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.
+ Production deployment of Tractatus governance on locally-trained open-source models, demonstrating framework portability beyond Claude. +
++ 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. +
+ +// 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' };
+}
+ + Status: Inference governance operational. Sovereign training pipeline installation in progress. Production deployment at Village Home Trust validates governance portability across model architectures. +
++ 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) +
++ Extract activation differences between contrasting prompts, then add scaled vectors during inference to steer model behaviour toward desired attributes. +
++ Identify and manipulate internal representations of concepts (honesty, safety, fairness) through probing and directional manipulation of model activations. +
++ Identify bias-critical layers through probing classifiers, then apply targeted corrections at those specific layers rather than globally. +
++ 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). +
++ 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. +
++ Model configurations, fine-tuning data, and steering vectors are treated as taonga (treasures) with full provenance tracking, access control, and community consent requirements. +
++ Every data object tracks its full lineage. Communities retain the right to withdraw their data and configurations, with architectural enforcement of removal propagation. +
++ Extends PluralisticDeliberationOrchestrator with iwi/community as co-equal governance authorities. Access decisions require multi-party consensus. +
++ Architecturally enforces Collective benefit, Authority to control, Responsibility, and Ethics for indigenous data governance within the Tractatus framework. +
+Status: Research Phase
+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.
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. +
+Status: Conceptual / In Peer Review
++ Governed data objects with provenance tracking, withdrawal rights, and polycentric community governance. Extends PluralisticDeliberationOrchestrator for indigenous data sovereignty. +
++ 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. +
++ Current status: Inference governance operational. Training pipeline installation in progress. First non-Claude deployment surface for Tractatus governance. +
+ + ++ 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. +
++ 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. +
+ + ++ 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 → + +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.
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.
+ 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. +
+ ++ 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. +
+ + ++ STO-RES-0010 v0.1 | Status: DRAFT — in indigenous peer review +
++ 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. +
+ ++ 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. +
+ + ++ 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. +
+ ++ 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. +
+ + +After 6 months of development and ~500 Claude Code sessions, we have grounded evidence for:
+After 11+ months of development, 1,000+ Claude Code sessions, and production deployment at Village Home Trust, we have grounded evidence for:
✅ Instruction Persistence Works in Single-Session Context
+✅ Instruction Persistence Works Across Sessions
✅ Single-Project Governance Successful
+✅ Multi-Deployment Governance Successful
❌ Adversarial Robustness
❌ Cross-Platform Consistency
+⚠️ Cross-Platform Consistency (Partial)
❌ Rule Proliferation Impact
Audit Data
-1,130+ governance decisions in MongoDB (anonymized exports available)
+171,800+ governance decisions in MongoDB (anonymized exports available)
- 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.