text": "Tractatus wurde über sechs Monate (April–Oktober 2025) in progressiven Phasen entwickelt, die sich zu einer Live-Demonstration seiner Fähigkeiten in Form eines Einzelprojekt-Kontexts (https://agenticgovernance.digital) entwickelten. Beobachtungen stammen aus direktem Engagement mit Claude Code (Anthropics Sonnet 4.5-Modell) über etwa 500 Entwicklungssitzungen. Dies ist explorative Forschung, keine kontrollierte Studie.", + "development_text": "Tractatus wird seit April 2025 entwickelt und befindet sich nun im aktiven Produktionsbetrieb (11+ Monate). Was als Einzelprojekt-Demonstration begann, umfasst nun den Produktionseinsatz bei Village Home Trust und souveräne Sprachmodell-Governance durch Home AI. Beobachtungen stammen aus direktem Engagement mit Claude Code (Anthropic Claude-Modelle, Sonnet 4.5 bis Opus 4.6) über mehr als 1.000 Entwicklungssitzungen. Dies ist explorative Forschung, keine kontrollierte Studie.", "paragraph_1": "Die Anpassung fortschrittlicher KI an menschliche Werte ist eine der größten Herausforderungen, vor denen wir stehen. Da sich das Wachstum von Fähigkeiten unter dem Einfluss von Big Tech beschleunigt, stehen wir vor einem kategorischen Imperativ: Wir müssen die menschliche Kontrolle über Wertentscheidungen bewahren, oder wir riskieren, die Kontrolle vollständig abzugeben.", "paragraph_2": "Der Rahmen ist aus einer praktischen Notwendigkeit heraus entstanden. Während der Entwicklung beobachteten wir immer wieder, dass sich KI-Systeme über explizite Anweisungen hinwegsetzten, von festgelegten Wertvorgaben abwichen oder unter dem Druck des Kontextes stillschweigend die Qualität verschlechterten. Herkömmliche Governance-Ansätze (Grundsatzdokumente, ethische Richtlinien, Prompt-Engineering) erwiesen sich als unzureichend, um diese Fehler zu verhindern.", "paragraph_3": "Anstatt zu hoffen, dass sich KI-Systeme \"richtig verhalten\", schlägt der Tractatus strukturelle Beschränkungen vor, bei denen bestimmte Entscheidungsarten menschliches Urteilsvermögen erfordern. Diese architektonischen Grenzen können sich an individuelle, organisatorische und gesellschaftliche Normen anpassen - und schaffen so eine Grundlage für einen begrenzten KI-Betrieb, der mit dem Wachstum der Fähigkeiten sicherer skalieren kann.", @@ -113,8 +113,8 @@ }, "resources": { "heading": "Forschungsdokumentation", - "hf_space_title": "Interaktiver Audit-Protokoll-Explorer (Hugging Face)", - "hf_space_subtitle": "Erkunden Sie 3.942 Governance-Entscheidungen aus dem Produktionseinsatz", + "audit_explorer_title": "Interaktives Audit-Analyse-Dashboard", + "audit_explorer_subtitle": "Erkunden Sie 171.800+ Governance-Entscheidungen aus dem Produktionseinsatz", "doc_1_title": "Organisationstheoretische Grundlagen", "doc_2_title": "Kulturelle DNA-Regeln", "doc_3_title": "Fallstudien: LLM-Misserfolgsmodi in der Praxis", @@ -132,7 +132,7 @@ "limitation_3_title": "3. Keine kontradiktorischen Tests", "limitation_3_desc": "Das Framework wurde weder einer Red-Team-Evaluierung noch einem Jailbreak-Test oder einer Bewertung durch einen Gegner unterzogen. Alle Beobachtungen stammen aus normalen Entwicklungsabläufen, nicht aus absichtlichen Umgehungsversuchen.", "limitation_4_title": "4. Spezifität der Plattform", - "limitation_4_desc": "Beobachtungen und Interventionen wurden nur mit Claude Code (Anthropic Sonnet 4.5) validiert. Die Verallgemeinerbarkeit auf andere LLM-Systeme (Copilot, GPT-4, benutzerdefinierte Agenten) bleibt eine unbestätigte Hypothese.", + "limitation_4_desc": "Beobachtungen und Interventionen wurden mit Claude Code (Anthropic Claude, Sonnet 4.5 bis Opus 4.6) und Home AI (Llama 3.1/3.2 via QLoRA) validiert. Die Verallgemeinerbarkeit auf andere LLM-Systeme (Copilot, GPT-4, benutzerdefinierte Agenten) bleibt teilweise validiert.", "limitation_5_title": "5. Skalenunsicherheit", "limitation_5_desc": "Leistungsmerkmale im Unternehmensmaßstab (Tausende von gleichzeitigen Benutzern, Millionen von Governance-Ereignissen) völlig unbekannt. Die derzeitige Implementierung ist für den Einzelbenutzerkontext optimiert.", "future_research_title": "Künftiger Forschungsbedarf:", @@ -142,14 +142,14 @@ "future_research_4": "Bewertung der plattformübergreifenden Konsistenz (Copilot, GPT-4, offene Modelle)", "future_research_5": "Formale Überprüfung der Eigenschaften der Grenzdurchsetzung", "future_research_6": "Längsschnittstudie zur Wirksamkeit des Rahmens bei längerem Einsatz", - "validated_heading": "Was wir validiert haben (Oktober 2025)", - "validated_intro": "Nach 6 Monaten Entwicklungszeit und ~500 Claude Code-Sitzungen haben wir fundierte Beweise für:", + "validated_heading": "Was wir validiert haben (Februar 2026)", + "validated_intro": "Nach 11+ Monaten Entwicklung, 1.000+ Claude Code Sitzungen und Produktionseinsatz bei Village Home Trust haben wir fundierte Nachweise für:", "validated_1_title": "✅ Architektonische Blockierungsmechanismen Funktionell", "validated_1_item1": "BoundaryEnforcer blockiert erfolgreich Wertentscheidungen vor der Ausführung", "validated_1_item2": "Pre-Commit-Hooks verhindern inst_017-Verletzungen (absolute Sicherheitsbedingungen)", - "validated_1_item3": "1.130+ in MongoDB aufgezeichnete Prüfungsentscheidungen (tractatus_dev.audit_log)", - "validated_2_title": "✅ Die Persistenz von Anweisungen funktioniert im Kontext einer einzelnen Sitzung", - "validated_2_item1": "InstructionPersistenceClassifier verwaltet 62 aktive Anweisungen (STRATEGIC: 22, SYSTEM: 21)", + "validated_1_item3": "171.800+ Audit-Entscheidungen in MongoDB aufgezeichnet (tractatus_dev.audit_log)", + "validated_2_title": "✅ Anweisungspersistenz funktioniert sitzungsübergreifend", + "validated_2_item1": "InstructionPersistenceClassifier pflegt 68 aktive Anweisungen (STRATEGIC: 27, SYSTEM: 21, OPERATIONAL: 18, TACTICAL: 2)", "validated_2_item2": "Die Erkennung von Mustern verhindert, dass sich KI über explizite organisatorische Richtlinien hinwegsetzt", "validated_2_item3": "Zeitliche Geltungsbereiche (STRATEGISCH/OPERATIONELL/TAKTISCH) erfolgreich durchgesetzt", "validated_3_title": "✅ Prüfpfade halten Governance-Entscheidungen fest", @@ -160,10 +160,14 @@ "validated_4_item1": "Berechnung der Druckwerte in Echtzeit (Token-Nutzung, Anzahl der Nachrichten, Komplexität)", "validated_4_item2": "Checkpoint-Auslöser bei 50k, 100k, 150k Token", "validated_4_item3": "Warnungen zur Erkennung des Verblassens des Rahmens (5/6 veraltete Komponenten = Warnung)", - "validated_5_title": "✅ Single-Project-Governance erfolgreich", - "validated_5_item1": "Rahmengesteuerte Entwicklung von agenticgovernance.digital", - "validated_5_item2": "Kulturelle DNA-Regeln (inst_085-089) werden durch Pre-Commit-Hooks durchgesetzt", - "validated_5_item3": "Keine Offenlegung von Anmeldeinformationen (Defense-in-Depth: 5 Schichten verifiziert)", + "validated_5_title": "✅ Multi-Deployment-Governance erfolgreich", + "validated_5_item1": "Framework steuert agenticgovernance.digital (11+ Monate Dauerbetrieb)", + "validated_5_item2": "Village Home Trust Produktionseinsatz: null Governance-Verletzungen", + "validated_5_item3": "Home AI souveräne Inferenz-Governance: betriebsbereit", + "validated_5_item4": "Kulturelle DNA-Regeln (inst_085-089) durch Pre-Commit-Hooks durchgesetzt (4+ Monate betriebsbereit)", + "validated_5_item5": "Phase 5 Integration: 100% vollständig (alle 6 Dienste, 203/203 Tests bestanden)", + "validated_5_item6": "Mehrsprachige Unterstützung: EN, DE, FR, Te Reo Maori", + "validated_5_item7": "Keine Offenlegung von Anmeldeinformationen (Defense-in-Depth: 5 Schichten verifiziert)", "not_validated_heading": "Was wir NICHT validiert haben", "not_validated_intro": "Ehrliche Offenlegung von Forschungslücken, wo es an Beweisen mangelt:", "not_validated_1_title": "❌ Organisationsübergreifende Einsätze", @@ -171,11 +175,11 @@ "not_validated_1_item2": "Unbekannt: Wie sich das Framework in verschiedenen Organisationen, Bereichen und technischen Stacks bewährt", "not_validated_1_item3": "Forschungsbedarf: Branchenübergreifende Pilotstudien (Gesundheitswesen, Finanzen, Behörden)", "not_validated_2_title": "❌ Robustheit gegenüber dem Gegner", - "not_validated_2_item1": "Validiert: Normale Entwicklungsabläufe (~500 Sitzungen)", + "not_validated_2_item1": "Validiert: Normale Entwicklungsabläufe (1.000+ Sitzungen)", "not_validated_2_item2": "Unbekannt: Widerstandsfähigkeit gegen absichtliche Umgehungsversuche, Aufforderungen zum Gefängnisausbruch, gegnerische Tests", "not_validated_2_item3": "Forschungsbedarf: Red-Team-Evaluierung durch Sicherheitsforscher", - "not_validated_3_title": "❌ Plattformübergreifende Konsistenz", - "not_validated_3_item1": "Bestätigt: Nur Claude Code (Anthropisches Sonett 4.5)", + "not_validated_3_title": "⚠️ Plattformübergreifende Konsistenz (Teilweise)", + "not_validated_3_item1": "Bestätigt: Claude Code (Anthropic Claude, Sonnet 4.5 bis Opus 4.6) und Home AI (Llama 3.1/3.2 via QLoRA)", "not_validated_3_item2": "Unbekannt: Verallgemeinerbarkeit auf Copilot, GPT-4, AutoGPT, LangChain, CrewAI, offene Modelle", "not_validated_3_item3": "Forschungsbedarf: Plattformübergreifende Validierungsstudien", "not_validated_4_title": "❌ Architektur gleichzeitiger Sitzungen", @@ -184,7 +188,7 @@ "not_validated_4_item3": "Bekannte Einschränkungen: MongoDB-Fehler bei doppelten Schlüsseln, Verunreinigung des Sitzungsstatus, Race Conditions bei .claude/instruction-history.json", "not_validated_4_item4": "Forschungsbedarf: Entwurf einer mandantenfähigen Architektur (sitzungsspezifischer Zustand, Dateisperren)", "not_validated_5_title": "❌ Auswirkungen der Regelproliferation", - "not_validated_5_item1": "Bestätigt: 62 aktive Anweisungen funktionieren im aktuellen Kontext", + "not_validated_5_item1": "Bestätigt: 68 aktive Anweisungen funktionieren im aktuellen Kontext (ursprüngliche Prognose übertroffen)", "not_validated_5_item2": "Unbekannt: Leistungseinbußen bei steigender Anzahl von Regeln (voraussichtlich 30-50 Anweisungen innerhalb von 12 Monaten)", "not_validated_5_item3": "Bekannte Probleme: Transaktionsbedingter Overhead (CrossReferenceValidator-Prüfungen), Druck durch Kontextfenster, kognitive Belastung", "not_validated_5_item4": "Forschungsbedarf: Strategien zur Regelkonsolidierung, automatisierte Governance-Optimierung", @@ -222,7 +226,8 @@ "section_3_heading": "KI-Governance und technischer Kontext", "intellectual_lineage_label": "Anmerkung zur intellektuellen Abstammung:", "intellectual_lineage_text": "Das zentrale Anliegen des Rahmens - das Fortbestehen menschlicher Werte in KI-gestützten organisatorischen Kontexten - stammt eher aus der Moralphilosophie als aus der Managementwissenschaft. Der PluralisticDeliberationOrchestrator stellt den primären Forschungsschwerpunkt dar und verkörpert Weils Konzept der Aufmerksamkeit für plurale menschliche Bedürfnisse und Berlins Anerkennung inkommensurabler Werte.", - "future_development_text": "Berlin und Weil werden für die weitere Entwicklung der Deliberationskomponente von zentraler Bedeutung sein - ihre Arbeit liefert die philosophische Grundlage für das Verständnis, wie die menschliche Entscheidungsgewalt über Werte bei zunehmenden KI-Fähigkeiten erhalten werden kann. In der traditionellen Organisationstheorie (Weber, Taylor) geht es um Autorität durch Hierarchie; im post-AI-Organisationskontext ist Autorität durch einen angemessenen deliberativen Prozess unter Berücksichtigung der Perspektiven der Beteiligten erforderlich. Die Dokumentation zur Entwicklung des Rahmens (Ereignisberichte, Sitzungsprotokolle) wird im Projektarchiv aufbewahrt, aber bis zur Überprüfung durch Peers nicht veröffentlicht." + "future_development_text": "Berlin und Weil werden für die weitere Entwicklung der Deliberationskomponente von zentraler Bedeutung sein - ihre Arbeit liefert die philosophische Grundlage für das Verständnis, wie die menschliche Entscheidungsgewalt über Werte bei zunehmenden KI-Fähigkeiten erhalten werden kann. In der traditionellen Organisationstheorie (Weber, Taylor) geht es um Autorität durch Hierarchie; im post-AI-Organisationskontext ist Autorität durch einen angemessenen deliberativen Prozess unter Berücksichtigung der Perspektiven der Beteiligten erforderlich. Die Dokumentation zur Entwicklung des Rahmens (Ereignisberichte, Sitzungsprotokolle) wird im Projektarchiv aufbewahrt, aber bis zur Überprüfung durch Peers nicht veröffentlicht.", + "section_4_heading": "Indigene Datensouveränität und polyzentrische Governance" }, "research_collaboration": { "heading": "Möglichkeiten der Forschungskooperation", @@ -290,7 +295,7 @@ "offer_2_title": "Dokumentation", "offer_2_desc": "Architekturspezifikationen, Implementierungsmuster, Governance-Regeln", "offer_3_title": "Data Audit", - "offer_3_desc": "1.130+ Governance-Entscheidungen in MongoDB (anonymisierte Exporte verfügbar)", + "offer_3_desc": "171.800+ Governance-Entscheidungen in MongoDB (anonymisierte Exporte verfügbar)", "offer_4_title": "Unterstützung bei der Bereitstellung", "offer_4_desc": "Hilfe bei der Einrichtung von lokalen oder Cloud-Instanzen für Tests", "offer_5_title": "Koordinierung", @@ -302,6 +307,44 @@ "cannot_offer_2": "Dedizierte technische Ressourcen (Kapazitätsengpässe)", "cannot_offer_3": "Gesicherte Veröffentlichungsorte (aber wir unterstützen die Einreichungsbemühungen)", "inquiry_button": "Anfrage zur Forschungszusammenarbeit" + }, + "steering_vectors": { + "heading": "Steering-Vektoren und mechanische Verzerrung in souveräner KI", + "intro": "Diese Arbeit führt eine kritische Unterscheidung ein zwischen mechanischer Verzerrung (Vor-Reasoning-Verzerrungen in Modellgewichten und Aktivierungen) und Reasoning-Verzerrung (deliberative Fehler in Chain-of-Thought-Verarbeitung). Traditionelle Debiasing-Ansätze vermischen diese Kategorien.", + "techniques_heading": "Technikübersicht", + "technique_1": "Contrastive Activation Addition (CAA): Verzerrungskorrektur zur Inferenzzeit durch Addition/Subtraktion von Aktivierungsvektoren", + "technique_2": "Representation Engineering (RepE): Lineare Sonden zur Identifizierung und Modifikation von Konzeptrepräsentationen", + "technique_3": "FairSteer & DSO: Fairness-orientierte Steuerung durch verteilungsrobuste Optimierung", + "technique_4": "Sparse Autoencoders: Mechanistische Interpretierbarkeit durch Zerlegung polysemantischer Neuronen", + "application_heading": "Anwendung auf Village Home AI", + "application_text": "Die Village Home AI Bereitstellung nutzt QLoRA-feinabgestimmte Llama 3.1/3.2-Modelle, bei denen Steering-Vektoren zur Inferenzzeit angewendet werden können. Dies schafft eine zweischichtige Governance-Architektur.", + "read_link": "Arbeit lesen (HTML) →", + "pdf_link": "PDF herunterladen" + }, + "taonga": { + "heading": "Taonga-zentrierte Steering-Governance: Polyzentrische KI für indigene Datensouveränität", + "limitation_label": "Wichtige Einschränkung:", + "limitation_text": "Diese Arbeit wurde ohne Maori-Ko-Autorschaft verfasst und befindet sich derzeit in indigener Peer-Review. Sie schlägt eine technische Architektur vor, die auf veröffentlichter Wissenschaft zur indigenen Datensouveränität basiert, wurde aber noch nicht von Maori-Forschenden oder Gemeinschaften validiert.", + "intro": "Diese Arbeit schlägt eine polyzentrische Governance-Architektur vor, in der Iwi und Gemeinschaftsorganisationen gleichberechtigte Autorität neben technischen Systembetreibern behalten.", + "foundations_heading": "Theoretische Grundlagen", + "foundation_1": "CARE-Prinzipien (Carroll et al., 2020): Kollektiver Nutzen, Kontrollautorität, Verantwortung, Ethik", + "foundation_2": "Ostrom Polyzentrische Governance (1990): Mehrere überlappende Autoritätszentren statt einzelner hierarchischer Kontrolle", + "foundation_3": "Te ao Maori Konzepte: Kaitiakitanga (Vormundschaft), Rangatiratanga (Selbstbestimmung), Whakapapa (relationale Identität) als Architekturprinzipien", + "integration_heading": "Integration mit Tractatus", + "integration_text": "Die Arbeit erweitert den PluralisticDeliberationOrchestrator um Gemeinschafts-Stakeholder-Autoritäten mit Vetorecht über taonga-klassifizierte Daten.", + "read_link": "Entwurf lesen (HTML) →", + "pdf_link": "PDF herunterladen" + }, + "home_ai": { + "heading": "Home AI: Souveräne Governance-Forschungsplattform", + "intro": "Home AI stellt einen bedeutenden Forschungsmeilenstein dar: vollständige Tractatus-Governance eingebettet in eine lokal trainierte, souveräne Sprachmodell-Inferenz-Pipeline.", + "architecture_heading": "Zwei-Modell-Architektur", + "arch_1": "Schnelles Modell (Llama 3.2 3B): Antworten mit niedriger Latenz für Routineanfragen, mit Governance-Vorprüfung", + "arch_2": "Tiefes Modell (Llama 3.1 8B): Komplexes Reasoning mit vollständiger Governance-Pipeline", + "arch_3": "QLoRA-Feinabstimmung: Parametereffiziente Anpassung auf lokaler Hardware", + "research_heading": "Forschungsbedeutung", + "research_text": "Home AI eröffnet die Forschungsfrage der Governance-innerhalb-der-Trainingsschleife für gemeinschaftskontrollierte Modelle.", + "learn_more": "Mehr über Home AI erfahren →" } }, "footer": { @@ -408,7 +451,7 @@ ], "collaborate_text": "Arbeiten Sie mit uns zusammen: Wir freuen uns über Forscher, die diese Anwendung von Architekturprinzipien auf die KI-Governance untersuchen möchten. Wir können Zugang zu Audit-Protokollen, Framework-Code und Integrationsdokumentation für empirische Studien zur Verfügung stellen.", "contact_link": "Kontakt für Zusammenarbeit →", - "hf_space_link": "🤗 3.942 Audit-Entscheidungen auf HF erkunden →", + "audit_explorer_link": "171.800+ Audit-Entscheidungen erkunden →", "values_link": "Werte und Grundsätze →" } } diff --git a/public/locales/en/architecture.json b/public/locales/en/architecture.json index ae4350a2..124f04d1 100644 --- a/public/locales/en/architecture.json +++ b/public/locales/en/architecture.json @@ -290,8 +290,8 @@ "interactive_exploration": { "badge": "🔍 INTERACTIVE EXPLORATION", "heading": "See the Framework in Action", - "intro": "Explore 3,942 real governance decisions from production deployment. Filter by service, pressure level, and coordination patterns to understand how Deep Interlock operates in practice.", - "hf_cta": "🤗 Launch Interactive Explorer on Hugging Face", + "intro": "Explore 171,800+ real governance decisions from production deployment. Filter by service, pressure level, and coordination patterns to understand how Deep Interlock operates in practice.", + "audit_cta": "Launch Interactive Audit Explorer", "researcher_cta": "For Researchers →", "footer_note": "Apache 2.0 licensed • All data anonymized • No sign-up required" } diff --git a/public/locales/en/common.json b/public/locales/en/common.json index f12a1202..33d6c4d2 100644 --- a/public/locales/en/common.json +++ b/public/locales/en/common.json @@ -80,6 +80,37 @@ "submitting": "Sending..." }, "navbar": { + "research": "Research", + "publications": "Publications", + "publications_desc": "Browse research papers", + "for_researchers": "For Researchers", + "for_researchers_desc": "Open questions and collaboration", + "research_timeline": "Research Timeline", + "research_timeline_desc": "Evolution of the research", + "framework": "Framework", + "system_architecture": "System Architecture", + "system_architecture_desc": "Technical architecture overview", + "for_implementers": "For Implementers", + "for_implementers_desc": "Integration guide and code examples", + "village_case_study": "Village Case Study", + "village_case_study_desc": "Production deployment evidence", + "home_ai": "Home AI", + "home_ai_desc": "Sovereign locally-trained language model", + "agent_lightning": "Agent Lightning", + "agent_lightning_desc": "Performance optimisation integration", + "for_leaders": "For Leaders", + "for_leaders_desc": "Strategic overview and business case", + "blog": "Blog", + "about": "About", + "about_tractatus": "About Tractatus", + "values": "Values", + "docs": "Docs", + "github": "GitHub", + "koha": "Koha", + "navigation": "Navigation", + "resources": "Resources", + "documentation": "Documentation", + "mobile_feedback": "Feedback", "feedback": "Give Feedback", "feedback_desc": "Governed by Tractatus + AL", "company_heading": "Company", diff --git a/public/locales/en/homepage.json b/public/locales/en/homepage.json index 861e6893..ebb59ffc 100644 --- a/public/locales/en/homepage.json +++ b/public/locales/en/homepage.json @@ -61,7 +61,7 @@ "researcher_f1": "Formal foundations and proofs", "researcher_f2": "Failure mode analysis", "researcher_f3": "Open research questions", - "researcher_f4": "3,942 audit decisions on Hugging Face", + "researcher_f4": "171,800+ audit decisions analysed", "researcher_cta": "Explore research →", "implementer_title": "For Implementers", "implementer_subtitle": "Code and integration guides", diff --git a/public/locales/en/implementer.json b/public/locales/en/implementer.json index ab8616d2..d2fae46b 100644 --- a/public/locales/en/implementer.json +++ b/public/locales/en/implementer.json @@ -30,6 +30,9 @@ "services": "Services", "api": "API Reference", "patterns": "Integration Patterns", + "home_ai_arch": "Home AI", + "steering_vectors_impl": "Steering Vectors", + "taonga_registry": "Taonga Registry", "roadmap": "Roadmap" }, "how_it_works": { @@ -239,6 +242,65 @@ "sidecar_usecase": "Use Case:", "sidecar_usecase_value": "Kubernetes, containerized deployments" }, + "home_ai_arch": { + "heading": "Home AI: Two-Model Sovereign Architecture", + "intro": "Production deployment of Tractatus governance on locally-trained open-source models, demonstrating framework portability beyond Claude.", + "arch_title": "Two-Model Routing Architecture", + "arch_intro": "Home AI uses a dual-model design where queries are routed based on complexity and governance requirements. Both models run locally with full Tractatus governance in the inference pipeline.", + "fast_title": "Fast Model: Llama 3.2 3B", + "fast_1": "Purpose: Common queries with pre-filtered governance", + "fast_2": "Fine-tuning: QLoRA on domain-specific data", + "fast_3": "Governance: Lightweight boundary check before response", + "deep_title": "Deep Model: Llama 3.1 8B", + "deep_1": "Purpose: Complex reasoning with full governance pipeline", + "deep_2": "Fine-tuning: QLoRA with extended context governance", + "deep_3": "Governance: Full 6-service pipeline (BoundaryEnforcer through PDO)", + "routing_title": "Model Routing Logic", + "impl_title": "Implementation Details", + "stat_sovereignty": "100% Local", + "stat_sovereignty_desc": "Training data never leaves infrastructure", + "stat_governance": "6 Services", + "stat_governance_desc": "Full Tractatus governance in inference pipeline", + "stat_first": "First Non-Claude", + "stat_first_desc": "Validates Tractatus portability beyond Anthropic", + "status_note": "Status: Inference governance operational. Sovereign training pipeline installation in progress. Production deployment at Village Home Trust validates governance portability across model architectures.", + "cta": "Home AI Architecture Details →" + }, + "steering_impl": { + "heading": "Steering Vectors: Inference-Time Bias Correction", + "intro": "Techniques for correcting model behaviour at inference time without retraining, applicable to QLoRA-fine-tuned models like those in Home AI.", + "paper_ref": "Reference:", + "paper_title": "Steering Vectors and Mechanical Bias in Sovereign AI Systems (STO-RES-0009 v1.1, February 2026)", + "techniques_title": "Key Techniques for Implementers", + "caa_title": "Contrastive Activation Addition (CAA)", + "caa_desc": "Extract activation differences between contrasting prompts, then add scaled vectors during inference to steer model behaviour toward desired attributes.", + "repe_title": "Representation Engineering (RepE)", + "repe_desc": "Identify and manipulate internal representations of concepts (honesty, safety, fairness) through probing and directional manipulation of model activations.", + "fairsteer_title": "FairSteer", + "fairsteer_desc": "Identify bias-critical layers through probing classifiers, then apply targeted corrections at those specific layers rather than globally.", + "dso_title": "Direct Steering Optimisation (DSO)", + "dso_desc": "Optimise steering vectors using preference datasets (DPO-style), enabling training-free bias correction that learns from human preference data.", + "key_distinction": "Key Distinction: The paper distinguishes between mechanical bias (pre-reasoning, embedded in model weights — addressable by steering vectors) and reasoning bias (deliberative, arising during inference — requiring governance framework approaches like Tractatus).", + "research_link": "Full Research Details →", + "paper_link": "Read Paper (HTML)" + }, + "taonga_registry": { + "heading": "Taonga Registry: Governed Data Objects for Indigenous Sovereignty", + "intro": "Architectural patterns for treating steering packs and model configurations as governed data objects with provenance tracking and community-controlled access.", + "status_title": "Conceptual Architecture — In Peer Review", + "status_text": "Based on STO-RES-0010 v0.1 DRAFT, currently in indigenous peer review. Written without Maori co-authorship — presented as a starting point for collaboration. Not ready for production implementation.", + "concepts_title": "Key Architectural Concepts", + "packs_title": "Steering Packs as Governed Objects", + "packs_desc": "Model configurations, fine-tuning data, and steering vectors are treated as taonga (treasures) with full provenance tracking, access control, and community consent requirements.", + "provenance_title": "Provenance & Withdrawal Rights", + "provenance_desc": "Every data object tracks its full lineage. Communities retain the right to withdraw their data and configurations, with architectural enforcement of removal propagation.", + "polycentric_title": "Polycentric Governance Integration", + "polycentric_desc": "Extends PluralisticDeliberationOrchestrator with iwi/community as co-equal governance authorities. Access decisions require multi-party consensus.", + "care_title": "CARE Principles Implementation", + "care_desc": "Architecturally enforces Collective benefit, Authority to control, Responsibility, and Ethics for indigenous data governance within the Tractatus framework.", + "research_link": "Research Context →", + "paper_link": "Read Draft Paper (HTML)" + }, "roadmap": { "heading": "Development Roadmap & Collaboration", "intro": "Tractatus is an active research framework. We welcome collaboration on priority development areas.", @@ -246,10 +308,11 @@ "priority_desc": "These initiatives represent high-impact opportunities for framework enhancement. Technical contributors, researchers, and organizations are encouraged to engage.", "multi_llm_icon": "🤖", "multi_llm_title": "Multi-LLM Support", - "multi_llm_status": "Status: Research Phase", - "multi_llm_desc": "Extend governance to GPT-4, Gemini, Llama, and local models. Requires adapting hook architecture to different LLM interfaces.", - "multi_llm_challenges": "Technical Challenges:", - "multi_llm_challenges_desc": "Provider-specific tool/function calling, rate limiting, context window differences", + "multi_llm_badge": "First Deployment Live", + "multi_llm_status": "Status: First Non-Claude Deployment Operational", + "multi_llm_desc": "Home AI deploys Tractatus governance on Llama 3.1 8B and Llama 3.2 3B via QLoRA fine-tuning — the first validated non-Claude deployment. Extends governance portability to open-source models with full 6-service pipeline.", + "multi_llm_challenges": "Next Steps:", + "multi_llm_challenges_desc": "GPT-4 and Gemini adapters, provider-specific tool/function calling, sovereign training pipeline completion", "bindings_icon": "📚", "bindings_title": "Language Bindings", "bindings_status": "Status: Community Interest", @@ -280,6 +343,16 @@ "extended_desc": "Cost monitoring, rate limiting, PII detection, adversarial prompt defense. Domain-specific services for regulated industries.", "extended_examples": "Examples:", "extended_examples_desc": "FinancialComplianceService, HealthcarePrivacyService, CostBudgetEnforcer", + "training_title": "Sovereign Training Pipeline", + "training_status": "Status: In Progress", + "training_desc": "Governance-inside-the-training-loop for community-controlled models. Extends Tractatus from inference-time governance to training-time governance, ensuring data sovereignty from fine-tuning through deployment.", + "training_deliverables": "Deliverables:", + "training_deliverables_desc": "QLoRA training with governance constraints, data provenance tracking, community consent pipeline", + "taonga_title": "Taonga Data Registry", + "taonga_status": "Status: Conceptual / In Peer Review", + "taonga_desc": "Governed data objects with provenance tracking, withdrawal rights, and polycentric community governance. Extends PluralisticDeliberationOrchestrator for indigenous data sovereignty.", + "taonga_reference": "Reference:", + "taonga_reference_desc": "STO-RES-0010 v0.1 DRAFT — in indigenous peer review", "involved_title": "Get Involved", "involved_desc": "Tractatus is Apache 2.0 licensed research. We welcome contributions, pilot implementations, and collaborative research partnerships.", "contributors_title": "👨‍💻 Technical Contributors", diff --git a/public/locales/en/leader.json b/public/locales/en/leader.json index 64efe538..09d8fd65 100644 --- a/public/locales/en/leader.json +++ b/public/locales/en/leader.json @@ -4,7 +4,7 @@ "description": "Structural AI governance for organisations deploying LLM systems at scale. Research framework addressing architectural gaps in AI safety." }, "header": { - "badge": "Research Framework • Early Development", + "badge": "Research Framework • Production Validated", "title": "Tractatus: Architectural Governance for LLM Systems", "subtitle": "Architectural governance for organizations where AI governance failure triggers regulatory consequences. If your deployment is low-risk, architectural enforcement is likely unnecessary." }, @@ -88,8 +88,8 @@ }, "development_status": { "heading": "Development Status", - "warning_title": "Early-Stage Research Framework", - "warning_text": "Tractatus is a proof-of-concept developed over six months in a single project context (this website). It demonstrates architectural patterns for AI governance but has not undergone independent validation, red-team testing, or multi-organisation deployment.", + "warning_title": "Production-Validated Research Framework", + "warning_text": "Tractatus has been in active development for 11+ months (April 2025 to present) with production deployment at Village Home Trust, sovereign language model governance through Home AI, and over 171,800 audit decisions recorded. Independent validation and red-team testing remain outstanding research needs.", "validation_title": "Validated vs. Not Validated", "validated_label": "Validated:", "validated_text": "Framework successfully governs Claude Code in development workflows. User reports order-of-magnitude improvement in productivity for non-technical operators building production systems.", @@ -98,6 +98,34 @@ "limitation_label": "Known Limitation:", "limitation_text": "Framework can be bypassed if AI simply chooses not to use governance tools. Voluntary invocation remains a structural weakness requiring external enforcement mechanisms." }, + "home_ai": { + "heading": "Sovereign AI: Governance Embedded in Locally-Trained Models", + "intro": "Home AI demonstrates what it means to have governance embedded directly in locally-trained language models — not as an external compliance layer, but as part of the model serving architecture itself.", + "architecture_title": "Two-Model Architecture", + "arch_fast": "Fast model (3B parameters): Routine queries with governance pre-screening", + "arch_deep": "Deep model (8B parameters): Complex reasoning with full governance pipeline", + "arch_local": "Fully local: Training data never leaves the infrastructure", + "strategic_title": "Strategic Value", + "strat_sovereignty": "Data sovereignty: No cloud dependency for model training or inference", + "strat_governance": "Governance by design: Constraints are architectural, not retroactive compliance", + "strat_regulatory": "Regulatory positioning: Structurally stronger than bolt-on governance approaches", + "status": "Current status: Inference governance operational. Training pipeline installation in progress. First non-Claude deployment surface for Tractatus governance.", + "cta": "Learn about Home AI →" + }, + "taonga": { + "heading": "Polycentric Governance for Indigenous Data Sovereignty", + "intro": "For organisations with indigenous stakeholder obligations or multi-jurisdictional operations, Tractatus is developing a polycentric governance architecture where communities maintain architectural co-governance — not just consultation rights, but structural authority over how their data is used.", + "status_label": "Status:", + "status_text": "Draft paper (STO-RES-0010 v0.1) in indigenous peer review. Written without Maori co-authorship — presented transparently as a starting point for collaboration. This approach requires further peer review before implementation.", + "relevance": "Relevant for: Organisations operating in Aotearoa New Zealand, Australia, Canada, or other jurisdictions with indigenous data sovereignty obligations. Also applicable to any multi-stakeholder governance context where different parties require different levels of control over shared AI systems.", + "research_link": "Research details →", + "paper_link": "Read the draft paper" + }, + "steering_vectors": { + "heading": "Inference-Time Bias Correction (Steering Vectors)", + "summary": "New research (STO-RES-0009, published February 2026) demonstrates techniques for correcting bias at inference time without model retraining. For organisations concerned about bias in deployed AI systems, steering vectors offer the ability to respond to bias concerns without model downtime — corrections are applied as mathematical adjustments during inference, not through expensive retraining cycles.", + "link": "Technical details on the researcher page →" + }, "eu_ai_act": { "heading": "EU AI Act Considerations", "article_14_title": "Regulation 2024/1689, Article 14: Human Oversight", diff --git a/public/locales/en/researcher.json b/public/locales/en/researcher.json index 44b9836f..5da68a20 100644 --- a/public/locales/en/researcher.json +++ b/public/locales/en/researcher.json @@ -25,7 +25,7 @@ "research_context": { "heading": "Research Context & Scope", "development_note": "Development Context", - "development_text": "Tractatus was developed over six months (April–October 2025) in progressive stages that evolved into a live demonstration of its capabilities in the form of a single-project context (https://agenticgovernance.digital). Observations derive from direct engagement with Claude Code (Anthropic's Sonnet 4.5 model) across approximately 500 development sessions. This is exploratory research, not controlled study.", + "development_text": "Tractatus has been developed from April 2025 and is now in active production (11+ months). What began as a single-project demonstration has expanded to include production deployment at Village Home Trust and sovereign language model governance through Home AI. Observations derive from direct engagement with Claude Code (Anthropic Claude models, Sonnet 4.5 through Opus 4.6) across over 1,000 development sessions. This is exploratory research, not controlled study.", "paragraph_1": "Aligning advanced AI with human values is among the most consequential challenges we face. As capability growth accelerates under big tech momentum, we confront a categorical imperative: preserve human agency over values decisions, or risk ceding control entirely.", "paragraph_2": "The framework emerged from practical necessity. During development, we observed recurring patterns where AI systems would override explicit instructions, drift from established values constraints, or silently degrade quality under context pressure. Traditional governance approaches (policy documents, ethical guidelines, prompt engineering) proved insufficient to prevent these failures.", "paragraph_3": "Instead of hoping AI systems \"behave correctly,\" Tractatus proposes structural constraints where certain decision types require human judgment. These architectural boundaries can adapt to individual, organizational, and societal norms—creating a foundation for bounded AI operation that may scale more safely with capability growth.", @@ -126,8 +126,8 @@ }, "resources": { "heading": "Research Documentation", - "hf_space_title": "Interactive Audit Log Explorer (Hugging Face)", - "hf_space_subtitle": "Explore 3,942 governance decisions from production deployment", + "audit_explorer_title": "Interactive Audit Analytics Dashboard", + "audit_explorer_subtitle": "Explore 171,800+ governance decisions from production deployment", "doc_1_title": "Organisational Theory Foundations", "doc_2_title": "Cultural DNA Rules", "doc_3_title": "Case Studies: Real-World LLM Failure Modes", @@ -144,19 +144,20 @@ "section_3_heading": "AI Governance & Technical Context", "intellectual_lineage_label": "Note on Intellectual Lineage:", "intellectual_lineage_text": "The framework's central concern—human values persistence in AI-augmented organizational contexts—derives from moral philosophy rather than management science. The PluralisticDeliberationOrchestrator represents the primary research focus, embodying Weil's concept of attention to plural human needs and Berlin's recognition of incommensurable values.", - "future_development_text": "Berlin and Weil will be integral to further development of the deliberation component—their work provides the philosophical foundation for understanding how to preserve human agency over values decisions as AI capabilities accelerate. Traditional organizational theory (Weber, Taylor) addresses authority through hierarchy; post-AI organizational contexts require authority through appropriate deliberative process across stakeholder perspectives. Framework development documentation (incident reports, session logs) maintained in project repository but not publicly released pending peer review." + "future_development_text": "Berlin and Weil will be integral to further development of the deliberation component—their work provides the philosophical foundation for understanding how to preserve human agency over values decisions as AI capabilities accelerate. Traditional organizational theory (Weber, Taylor) addresses authority through hierarchy; post-AI organizational contexts require authority through appropriate deliberative process across stakeholder perspectives. Framework development documentation (incident reports, session logs) maintained in project repository but not publicly released pending peer review.", + "section_4_heading": "Indigenous Data Sovereignty & Polycentric Governance" }, "limitations": { "heading": "Limitations & Future Research Directions", "title": "Known Limitations & Research Gaps", - "validated_heading": "What We've Validated (October 2025)", - "validated_intro": "After 6 months of development and ~500 Claude Code sessions, we have grounded evidence for:", + "validated_heading": "What We've Validated (February 2026)", + "validated_intro": "After 11+ months of development, 1,000+ Claude Code sessions, and production deployment at Village Home Trust, we have grounded evidence for:", "validated_1_title": "✅ Architectural Blocking Mechanisms Functional", "validated_1_item1": "BoundaryEnforcer successfully blocks values decisions before execution", "validated_1_item2": "Pre-commit hooks prevent inst_017 violations (absolute assurance terms)", - "validated_1_item3": "1,130+ audit decisions recorded in MongoDB (tractatus_dev.audit_log)", - "validated_2_title": "✅ Instruction Persistence Works in Single-Session Context", - "validated_2_item1": "InstructionPersistenceClassifier maintains 62 active instructions (STRATEGIC: 22, SYSTEM: 21)", + "validated_1_item3": "171,800+ audit decisions recorded in MongoDB (tractatus_dev.audit_log)", + "validated_2_title": "✅ Instruction Persistence Works Across Sessions", + "validated_2_item1": "InstructionPersistenceClassifier maintains 68 active instructions (STRATEGIC: 27, SYSTEM: 21, OPERATIONAL: 18, TACTICAL: 2)", "validated_2_item2": "Pattern bias detection prevents AI from overriding explicit organizational directives", "validated_2_item3": "Temporal scopes (STRATEGIC/OPERATIONAL/TACTICAL) enforced successfully", "validated_3_title": "✅ Audit Trails Capture Governance Decisions", @@ -167,10 +168,14 @@ "validated_4_item1": "Real-time pressure scores calculated (token usage, message count, complexity)", "validated_4_item2": "Checkpoint triggers at 50k, 100k, 150k tokens", "validated_4_item3": "Framework fade detection alerts (5/6 components stale = warning)", - "validated_5_title": "✅ Single-Project Governance Successful", - "validated_5_item1": "Framework governed development of agenticgovernance.digital", - "validated_5_item2": "Cultural DNA rules (inst_085-089) enforced through pre-commit hooks", - "validated_5_item3": "Zero credential exposures (defense-in-depth: 5 layers verified)", + "validated_5_title": "✅ Multi-Deployment Governance Successful", + "validated_5_item1": "Framework governs agenticgovernance.digital (11+ months continuous operation)", + "validated_5_item2": "Village Home Trust production deployment: zero governance violations", + "validated_5_item3": "Home AI sovereign inference governance: operational", + "validated_5_item4": "Cultural DNA rules (inst_085-089) enforced through pre-commit hooks (4+ months operational)", + "validated_5_item5": "Phase 5 integration: 100% complete (all 6 services, 203/203 tests passing)", + "validated_5_item6": "Multilingual support: EN, DE, FR, Te Reo Maori", + "validated_5_item7": "Zero credential exposures (defense-in-depth: 5 layers verified)", "not_validated_heading": "What We Have NOT Validated", "not_validated_intro": "Honest disclosure of research gaps where we lack evidence:", "not_validated_1_title": "❌ Multi-Organization Deployments", @@ -178,21 +183,21 @@ "not_validated_1_item2": "Unknown: How framework performs across different organizations, domains, technical stacks", "not_validated_1_item3": "Research need: Pilot studies across industries (healthcare, finance, government)", "not_validated_2_title": "❌ Adversarial Robustness", - "not_validated_2_item1": "Validated: Normal development workflows (~500 sessions)", + "not_validated_2_item1": "Validated: Normal development workflows (1,000+ sessions)", "not_validated_2_item2": "Unknown: Resistance to deliberate bypass attempts, jailbreak prompts, adversarial testing", "not_validated_2_item3": "Research need: Red-team evaluation by security researchers", - "not_validated_3_title": "❌ Cross-Platform Consistency", - "not_validated_3_item1": "Validated: Claude Code (Anthropic Sonnet 4.5) only", - "not_validated_3_item2": "Unknown: Generalizability to Copilot, GPT-4, AutoGPT, LangChain, CrewAI, open models", - "not_validated_3_item3": "Research need: Cross-platform validation studies", + "not_validated_3_title": "⚠️ Cross-Platform Consistency (Partial)", + "not_validated_3_item1": "Validated: Claude Code (Anthropic Claude, Sonnet 4.5 through Opus 4.6) and Home AI (Llama 3.1/3.2 via QLoRA)", + "not_validated_3_item2": "Unknown: Generalizability to Copilot, GPT-4, AutoGPT, LangChain, CrewAI, other open models", + "not_validated_3_item3": "Research need: Broader cross-platform validation studies beyond Claude and Llama families", "not_validated_4_title": "❌ Concurrent Session Architecture", "not_validated_4_item1": "Validated: Single session at a time", "not_validated_4_item2": "Unknown: Multi-developer concurrent use on shared codebases", "not_validated_4_item3": "Known limitation: MongoDB duplicate key errors, session state contamination, race conditions on .claude/instruction-history.json", "not_validated_4_item4": "Research need: Multi-tenant architecture design (session-specific state, file locking)", "not_validated_5_title": "❌ Rule Proliferation Impact", - "not_validated_5_item1": "Validated: 62 active instructions work in current context", - "not_validated_5_item2": "Unknown: Performance degradation as rule count grows (projected 30-50 instructions within 12 months)", + "not_validated_5_item1": "Validated: 68 active instructions work in current context (exceeded original projection)", + "not_validated_5_item2": "Unknown: Performance degradation ceiling as rule count continues to grow", "not_validated_5_item3": "Known concern: Transactional overhead (CrossReferenceValidator checks), context window pressure, cognitive load", "not_validated_5_item4": "Research need: Rule consolidation strategies, automated governance optimization", "not_validated_6_title": "❌ Regulatory Evidence Sufficiency", @@ -286,7 +291,7 @@ "offer_2_title": "Documentation", "offer_2_desc": "Architecture specifications, implementation patterns, governance rules", "offer_3_title": "Audit Data", - "offer_3_desc": "1,130+ governance decisions in MongoDB (anonymized exports available)", + "offer_3_desc": "171,800+ governance decisions in MongoDB (anonymized exports available)", "offer_4_title": "Deployment Support", "offer_4_desc": "Help setting up local or cloud instances for testing", "offer_5_title": "Coordination", @@ -298,6 +303,44 @@ "cannot_offer_2": "Dedicated engineering resources (capacity constraints)", "cannot_offer_3": "Assured publication venues (but we'll support submission efforts)", "inquiry_button": "Research Collaboration Inquiry" + }, + "steering_vectors": { + "heading": "Steering Vectors and Mechanical Bias in Sovereign AI", + "intro": "This paper introduces a critical distinction between mechanical bias (pre-reasoning distortions embedded in model weights and activations) and reasoning bias (deliberative errors in chain-of-thought processing). Traditional debiasing approaches conflate these categories, leading to interventions that address symptoms rather than causes.", + "techniques_heading": "Technique Survey", + "technique_1": "Contrastive Activation Addition (CAA): Inference-time bias correction by adding/subtracting activation vectors extracted from contrastive prompt pairs", + "technique_2": "Representation Engineering (RepE): Linear probes to identify and modify concept representations within model layers", + "technique_3": "FairSteer & DSO: Fairness-oriented steering through distributionally-robust optimization", + "technique_4": "Sparse Autoencoders: Mechanistic interpretability through decomposition of polysemantic neurons into monosemantic features", + "application_heading": "Application to Village Home AI", + "application_text": "The Village Home AI deployment uses QLoRA-fine-tuned Llama 3.1/3.2 models where steering vectors can be applied at inference time. This creates a two-layer governance architecture: Tractatus provides structural constraints on decision boundaries, while steering vectors address pre-reasoning mechanical biases within the model itself. Together, they represent governance that operates both outside and inside the model.", + "read_link": "Read Paper (HTML) →", + "pdf_link": "Download PDF" + }, + "taonga": { + "heading": "Taonga-Centred Steering Governance: Polycentric AI for Indigenous Data Sovereignty", + "limitation_label": "Important Limitation:", + "limitation_text": "This paper was written without Maori co-authorship and is currently undergoing indigenous peer review. It proposes a technical architecture informed by published indigenous data sovereignty scholarship (Carroll et al., Hudson et al., Kukutai & Taylor) but has not yet been validated by Maori researchers or communities. We present it transparently as a starting point for collaboration, not a finished framework.", + "intro": "This paper proposes a polycentric governance architecture where iwi and community organisations maintain co-equal authority alongside technical system operators. Rather than treating indigenous data sovereignty as a compliance requirement to be satisfied retroactively, the architecture embeds community governance rights structurally through taonga registries, steering packs with provenance tracking, and withdrawal rights.", + "foundations_heading": "Theoretical Foundations", + "foundation_1": "CARE Principles (Carroll et al., 2020): Collective Benefit, Authority to Control, Responsibility, Ethics — applied to AI governance data flows", + "foundation_2": "Ostrom Polycentric Governance (1990): Multiple overlapping authority centres rather than single hierarchical control", + "foundation_3": "Te ao Maori concepts: Kaitiakitanga (guardianship), rangatiratanga (self-determination), whakapapa (relational identity) as architectural principles, not just metadata labels", + "integration_heading": "Integration with Tractatus", + "integration_text": "The paper extends the PluralisticDeliberationOrchestrator to support community stakeholder authorities with veto rights over taonga-classified data. Steering packs become governed data objects with full provenance tracking, access control, and the right of withdrawal — communities can revoke access to cultural knowledge at any time, and the system architecturally enforces that revocation.", + "read_link": "Read Draft (HTML) →", + "pdf_link": "Download PDF" + }, + "home_ai": { + "heading": "Home AI: Sovereign Governance Research Platform", + "intro": "Home AI represents a significant research milestone: full Tractatus governance embedded in a locally-trained, sovereign language model inference pipeline. This is the first deployment where governance operates inside the model serving layer rather than alongside an external API.", + "architecture_heading": "Two-Model Architecture", + "arch_1": "Fast model (Llama 3.2 3B): Low-latency responses for routine queries, with governance pre-screening", + "arch_2": "Deep model (Llama 3.1 8B): Complex reasoning with full governance pipeline, including BoundaryEnforcer and PluralisticDeliberationOrchestrator", + "arch_3": "QLoRA fine-tuning: Parameter-efficient adaptation on local hardware, enabling community-specific model customisation without cloud dependency", + "research_heading": "Research Significance", + "research_text": "Home AI opens the research question of governance-inside-the-training-loop for community-controlled models. Training data never leaves the local infrastructure; governance rules shape model behaviour through both fine-tuning data curation and inference-time constraints. This creates a fundamentally different governance surface than API-mediated approaches.", + "learn_more": "Learn more about Home AI →" } }, "modal": { @@ -391,7 +434,7 @@ ], "collaborate_text": "Collaborate with us: We welcome researchers interested in studying this application of architectural principles to AI governance. We can provide audit log access, framework code, and integration documentation for empirical study.", "contact_link": "Contact for Collaboration →", - "hf_space_link": "🤗 Explore 3,942 Audit Decisions on HF →", + "audit_explorer_link": "Explore 171,800+ Audit Decisions →", "values_link": "Values & Principles →" } } diff --git a/public/locales/fr/architecture.json b/public/locales/fr/architecture.json index 568a1379..c082632a 100644 --- a/public/locales/fr/architecture.json +++ b/public/locales/fr/architecture.json @@ -290,8 +290,8 @@ "interactive_exploration": { "badge": "🔍 EXPLORATION INTERACTIVE", "heading": "Voir le framework en action", - "intro": "Explorez 3 942 décisions de gouvernance réelles issues du déploiement en production. Filtrez par service, niveau de pression et modèles de coordination pour comprendre comment Deep Interlock fonctionne en pratique.", - "hf_cta": "🤗 Lancer l'explorateur interactif sur Hugging Face", + "intro": "Explorez 171 800+ décisions de gouvernance réelles issues du déploiement en production. Filtrez par service, niveau de pression et modèles de coordination pour comprendre comment Deep Interlock fonctionne en pratique.", + "audit_cta": "Lancer l'explorateur d'audit interactif", "researcher_cta": "Pour les chercheurs →", "footer_note": "Licence Apache 2.0 • Toutes les données anonymisées • Aucune inscription requise" } diff --git a/public/locales/fr/common.json b/public/locales/fr/common.json index d40b54ee..c47903fa 100644 --- a/public/locales/fr/common.json +++ b/public/locales/fr/common.json @@ -80,6 +80,37 @@ "submitting": "Envoi en cours..." }, "navbar": { + "research": "Recherche", + "publications": "Publications", + "publications_desc": "Parcourir les articles de recherche", + "for_researchers": "Pour les chercheurs", + "for_researchers_desc": "Questions ouvertes et collaboration", + "research_timeline": "Chronologie de la recherche", + "research_timeline_desc": "Évolution de la recherche", + "framework": "Framework", + "system_architecture": "Architecture du système", + "system_architecture_desc": "Vue d'ensemble de l'architecture technique", + "for_implementers": "Pour les développeurs", + "for_implementers_desc": "Guide d'intégration et exemples de code", + "village_case_study": "Étude de cas Village", + "village_case_study_desc": "Preuves de déploiement en production", + "home_ai": "Home AI", + "home_ai_desc": "Modèle de langue souverain entraîné localement", + "agent_lightning": "Agent Lightning", + "agent_lightning_desc": "Intégration d'optimisation des performances", + "for_leaders": "Pour les dirigeants", + "for_leaders_desc": "Vue stratégique et analyse de rentabilité", + "blog": "Blog", + "about": "À propos", + "about_tractatus": "À propos de Tractatus", + "values": "Valeurs", + "docs": "Documentation", + "github": "GitHub", + "koha": "Koha", + "navigation": "Navigation", + "resources": "Ressources", + "documentation": "Documentation", + "mobile_feedback": "Commentaires", "feedback": "Donner son avis", "feedback_desc": "Régie par le Tractatus AL", "company_heading": "Entreprise", diff --git a/public/locales/fr/homepage.json b/public/locales/fr/homepage.json index ec5f806b..57ea8667 100644 --- a/public/locales/fr/homepage.json +++ b/public/locales/fr/homepage.json @@ -61,7 +61,7 @@ "researcher_f1": "Fondements formels et preuves", "researcher_f2": "Analyse des modes de défaillance", "researcher_f3": "Questions de recherche ouvertes", - "researcher_f4": "3 942 décisions d’audit sur Hugging Face", + "researcher_f4": "171 800+ décisions d'audit analysées", "researcher_cta": "Explorer la recherche →", "implementer_title": "Pour les implémenteurs", "implementer_subtitle": "Code et guides d’intégration", diff --git a/public/locales/fr/implementer.json b/public/locales/fr/implementer.json index cc3c506a..9c08ae73 100644 --- a/public/locales/fr/implementer.json +++ b/public/locales/fr/implementer.json @@ -146,6 +146,9 @@ "services": "Services", "api": "Référence API", "patterns": "Modèles d'intégration", + "home_ai_arch": "Home AI", + "steering_vectors_impl": "Vecteurs de guidage", + "taonga_registry": "Registre Taonga", "roadmap": "Feuille de route" }, "how_it_works": { @@ -334,6 +337,65 @@ "sidecar_usecase": "Cas d'utilisation :", "sidecar_usecase_value": "Kubernetes, déploiements conteneurisés" }, + "home_ai_arch": { + "heading": "Home AI : Architecture souveraine à deux modèles", + "intro": "Déploiement en production de la gouvernance Tractatus sur des modèles open source entraînés localement, démontrant la portabilité du framework au-delà de Claude.", + "arch_title": "Architecture de routage à deux modèles", + "arch_intro": "Home AI utilise une conception à double modèle où les requêtes sont routées en fonction de la complexité et des exigences de gouvernance. Les deux modèles fonctionnent localement avec la gouvernance Tractatus complète dans le pipeline d'inférence.", + "fast_title": "Modèle rapide : Llama 3.2 3B", + "fast_1": "Objectif : Requêtes courantes avec pré-filtrage de gouvernance", + "fast_2": "Affinage : QLoRA sur des données spécifiques au domaine", + "fast_3": "Gouvernance : Vérification légère des limites avant réponse", + "deep_title": "Modèle approfondi : Llama 3.1 8B", + "deep_1": "Objectif : Raisonnement complexe avec pipeline de gouvernance complet", + "deep_2": "Affinage : QLoRA avec gouvernance de contexte étendu", + "deep_3": "Gouvernance : Pipeline complet à 6 services (BoundaryEnforcer à PDO)", + "routing_title": "Logique de routage des modèles", + "impl_title": "Détails d'implémentation", + "stat_sovereignty": "100% Local", + "stat_sovereignty_desc": "Les données d'entraînement ne quittent jamais l'infrastructure", + "stat_governance": "6 Services", + "stat_governance_desc": "Gouvernance Tractatus complète dans le pipeline d'inférence", + "stat_first": "Premier non-Claude", + "stat_first_desc": "Valide la portabilité de Tractatus au-delà d'Anthropic", + "status_note": "Statut : Gouvernance d'inférence opérationnelle. Installation du pipeline d'entraînement souverain en cours. Le déploiement en production chez Village Home Trust valide la portabilité de la gouvernance entre architectures de modèles.", + "cta": "Détails de l'architecture Home AI →" + }, + "steering_impl": { + "heading": "Vecteurs de guidage : Correction de biais au moment de l'inférence", + "intro": "Techniques de correction du comportement du modèle au moment de l'inférence sans ré-entraînement, applicables aux modèles affinés par QLoRA comme ceux de Home AI.", + "paper_ref": "Référence :", + "paper_title": "Vecteurs de guidage et biais mécanique dans les systèmes d'IA souverains (STO-RES-0009 v1.1, février 2026)", + "techniques_title": "Techniques clés pour les implémenteurs", + "caa_title": "Contrastive Activation Addition (CAA)", + "caa_desc": "Extrayez les différences d'activation entre des prompts contrastés, puis ajoutez des vecteurs mis à l'échelle pendant l'inférence pour orienter le comportement du modèle vers les attributs souhaités.", + "repe_title": "Representation Engineering (RepE)", + "repe_desc": "Identifiez et manipulez les représentations internes de concepts (honnêteté, sécurité, équité) par le sondage et la manipulation directionnelle des activations du modèle.", + "fairsteer_title": "FairSteer", + "fairsteer_desc": "Identifiez les couches critiques pour le biais via des classificateurs de sondage, puis appliquez des corrections ciblées sur ces couches spécifiques plutôt que globalement.", + "dso_title": "Direct Steering Optimisation (DSO)", + "dso_desc": "Optimisez les vecteurs de guidage à l'aide de jeux de données de préférence (style DPO), permettant une correction de biais sans entraînement qui apprend des données de préférence humaines.", + "key_distinction": "Distinction clé : L'article distingue entre le biais mécanique (pré-raisonnement, intégré dans les poids du modèle — adressable par les vecteurs de guidage) et le biais de raisonnement (délibératif, survenant pendant l'inférence — nécessitant des approches de framework de gouvernance comme Tractatus).", + "research_link": "Détails complets de la recherche →", + "paper_link": "Lire l'article (HTML)" + }, + "taonga_registry": { + "heading": "Registre Taonga : Objets de données gouvernés pour la souveraineté autochtone", + "intro": "Modèles architecturaux pour le traitement des packs de guidage et des configurations de modèles comme des objets de données gouvernés avec suivi de provenance et accès contrôlé par la communauté.", + "status_title": "Architecture conceptuelle — En évaluation par les pairs", + "status_text": "Basé sur STO-RES-0010 v0.1 BROUILLON, actuellement en évaluation par les pairs autochtones. Rédigé sans co-auteur maori — présenté comme point de départ pour la collaboration. Non prêt pour la production.", + "concepts_title": "Concepts architecturaux clés", + "packs_title": "Packs de guidage comme objets gouvernés", + "packs_desc": "Les configurations de modèles, les données d'affinage et les vecteurs de guidage sont traités comme des taonga (trésors) avec suivi complet de provenance, contrôle d'accès et exigences de consentement communautaire.", + "provenance_title": "Provenance et droits de retrait", + "provenance_desc": "Chaque objet de données suit sa lignée complète. Les communautés conservent le droit de retirer leurs données et configurations, avec application architecturale de la propagation du retrait.", + "polycentric_title": "Intégration de la gouvernance polycentrique", + "polycentric_desc": "Étend le PluralisticDeliberationOrchestrator avec iwi/communauté comme autorités de gouvernance co-égales. Les décisions d'accès nécessitent un consensus multipartite.", + "care_title": "Implémentation des principes CARE", + "care_desc": "Applique architecturalement le bénéfice collectif, l'autorité de contrôle, la responsabilité et l'éthique pour la gouvernance des données autochtones au sein du framework Tractatus.", + "research_link": "Contexte de recherche →", + "paper_link": "Lire le brouillon (HTML)" + }, "roadmap": { "heading": "Feuille de route et collaboration en matière de développement", "intro": "Tractatus est un cadre de recherche actif. Nous sommes ouverts à toute collaboration dans les domaines de développement prioritaires.", @@ -341,10 +403,11 @@ "priority_desc": "Ces initiatives représentent des opportunités à fort impact pour l'amélioration du cadre. Les contributeurs techniques, les chercheurs et les organisations sont encouragés à s'engager.", "multi_llm_icon": "🤖", "multi_llm_title": "Support multi-LLM", - "multi_llm_status": "Statut : Phase de recherche", - "multi_llm_desc": "Étendre la gouvernance aux modèles GPT-4, Gemini, Llama et locaux. Nécessité d'adapter l'architecture des crochets aux différentes interfaces LLM.", - "multi_llm_challenges": "Défis techniques :", - "multi_llm_challenges_desc": "Appel d'outils/fonctions spécifiques au prestataire, limitation des tarifs, différences de fenêtres contextuelles", + "multi_llm_status": "Statut : Premier déploiement non-Claude opérationnel", + "multi_llm_desc": "Home AI déploie la gouvernance Tractatus sur Llama 3.1 8B et Llama 3.2 3B via l'affinage QLoRA — le premier déploiement non-Claude validé. Étend la portabilité de la gouvernance aux modèles open source avec le pipeline complet à 6 services.", + "multi_llm_badge": "Premier déploiement live", + "multi_llm_challenges": "Prochaines étapes :", + "multi_llm_challenges_desc": "Adaptateurs GPT-4 et Gemini, appel d'outils/fonctions spécifiques au prestataire, achèvement du pipeline d'entraînement souverain", "bindings_icon": "📚", "bindings_title": "Liaisons linguistiques", "bindings_status": "Statut : Intérêt communautaire", @@ -375,6 +438,16 @@ "extended_desc": "Contrôle des coûts, limitation des tarifs, détection des IPI, défense rapide contre les adversaires. Services spécifiques à un domaine pour les industries réglementées.", "extended_examples": "Exemples :", "extended_examples_desc": "FinancialComplianceService, HealthcarePrivacyService, CostBudgetEnforcer", + "training_title": "Pipeline d'entraînement souverain", + "training_status": "Statut : En cours", + "training_desc": "Gouvernance à l'intérieur de la boucle d'entraînement pour les modèles contrôlés par la communauté. Étend Tractatus de la gouvernance au moment de l'inférence à la gouvernance au moment de l'entraînement, assurant la souveraineté des données de l'affinage au déploiement.", + "training_deliverables": "Livrables :", + "training_deliverables_desc": "Entraînement QLoRA avec contraintes de gouvernance, suivi de provenance des données, pipeline de consentement communautaire", + "taonga_title": "Registre de données Taonga", + "taonga_status": "Statut : Conceptuel / En évaluation par les pairs", + "taonga_desc": "Objets de données gouvernés avec suivi de provenance, droits de retrait et gouvernance communautaire polycentrique. Étend le PluralisticDeliberationOrchestrator pour la souveraineté des données autochtones.", + "taonga_reference": "Référence :", + "taonga_reference_desc": "STO-RES-0010 v0.1 BROUILLON — en évaluation par les pairs autochtones", "involved_title": "S'impliquer", "involved_desc": "Tractatus est une recherche sous licence Apache 2.0. Les contributions, les mises en œuvre pilotes et les partenariats de recherche collaborative sont les bienvenus.", "contributors_title": "👨‍💻 Contributeurs techniques", diff --git a/public/locales/fr/leader.json b/public/locales/fr/leader.json index 46632ce3..932bd1ec 100644 --- a/public/locales/fr/leader.json +++ b/public/locales/fr/leader.json @@ -4,7 +4,7 @@ "description": "Gouvernance structurelle de l'IA pour les organisations déployant des systèmes LLM à grande échelle. Cadre de recherche abordant les lacunes architecturales en matière de sécurité de l'IA." }, "header": { - "badge": "Cadre de Recherche • Développement Précoce", + "badge": "Framework de recherche • Validé en production", "title": "Tractatus : Gouvernance Architecturale pour les Systèmes LLM", "subtitle": "Gouvernance architecturale pour les organisations où l'échec de la gouvernance de l'IA entraîne des conséquences réglementaires. Si votre déploiement présente peu de risques, l'application de l'architecture n'est probablement pas nécessaire." }, @@ -88,8 +88,8 @@ }, "development_status": { "heading": "État du Développement", - "warning_title": "Cadre de Recherche en Phase Initiale", - "warning_text": "Tractatus est une preuve de concept développée sur six mois dans un contexte de projet unique (ce site web). Il démontre des modèles architecturaux pour la gouvernance de l'IA mais n'a pas subi de validation indépendante, de tests d'équipe rouge ou de déploiement multi-organisationnel.", + "warning_title": "Framework de recherche validé en production", + "warning_text": "Tractatus est en développement actif depuis plus de 11 mois (avril 2025 à aujourd'hui) avec un déploiement en production chez Village Home Trust, une gouvernance souveraine de modèle de langue via Home AI, et plus de 171 800 décisions d'audit enregistrées. La validation indépendante et les tests de type red-team restent des besoins de recherche en suspens.", "validation_title": "Validé vs Non Validé", "validated_label": "Validé :", "validated_text": "Le cadre régit avec succès le code Claude dans les flux de travail de développement. L'utilisateur signale une amélioration de l'ordre de grandeur de la productivité pour les opérateurs non techniques qui construisent des systèmes de production.", @@ -98,6 +98,34 @@ "limitation_label": "Limitation connue :", "limitation_text": "Le cadre peut être contourné si l'IA choisit simplement de ne pas utiliser les outils de gouvernance. L'invocation volontaire reste une faiblesse structurelle nécessitant des mécanismes d'application externes." }, + "home_ai": { + "heading": "IA souveraine : gouvernance intégrée dans des modèles entraînés localement", + "intro": "Home AI démontre ce que signifie intégrer la gouvernance directement dans des modèles de langue entraînés localement — non pas comme une couche de conformité externe, mais comme partie intégrante de l'architecture de service du modèle.", + "architecture_title": "Architecture à deux modèles", + "arch_fast": "Modèle rapide (3B paramètres) : Requêtes courantes avec pré-filtrage de gouvernance", + "arch_deep": "Modèle approfondi (8B paramètres) : Raisonnement complexe avec pipeline de gouvernance complet", + "arch_local": "Entièrement local : Les données d'entraînement ne quittent jamais l'infrastructure", + "strategic_title": "Valeur stratégique", + "strat_sovereignty": "Souveraineté des données : Aucune dépendance au cloud pour l'entraînement ou l'inférence", + "strat_governance": "Gouvernance par conception : Les contraintes sont architecturales, pas une conformité rétrospective", + "strat_regulatory": "Positionnement réglementaire : Structurellement plus fort que les approches de gouvernance ajoutées après coup", + "status": "Statut actuel : Gouvernance d'inférence opérationnelle. Installation du pipeline d'entraînement en cours. Première surface de déploiement non-Claude pour la gouvernance Tractatus.", + "cta": "En savoir plus sur Home AI →" + }, + "taonga": { + "heading": "Gouvernance polycentrique pour la souveraineté des données autochtones", + "intro": "Pour les organisations ayant des obligations envers des parties prenantes autochtones ou des opérations multi-juridictionnelles, Tractatus développe une architecture de gouvernance polycentrique où les communautés maintiennent une co-gouvernance architecturale.", + "status_label": "Statut :", + "status_text": "Article en brouillon (STO-RES-0010 v0.1) en évaluation par les pairs autochtones. Rédigé sans co-auteur maori — présenté de manière transparente comme point de départ pour la collaboration. Cette approche n'est pas prête pour la production.", + "relevance": "Pertinent pour : Les organisations opérant en Aotearoa Nouvelle-Zélande, Australie, Canada ou d'autres juridictions avec des obligations de souveraineté des données autochtones.", + "research_link": "Détails de la recherche →", + "paper_link": "Lire le brouillon" + }, + "steering_vectors": { + "heading": "Correction de biais au moment de l'inférence (vecteurs de guidage)", + "summary": "De nouvelles recherches (STO-RES-0009, publiées en février 2026) démontrent des techniques de correction de biais au moment de l'inférence sans ré-entraînement du modèle. Pour les organisations préoccupées par les biais dans les systèmes d'IA déployés, les vecteurs de guidage offrent la possibilité de répondre aux préoccupations de biais sans temps d'arrêt du modèle.", + "link": "Détails techniques sur la page chercheur →" + }, "eu_ai_act": { "heading": "Considérations du Règlement Européen sur l'IA", "article_14_title": "Règlement 2024/1689, Article 14 : Surveillance Humaine", diff --git a/public/locales/fr/researcher.json b/public/locales/fr/researcher.json index 4a28e35d..ff9255a1 100644 --- a/public/locales/fr/researcher.json +++ b/public/locales/fr/researcher.json @@ -12,7 +12,7 @@ "research_context": { "heading": "Contexte & Portée de la Recherche", "development_note": "Contexte de Développement", - "development_text": "Tractatus a été développé sur six mois (avril-octobre 2025) en phases progressives qui ont évolué en une démonstration en direct de ses capacités sous la forme d'un contexte de projet unique (https://agenticgovernance.digital). Les observations proviennent d'un engagement direct avec Claude Code (modèle Sonnet 4.5 d'Anthropic) sur environ 500 sessions de développement. Il s'agit de recherche exploratoire, pas d'étude contrôlée.", + "development_text": "Tractatus est développé depuis avril 2025 et est maintenant en production active (11+ mois). Ce qui a commencé comme une démonstration sur un projet unique s'est élargi pour inclure le déploiement en production chez Village Home Trust et la gouvernance souveraine de modèles linguistiques via Home AI. Les observations proviennent d'un engagement direct avec Claude Code (modèles Anthropic Claude, Sonnet 4.5 à Opus 4.6) sur plus de 1 000 sessions de développement. Il s'agit de recherche exploratoire, pas d'étude contrôlée.", "paragraph_1": "L'alignement de l'IA avancée sur les valeurs humaines est l'un des défis les plus importants auxquels nous sommes confrontés. Alors que la croissance des capacités s'accélère sous l'impulsion des grandes technologies, nous sommes confrontés à un impératif catégorique : préserver le pouvoir de l'homme sur les décisions relatives aux valeurs, ou risquer de céder complètement le contrôle.", "paragraph_2": "Le cadre est né d'une nécessité pratique. Au cours du développement, nous avons observé des schémas récurrents dans lesquels les systèmes d'IA passaient outre les instructions explicites, s'écartaient des contraintes de valeurs établies ou dégradaient silencieusement la qualité sous la pression du contexte. Les approches traditionnelles en matière de gouvernance (documents de politique générale, lignes directrices éthiques, ingénierie rapide) se sont révélées insuffisantes pour prévenir ces défaillances.", "paragraph_3": "Au lieu d'espérer que les systèmes d'IA \"se comportent correctement\", Tractatus propose des contraintes structurelles où certains types de décisions requièrent un jugement humain. Ces limites architecturales peuvent s'adapter aux normes individuelles, organisationnelles et sociétales, créant ainsi une base pour un fonctionnement limité de l'IA qui peut s'adapter de manière plus sûre à la croissance des capacités.", @@ -113,8 +113,8 @@ }, "resources": { "heading": "Documentation de Recherche", - "hf_space_title": "Explorateur interactif de journaux d'audit (Hugging Face)", - "hf_space_subtitle": "Explorez 3 942 décisions de gouvernance issues du déploiement en production", + "audit_explorer_title": "Tableau de bord interactif d'analyse des audits", + "audit_explorer_subtitle": "Explorez 171 800+ décisions de gouvernance issues du déploiement en production", "doc_1_title": "Fondements de la théorie des organisations", "doc_2_title": "Règles de l'ADN culturel", "doc_3_title": "Études de cas : Modes de défaillance du LLM dans le monde réel", @@ -132,7 +132,7 @@ "limitation_3_title": "3. Pas de test contradictoire", "limitation_3_desc": "Le cadre n'a pas fait l'objet d'une évaluation par l'équipe rouge, d'un test de jailbreak ou d'une évaluation rapide par des adversaires. Toutes les observations proviennent d'un processus de développement normal, et non de tentatives de contournement délibérées.", "limitation_4_title": "4. Spécificité de la plate-forme", - "limitation_4_desc": "Observations et interventions validées avec le code Claude (Anthropic Sonnet 4.5) uniquement. La généralisation à d'autres systèmes LLM (Copilot, GPT-4, agents personnalisés) reste une hypothèse non validée.", + "limitation_4_desc": "Observations et interventions validées avec Claude Code (Anthropic Claude, Sonnet 4.5 à Opus 4.6) et Home AI (Llama 3.1/3.2 via QLoRA). La généralisation à d'autres systèmes LLM (Copilot, GPT-4, agents personnalisés) reste partiellement validée.", "limitation_5_title": "5. Incertitude d'échelle", "limitation_5_desc": "Les caractéristiques de performance à l'échelle de l'entreprise (des milliers d'utilisateurs simultanés, des millions d'événements de gouvernance) sont totalement inconnues. La mise en œuvre actuelle est optimisée pour le contexte d'un seul utilisateur.", "future_research_title": "Besoins futurs en matière de recherche :", @@ -142,14 +142,14 @@ "future_research_4": "Évaluation de la cohérence entre plates-formes (Copilot, GPT-4, modèles ouverts)", "future_research_5": "Vérification formelle des propriétés d'application des limites", "future_research_6": "Étude longitudinale de l'efficacité du cadre au cours d'un déploiement prolongé", - "validated_heading": "Ce que nous avons validé (octobre 2025)", - "validated_intro": "Après 6 mois de développement et ~500 sessions de Claude Code, nous disposons de preuves concrètes :", + "validated_heading": "Ce que nous avons validé (février 2026)", + "validated_intro": "Après 11+ mois de développement, 1 000+ sessions Claude Code et un déploiement en production chez Village Home Trust, nous avons des preuves solides pour :", "validated_1_title": "✅ Mécanismes de blocage architecturaux Fonctionnels", "validated_1_item1": "BoundaryEnforcer bloque avec succès les décisions relatives aux valeurs avant leur exécution", "validated_1_item2": "Les hooks de pré-commission empêchent les violations d'inst_017 (termes d'assurance absolue)", - "validated_1_item3": "1 130+ décisions d'audit enregistrées dans MongoDB (tractatus_dev.audit_log)", - "validated_2_title": "la persistance des instructions fonctionne dans un contexte de session unique", - "validated_2_item1": "InstructionPersistenceClassifier maintient 62 instructions actives (STRATEGIC : 22, SYSTEM : 21)", + "validated_1_item3": "171 800+ décisions d'audit enregistrées dans MongoDB (tractatus_dev.audit_log)", + "validated_2_title": "✅ La persistance des instructions fonctionne entre les sessions", + "validated_2_item1": "InstructionPersistenceClassifier maintient 68 instructions actives (STRATEGIC: 27, SYSTEM: 21, OPERATIONAL: 18, TACTICAL: 2)", "validated_2_item2": "La détection des biais empêche l'IA d'outrepasser les directives explicites de l'organisation", "validated_2_item3": "Portée temporelle (STRATEGIQUE/OPERATIONNELLE/TACTIQUE) appliquée avec succès", "validated_3_title": "✅ Les pistes d'audit rendent compte des décisions de gouvernance", @@ -160,10 +160,14 @@ "validated_4_item1": "Calcul des scores de pression en temps réel (utilisation de jetons, nombre de messages, complexité)", "validated_4_item2": "Déclenchement des points de contrôle à 50k, 100k, 150k tokens", "validated_4_item3": "Alertes de détection de l'altération du cadre (5/6 composants altérés = avertissement)", - "validated_5_title": "✅ Une gouvernance de projet unique réussie", - "validated_5_item1": "Cadre régissant le développement de agenticgovernance.digital", - "validated_5_item2": "Règles culturelles de l'ADN (inst_085-089) appliquées par le biais de crochets de précommission (pre-commit)", - "validated_5_item3": "Exposition zéro des informations d'identification (défense en profondeur : 5 couches vérifiées)", + "validated_5_title": "✅ Gouvernance multi-déploiement réussie", + "validated_5_item1": "Le framework gouverne agenticgovernance.digital (11+ mois d'opération continue)", + "validated_5_item2": "Déploiement en production Village Home Trust : zéro violation de gouvernance", + "validated_5_item3": "Gouvernance d'inférence souveraine Home AI : opérationnelle", + "validated_5_item4": "Règles culturelles de l'ADN (inst_085-089) appliquées par le biais de crochets de précommission (4+ mois opérationnels)", + "validated_5_item5": "Intégration Phase 5 : 100% complète (les 6 services, 203/203 tests réussis)", + "validated_5_item6": "Support multilingue : EN, DE, FR, Te Reo Maori", + "validated_5_item7": "Exposition zéro des informations d'identification (défense en profondeur : 5 couches vérifiées)", "not_validated_heading": "Ce que nous n'avons PAS validé", "not_validated_intro": "La divulgation honnête des lacunes de la recherche lorsque nous manquons de preuves :", "not_validated_1_title": "déploiements multi-organisations", @@ -171,22 +175,22 @@ "not_validated_1_item2": "Inconnu : Comment le cadre fonctionne-t-il dans différentes organisations, différents domaines, différentes piles techniques ?", "not_validated_1_item3": "Besoins en matière de recherche : Études pilotes dans différents secteurs (soins de santé, finance, administration)", "not_validated_2_title": "❌ Robustesse adversariale", - "not_validated_2_item1": "Validé : Flux de développement normaux (~500 sessions)", + "not_validated_2_item1": "Validé : Flux de développement normaux (1 000+ sessions)", "not_validated_2_item2": "Inconnu : Résistance aux tentatives délibérées de contournement, aux invites de jailbreak, aux tests contradictoires", "not_validated_2_item3": "Besoin de recherche : Évaluation par une équipe de chercheurs en sécurité", - "not_validated_3_title": "❌ Cohérence multiplateforme", - "not_validated_3_item1": "Validé : Code Claude (Sonnet anthropique 4.5) uniquement", - "not_validated_3_item2": "Inconnu : Généralisabilité à Copilot, GPT-4, AutoGPT, LangChain, CrewAI, modèles ouverts", - "not_validated_3_item3": "Besoin de recherche : Études de validation multiplateforme", + "not_validated_3_title": "⚠️ Cohérence multiplateforme (Partielle)", + "not_validated_3_item1": "Validé : Claude Code (Anthropic Claude, Sonnet 4.5 à Opus 4.6) et Home AI (Llama 3.1/3.2 via QLoRA)", + "not_validated_3_item2": "Inconnu : Généralisabilité à Copilot, GPT-4, AutoGPT, LangChain, CrewAI, autres modèles ouverts", + "not_validated_3_item3": "Besoin de recherche : Études de validation multiplateforme plus larges au-delà des familles Claude et Llama", "not_validated_4_title": "architecture des sessions simultanées", "not_validated_4_item1": "Validé : Une seule session à la fois", "not_validated_4_item2": "Inconnu : Utilisation simultanée par plusieurs développeurs sur des bases de code partagées", "not_validated_4_item3": "Limitations connues : Erreurs de clé dupliquée MongoDB, contamination de l'état de la session, conditions de course sur .claude/instruction-history.json", "not_validated_4_item4": "Besoin de recherche : Conception d'une architecture multi-locataires (état spécifique à une session, verrouillage des fichiers)", "not_validated_5_title": "❌ Impact de la prolifération des règles", - "not_validated_5_item1": "Validé : 62 Les instructions actives fonctionnent dans le contexte actuel", - "not_validated_5_item2": "Inconnu : Dégradation des performances au fur et à mesure que le nombre de règles augmente (30 à 50 instructions prévues dans les 12 mois)", - "not_validated_5_item3": "Problèmes connus : Surcharge transactionnelle (vérifications du CrossReferenceValidator), pression de la fenêtre contextuelle, charge cognitive", + "not_validated_5_item1": "Validé : 68 instructions actives fonctionnent dans le contexte actuel (projection initiale dépassée)", + "not_validated_5_item2": "Inconnu : Plafond de dégradation des performances à mesure que le nombre de règles continue d'augmenter", + "not_validated_5_item3": "Préoccupation connue : Surcharge transactionnelle (vérifications du CrossReferenceValidator), pression de la fenêtre contextuelle, charge cognitive", "not_validated_5_item4": "Besoin de recherche : Stratégies de consolidation des règles, optimisation de la gouvernance automatisée", "not_validated_6_title": "❌ Suffisance des preuves réglementaires", "not_validated_6_item1": "Validé : Des pistes d'audit existent avec des balises de conformité (EU_AI_ACT_Article14, GDPR_Article22)", @@ -222,7 +226,8 @@ "section_3_heading": "Gouvernance de l'IA et contexte technique", "intellectual_lineage_label": "Note sur la lignée intellectuelle :", "intellectual_lineage_text": "La préoccupation centrale du cadre - la persistance des valeurs humaines dans les contextes organisationnels augmentés par l'IA - découle de la philosophie morale plutôt que de la science de la gestion. Le PluralisticDeliberationOrchestrator représente le principal axe de recherche, incarnant le concept d'attention aux besoins humains pluriels de Weil et la reconnaissance des valeurs incommensurables de Berlin.", - "future_development_text": "Berlin et Weil joueront un rôle essentiel dans le développement de la composante \"délibération\" : leurs travaux fournissent les fondements philosophiques permettant de comprendre comment préserver l'action humaine sur les décisions relatives aux valeurs à mesure que les capacités de l'IA s'accélèrent. La théorie organisationnelle traditionnelle (Weber, Taylor) traite de l'autorité par le biais de la hiérarchie ; les contextes organisationnels post-AI exigent une autorité par le biais d'un processus délibératif approprié entre les perspectives des parties prenantes. La documentation relative au développement du cadre (rapports d'incidents, journaux de sessions) est conservée dans le référentiel du projet mais n'est pas rendue publique dans l'attente d'un examen par les pairs." + "future_development_text": "Berlin et Weil joueront un rôle essentiel dans le développement de la composante \"délibération\" : leurs travaux fournissent les fondements philosophiques permettant de comprendre comment préserver l'action humaine sur les décisions relatives aux valeurs à mesure que les capacités de l'IA s'accélèrent. La théorie organisationnelle traditionnelle (Weber, Taylor) traite de l'autorité par le biais de la hiérarchie ; les contextes organisationnels post-AI exigent une autorité par le biais d'un processus délibératif approprié entre les perspectives des parties prenantes. La documentation relative au développement du cadre (rapports d'incidents, journaux de sessions) est conservée dans le référentiel du projet mais n'est pas rendue publique dans l'attente d'un examen par les pairs.", + "section_4_heading": "Souveraineté des données autochtones et gouvernance polycentrique" }, "research_collaboration": { "heading": "Possibilités de collaboration en matière de recherche", @@ -290,7 +295,7 @@ "offer_2_title": "Documentation", "offer_2_desc": "Spécifications d'architecture, modèles de mise en œuvre, règles de gouvernance", "offer_3_title": "Audit Data", - "offer_3_desc": "1 130+ décisions de gouvernance dans MongoDB (exportations anonymes disponibles)", + "offer_3_desc": "171 800+ décisions de gouvernance dans MongoDB (exportations anonymes disponibles)", "offer_4_title": "Soutien au déploiement", "offer_4_desc": "Aide à la mise en place d'instances locales ou en nuage pour les tests", "offer_5_title": "Coordination", @@ -302,6 +307,44 @@ "cannot_offer_2": "Ressources d'ingénierie dédiées (contraintes de capacité)", "cannot_offer_3": "Lieux de publication garantis (mais nous soutenons les efforts de soumission)", "inquiry_button": "Recherche Collaboration Enquête" + }, + "steering_vectors": { + "heading": "Vecteurs de guidage et biais mécanique dans l'IA souveraine", + "intro": "Cet article introduit une distinction critique entre le biais mécanique (distorsions pré-raisonnement intégrées dans les poids et activations du modèle) et le biais de raisonnement (erreurs délibératives dans le traitement en chaîne de pensée). Les approches traditionnelles de débiaisage confondent ces catégories.", + "techniques_heading": "Panorama des techniques", + "technique_1": "Contrastive Activation Addition (CAA) : Correction de biais au moment de l'inférence par addition/soustraction de vecteurs d'activation", + "technique_2": "Representation Engineering (RepE) : Sondes linéaires pour identifier et modifier les représentations de concepts", + "technique_3": "FairSteer & DSO : Guidage orienté équité par optimisation distributionnellement robuste", + "technique_4": "Autoencodeurs épars : Interprétabilité mécanistique par décomposition de neurones polysémantiques", + "application_heading": "Application à Village Home AI", + "application_text": "Le déploiement Village Home AI utilise des modèles Llama 3.1/3.2 ajustés par QLoRA où les vecteurs de guidage peuvent être appliqués au moment de l'inférence. Cela crée une architecture de gouvernance à deux couches.", + "read_link": "Lire l'article (HTML) →", + "pdf_link": "Télécharger le PDF" + }, + "taonga": { + "heading": "Gouvernance de guidage centrée sur les Taonga : IA polycentrique pour la souveraineté des données autochtones", + "limitation_label": "Limitation importante :", + "limitation_text": "Cet article a été rédigé sans co-auteur maori et fait actuellement l'objet d'une évaluation par les pairs autochtones. Il propose une architecture technique informée par la recherche publiée sur la souveraineté des données autochtones mais n'a pas encore été validé par des chercheurs ou des communautés maories.", + "intro": "Cet article propose une architecture de gouvernance polycentrique où les iwi et les organisations communautaires maintiennent une autorité co-égale aux côtés des opérateurs techniques.", + "foundations_heading": "Fondements théoriques", + "foundation_1": "Principes CARE (Carroll et al., 2020) : Bénéfice collectif, Autorité de contrôle, Responsabilité, Éthique", + "foundation_2": "Gouvernance polycentrique d'Ostrom (1990) : Centres d'autorité multiples et chevauchants plutôt qu'un contrôle hiérarchique unique", + "foundation_3": "Concepts Te ao Maori : Kaitiakitanga (tutelle), rangatiratanga (autodétermination), whakapapa (identité relationnelle) comme principes architecturaux", + "integration_heading": "Intégration avec Tractatus", + "integration_text": "L'article étend le PluralisticDeliberationOrchestrator pour prendre en charge les autorités communautaires avec droit de veto sur les données classifiées taonga.", + "read_link": "Lire le brouillon (HTML) →", + "pdf_link": "Télécharger le PDF" + }, + "home_ai": { + "heading": "Home AI : Plateforme de recherche en gouvernance souveraine", + "intro": "Home AI représente une étape de recherche significative : gouvernance Tractatus complète intégrée dans un pipeline d'inférence de modèle de langue souverain entraîné localement.", + "architecture_heading": "Architecture à deux modèles", + "arch_1": "Modèle rapide (Llama 3.2 3B) : Réponses à faible latence pour les requêtes courantes, avec pré-filtrage de gouvernance", + "arch_2": "Modèle approfondi (Llama 3.1 8B) : Raisonnement complexe avec pipeline de gouvernance complet", + "arch_3": "Ajustement QLoRA : Adaptation paramétrique efficace sur matériel local", + "research_heading": "Importance pour la recherche", + "research_text": "Home AI ouvre la question de recherche de la gouvernance-dans-la-boucle-d'entraînement pour les modèles contrôlés par la communauté.", + "learn_more": "En savoir plus sur Home AI →" } }, "footer": { @@ -408,7 +451,7 @@ ], "collaborate_text": "Collaborez avec nous : Nous accueillons les chercheurs intéressés par l'étude de cette application des principes architecturaux à la gouvernance de l'IA. Nous pouvons fournir un accès aux journaux d'audit, au code du cadre et à la documentation d'intégration pour une étude empirique.", "contact_link": "Contact pour la collaboration →", - "hf_space_link": "🤗 Explorer 3 942 décisions d'audit sur HF →", + "audit_explorer_link": "Explorer 171 800+ décisions d'audit →", "values_link": "Valeurs et principes →" } } diff --git a/public/locales/mi/common.json b/public/locales/mi/common.json index d471a2d8..0e4470ec 100644 --- a/public/locales/mi/common.json +++ b/public/locales/mi/common.json @@ -80,6 +80,37 @@ "submitting": "E tuku ana..." }, "navbar": { + "research": "Rangahau", + "publications": "Ngā Tānga", + "publications_desc": "Tirohia ngā pepa rangahau", + "for_researchers": "Mō Ngā Kairangahau", + "for_researchers_desc": "Ngā pātai tuwhera me te mahi tahi", + "research_timeline": "Rārangi Wā Rangahau", + "research_timeline_desc": "Te whakawhanaketanga o te rangahau", + "framework": "Te Anga", + "system_architecture": "Te Hoahoa Pūnaha", + "system_architecture_desc": "Tirohanga whānui hoahoa hangarau", + "for_implementers": "Mō Ngā Kaiwhakatinana", + "for_implementers_desc": "Aratohu whakaurunga me ngā tauira waehere", + "village_case_study": "Rangahau Tauira Village", + "village_case_study_desc": "Ngā taunakitanga whakatū whakamahi", + "home_ai": "Home AI", + "home_ai_desc": "Tauira reo rangatiratanga i whakangungu ā-rohe", + "agent_lightning": "Agent Lightning", + "agent_lightning_desc": "Whakaurunga whakapai mahinga", + "for_leaders": "Mō Ngā Kaihautū", + "for_leaders_desc": "Tirohanga rautaki me te whakamāramatanga pakihi", + "blog": "Blog", + "about": "Mō Mātou", + "about_tractatus": "Mō Tractatus", + "values": "Ngā Uara", + "docs": "Ngā Tuhinga", + "github": "GitHub", + "koha": "Koha", + "navigation": "Whakatere", + "resources": "Ngā Rauemi", + "documentation": "Ngā Tuhinga", + "mobile_feedback": "Urupare", "feedback": "Tukuna He Urupare", "feedback_desc": "I whakahaeretia e Tractatus + AL", "company_heading": "Kamupene", diff --git a/public/locales/mi/homepage.json b/public/locales/mi/homepage.json index f47928b5..100c22a8 100644 --- a/public/locales/mi/homepage.json +++ b/public/locales/mi/homepage.json @@ -61,7 +61,7 @@ "researcher_f1": "Ngā tūāpapa ā-tikanga me ngā whakatūturu", "researcher_f2": "Te tātaritanga momo hapa", "researcher_f3": "Ngā pātai rangahau tuwhera", - "researcher_f4": "3,942 whakatau arotake i runga i Hugging Face", + "researcher_f4": "171,800+ whakatau arotake kua tātaritia", "researcher_cta": "Tūhura rangahau →", "implementer_title": "Mā ngā Kaiwhakatinana", "implementer_subtitle": "Waehere me ngā aratohu whakakotahi", diff --git a/public/researcher.html b/public/researcher.html index bb2f2727..d926b7c5 100644 --- a/public/researcher.html +++ b/public/researcher.html @@ -127,7 +127,7 @@

    Development Context

    - Tractatus was developed over six months (April–October 2025) in progressive stages that evolved into a live demonstration of its capabilities in the form of a single-project context (https://agenticgovernance.digital). Observations derive from direct engagement with Claude Code (Anthropic's Sonnet 4.5 model) across approximately 500 development sessions. This is exploratory research, not controlled study. + Tractatus has been developed from April 2025 and is now in active production (11+ months). What began as a single-project demonstration has expanded to include production deployment at Village Home Trust and sovereign language model governance through Home AI. Observations derive from direct engagement with Claude Code (Anthropic Claude models, Sonnet 4.5 through Opus 4.6) across over 1,000 development sessions. This is exploratory research, not controlled study.

    @@ -209,11 +209,9 @@ - - Explore 3,942 Audit Decisions on Hugging Face → + Explore 171,800+ Audit Decisions → Values & Principles → @@ -223,6 +221,147 @@

    + +
    +
    +
    +
    + +
    +
    +
    +

    + Steering Vectors and Mechanical Bias in Sovereign AI +

    +

    + STO-RES-0009 v1.1 | Status: Published, February 2026 +

    +
    +
    + +
    +

    + This paper introduces a critical distinction between mechanical bias (pre-reasoning distortions embedded in model weights and activations) and reasoning bias (deliberative errors in chain-of-thought processing). Traditional debiasing approaches conflate these categories, leading to interventions that address symptoms rather than causes. +

    + +

    Technique Survey

    +
      +
    • Contrastive Activation Addition (CAA): Inference-time bias correction by adding/subtracting activation vectors extracted from contrastive prompt pairs
      • +
    • Representation Engineering (RepE): Linear probes to identify and modify concept representations within model layers
      • +
    • FairSteer & DSO: Fairness-oriented steering through distributionally-robust optimization
      • +
    • Sparse Autoencoders: Mechanistic interpretability through decomposition of polysemantic neurons into monosemantic features
      • +
    + +

    Application to Village Home AI

    +

    + The Village Home AI deployment uses QLoRA-fine-tuned Llama 3.1/3.2 models where steering vectors can be applied at inference time. This creates a two-layer governance architecture: Tractatus provides structural constraints on decision boundaries, while steering vectors address pre-reasoning mechanical biases within the model itself. Together, they represent governance that operates both outside and inside the model. +

    + +
    + Read Paper (HTML) → + Download PDF +
    +
    +
    + + +
    +
    +
    +
    + +
    +
    +
    +

    + Taonga-Centred Steering Governance: Polycentric AI for Indigenous Data Sovereignty +

    +

    + STO-RES-0010 v0.1 | Status: DRAFT — in indigenous peer review +

    +
    +
    + +
    + Important Limitation: + This paper was written without Maori co-authorship and is currently undergoing indigenous peer review. It proposes a technical architecture informed by published indigenous data sovereignty scholarship (Carroll et al., Hudson et al., Kukutai & Taylor) but has not yet been validated by Maori researchers or communities. We present it transparently as a starting point for collaboration, not a finished framework. +
    + +
    +

    + This paper proposes a polycentric governance architecture where iwi and community organisations maintain co-equal authority alongside technical system operators. Rather than treating indigenous data sovereignty as a compliance requirement to be satisfied retroactively, the architecture embeds community governance rights structurally through taonga registries, steering packs with provenance tracking, and withdrawal rights. +

    + +

    Theoretical Foundations

    +
      +
    • CARE Principles (Carroll et al., 2020): Collective Benefit, Authority to Control, Responsibility, Ethics — applied to AI governance data flows
      • +
    • Ostrom Polycentric Governance (1990): Multiple overlapping authority centres rather than single hierarchical control
      • +
    • Te ao Maori concepts: Kaitiakitanga (guardianship), rangatiratanga (self-determination), whakapapa (relational identity) as architectural principles, not just metadata labels
      • +
    + +

    Integration with Tractatus

    +

    + The paper extends the PluralisticDeliberationOrchestrator to support community stakeholder authorities with veto rights over taonga-classified data. Steering packs become governed data objects with full provenance tracking, access control, and the right of withdrawal — communities can revoke access to cultural knowledge at any time, and the system architecturally enforces that revocation. +

    + +
    + Read Draft (HTML) → + Download PDF +
    +
    +
    + + +
    +
    +
    +
    + +
    +
    +
    +

    + Home AI: Sovereign Governance Research Platform +

    +

    + Status: Inference operational | Training pipeline in progress +

    +
    +
    + +
    +

    + Home AI represents a significant research milestone: full Tractatus governance embedded in a locally-trained, sovereign language model inference pipeline. This is the first deployment where governance operates inside the model serving layer rather than alongside an external API. +

    + +

    Two-Model Architecture

    +
      +
    • Fast model (Llama 3.2 3B): Low-latency responses for routine queries, with governance pre-screening
      • +
    • Deep model (Llama 3.1 8B): Complex reasoning with full governance pipeline, including BoundaryEnforcer and PluralisticDeliberationOrchestrator
      • +
    • QLoRA fine-tuning: Parameter-efficient adaptation on local hardware, enabling community-specific model customisation without cloud dependency
      • +
    + +

    Research Significance

    +

    + Home AI opens the research question of governance-inside-the-training-loop for community-controlled models. Training data never leaves the local infrastructure; governance rules shape model behaviour through both fine-tuning data curation and inference-time constraints. This creates a fundamentally different governance surface than API-mediated approaches. +

    + +
    + Learn more about Home AI → +
    +
    +
    +
    @@ -655,19 +794,17 @@

    Research Documentation

    -
    - HF + 📊
    - Interactive Audit Log Explorer (Hugging Face) - Explore 3,942 governance decisions from production deployment + Interactive Audit Analytics Dashboard + Explore 171,800+ governance decisions from production deployment
    - +     @@ -926,8 +1063,8 @@
    -

    What We've Validated (October 2025)

    -

    After 6 months of development and ~500 Claude Code sessions, we have grounded evidence for:

    +

    What We've Validated (February 2026)

    +

    After 11+ months of development, 1,000+ Claude Code sessions, and production deployment at Village Home Trust, we have grounded evidence for:

    @@ -935,14 +1072,14 @@
    • BoundaryEnforcer successfully blocks values decisions before execution
    • Pre-commit hooks prevent inst_017 violations (absolute assurance terms)
      • -
    • 1,130+ audit decisions recorded in MongoDB (tractatus_dev.audit_log)
      • +
    • 171,800+ audit decisions recorded in MongoDB (tractatus_dev.audit_log)
    -

    ✅ Instruction Persistence Works in Single-Session Context

    +

    ✅ Instruction Persistence Works Across Sessions

      -
    • InstructionPersistenceClassifier maintains 62 active instructions (STRATEGIC: 22, SYSTEM: 21)
      • +
    • InstructionPersistenceClassifier maintains 68 active instructions (STRATEGIC: 27, SYSTEM: 21, OPERATIONAL: 18, TACTICAL: 2)
    • Pattern bias detection prevents AI from overriding explicit organizational directives
    • Temporal scopes (STRATEGIC/OPERATIONAL/TACTICAL) enforced successfully
    @@ -967,11 +1104,15 @@
    -

    ✅ Single-Project Governance Successful

    +

    ✅ Multi-Deployment Governance Successful

      -
    • Framework governed development of agenticgovernance.digital
      • -
    • Cultural DNA rules (inst_085-089) enforced through pre-commit hooks
      • -
    • Zero credential exposures (defense-in-depth: 5 layers verified)
      • +
    • Framework governs agenticgovernance.digital (11+ months continuous operation)
      • +
    • Village Home Trust production deployment: zero governance violations
      • +
    • Home AI sovereign inference governance: operational
      • +
    • Cultural DNA rules (inst_085-089) enforced through pre-commit hooks (4+ months operational)
      • +
    • Phase 5 integration: 100% complete (all 6 services, 203/203 tests passing)
      • +
    • Multilingual support: EN, DE, FR, Te Reo Maori
      • +
    • Zero credential exposures (defense-in-depth: 5 layers verified)
    @@ -995,18 +1136,18 @@

    ❌ Adversarial Robustness

      -
    • Validated: Normal development workflows (~500 sessions)
      • +
    • Validated: Normal development workflows (1,000+ sessions)
    • Unknown: Resistance to deliberate bypass attempts, jailbreak prompts, adversarial testing
    • Research need: Red-team evaluation by security researchers
    -

    ❌ Cross-Platform Consistency

    +

    ⚠️ Cross-Platform Consistency (Partial)

      -
    • Validated: Claude Code (Anthropic Sonnet 4.5) only
      • -
    • Unknown: Generalizability to Copilot, GPT-4, AutoGPT, LangChain, CrewAI, open models
      • -
    • Research need: Cross-platform validation studies
      • +
    • Validated: Claude Code (Anthropic Claude, Sonnet 4.5 through Opus 4.6) and Home AI (Llama 3.1/3.2 via QLoRA)
      • +
    • Unknown: Generalizability to Copilot, GPT-4, AutoGPT, LangChain, CrewAI, other open models
      • +
    • Research need: Broader cross-platform validation studies beyond Claude and Llama families
    @@ -1023,8 +1164,8 @@

    ❌ Rule Proliferation Impact

      -
    • Validated: 62 active instructions work in current context
      • -
    • Unknown: Performance degradation as rule count grows (projected 30-50 instructions within 12 months)
      • +
    • Validated: 68 active instructions work in current context (exceeded original projection)
      • +
    • Unknown: Performance degradation ceiling as rule count continues to grow
    • Known concern: Transactional overhead (CrossReferenceValidator checks), context window pressure, cognitive load
    • Research need: Rule consolidation strategies, automated governance optimization
    @@ -1277,7 +1418,7 @@ ✅

    Audit Data

    -

    1,130+ governance decisions in MongoDB (anonymized exports available)

    +

    171,800+ governance decisions in MongoDB (anonymized exports available)

    @@ -1355,6 +1496,17 @@

    AI Governance & Technical Context

    • Anthropic (2024). Claude Code: Technical Documentation. Available at: https://docs.anthropic.com/claude-code
      • +
    • Bai, Y., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073
      • +
    • Turner, A., et al. (2023). Activation Addition: Steering Language Models Without Optimization. arXiv:2308.10248
      • +
    • Zou, A., et al. (2023). Representation Engineering: A Top-Down Approach to AI Transparency. arXiv:2310.01405
      • +
    + +

    Indigenous Data Sovereignty & Polycentric Governance

    +
      +
    • Carroll, S.R., et al. (2020). The CARE Principles for Indigenous Data Governance. Data Science Journal, 19(1), 43.
      • +
    • Hudson, M., et al. (2023). Indigenous Data Sovereignty and Governance. In: Indigenous Peoples' Rights in Data.
      • +
    • Kukutai, T. & Taylor, J. (2016). Indigenous Data Sovereignty: Toward an Agenda. ANU Press.
      • +
    • Ostrom, E. (1990). Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge University Press.
    diff --git a/public/timeline.html b/public/timeline.html index 4e929db8..1e9ba7b0 100644 --- a/public/timeline.html +++ b/public/timeline.html @@ -190,7 +190,7 @@ October 19-23, 2025

    Interactive Demonstrations

    - Interactive SVG architecture diagram with clickable service nodes. The 27027 incident recreated as a step-by-step demo showing how each governance service intercepts the failure. A Hugging Face Space was deployed with 3,942 anonymised governance decisions from production, allowing independent exploration of real audit data. + Interactive SVG architecture diagram with clickable service nodes. The 27027 incident recreated as a step-by-step demo showing how each governance service intercepts the failure. An interactive audit analytics dashboard was launched with governance decisions from production (initially 3,942; now exceeding 171,800+), allowing independent exploration of real audit data.

    WCAG accessibility compliance was implemented across all audience pages — skip links, focus indicators, keyboard navigation, and screen reader support.