diff --git a/docs/markdown/research-timeline.md b/docs/markdown/research-timeline.md index 890b0aa3..bab2e2df 100644 --- a/docs/markdown/research-timeline.md +++ b/docs/markdown/research-timeline.md @@ -25,7 +25,7 @@ tags: timeline, research, governance, indigenous, steering-vectors, alexander, w ## Overview -This document traces the intellectual and technical evolution of the Tractatus Framework and its ecosystem of projects, from early experiments with AI-augmented project management in early 2025 through to sovereign small language model governance and indigenous-centred polycentric steering in February 2026. The narrative follows ideas as they developed across five interconnected projects: SyDigital, Digital Sovereignty Passport, Family History, Tractatus, and the Village Home AI platform. +This document traces the intellectual and technical evolution of the Tractatus Framework and its ecosystem of projects, from early experiments with AI-augmented project management in early 2025 through to sovereign small language model governance and indigenous-centred polycentric steering in February 2026. The narrative follows ideas as they developed across five interconnected projects: SyDigital, Digital Sovereignty Passport, Family History, Tractatus, and the Village Village AI platform. --- @@ -310,7 +310,7 @@ The passport project demonstrated that governance architecture need not slow dev --- -## Phase 5: Village Home AI and Sovereign Deployment +## Phase 5: Village Village AI and Sovereign Deployment **Period:** December 2025 -- February 2026 **Platform commits:** 1 December 2025 (homepage promotion), 9 December 2025 (case study and technical details) @@ -318,7 +318,7 @@ The passport project demonstrated that governance architecture need not slow dev ### The Sovereign Architecture -The Village Home AI platform represented a fundamental shift from governance of cloud-hosted AI services to governance of locally trained and served models. The architecture: +The Village Village AI platform represented a fundamental shift from governance of cloud-hosted AI services to governance of locally trained and served models. The architecture: - **Tier 1 (Platform Base):** Llama 3.1 8B -- deep reasoning, complex governance decisions - **Tier 2 (Per-Tenant Adapters):** Llama 3.2 3B -- fast inference, routine operations @@ -401,7 +401,7 @@ The paper surveyed five steering vector techniques: ### The Sovereign Advantage -None of these techniques are available through commercial API endpoints. Only sovereign deployments with full access to model weights and activations can extract, inject, and calibrate steering vectors. This makes the Village Home AI's QLoRA-fine-tuned Llama models uniquely positioned to address mechanical bias -- and it reframes the choice between cloud APIs and local deployment as a governance question, not merely a cost or performance question. +None of these techniques are available through commercial API endpoints. Only sovereign deployments with full access to model weights and activations can extract, inject, and calibrate steering vectors. This makes the Village Village AI's QLoRA-fine-tuned Llama models uniquely positioned to address mechanical bias -- and it reframes the choice between cloud APIs and local deployment as a governance question, not merely a cost or performance question. ### Decolonial Reading (v1.1) @@ -500,7 +500,7 @@ The evolution of indigenous values across the project timeline is not a separate **Tractatus (October 2025):** Cultural DNA rules (inst_085-089) architecturally enforced. CARE Principles compliance as structural requirement. Te Tiriti alignment in all cultural decisions. Language parity obligations. -**Village Home AI (December 2025):** Sovereign deployment specifically enables cultural governance impossible through commercial APIs. Training data never leaves community infrastructure. Governance by community design. +**Village Village AI (December 2025):** Sovereign deployment specifically enables cultural governance impossible through commercial APIs. Training data never leaves community infrastructure. Governance by community design. **Steering Vectors (February 2026):** Representational bias reframed as statistical encoding of colonial knowledge hierarchies. Mechanical bias in Western-dominated training corpora identified as epistemic colonialism. @@ -526,7 +526,7 @@ The trajectory is clear: from respectful acknowledgment, through structural inte | 28 Oct 2025 | Cultural DNA rules encoded (inst_085-089) | | 30 Oct 2025 | Christopher Alexander principles integrated (inst_090-094) | | 3 Nov 2025 | Agent Lightning integration and community launch | -| 1 Dec 2025 | Village Home AI platform announced | +| 1 Dec 2025 | Village Village AI platform announced | | 9 Dec 2025 | Village case study with sovereign two-model architecture details | | 19 Jan 2026 | Architectural alignment and korero counter-arguments deployed | | 7 Feb 2026 | Wittgenstein established as primary philosophical foundation | @@ -564,7 +564,7 @@ The immediate research trajectory includes: 1. **Indigenous peer review** of STO-RES-0010 -- the taonga governance paper cannot advance without Maori validation 2. **Sovereign training pipeline** -- extending governance from inference to training -3. **Steering vector implementation** on the Village Home AI platform (four-phase plan in STO-RES-0009) +3. **Steering vector implementation** on the Village Village AI platform (four-phase plan in STO-RES-0009) 4. **Multi-project governance scaling** -- extending architectural enforcement across the full project ecosystem 5. **Community engagement** -- broadening the framework's validation beyond a single-developer context @@ -575,5 +575,5 @@ The central question remains the one that started it all: **how do we work along **Document Metadata:** - **Version:** 1.0 - **Status:** Current -- **Projects Referenced:** SyDigital, Digital Sovereignty Passport, Family History, Tractatus, Village Home AI, Agent Lightning, Community, Platform Admin +- **Projects Referenced:** SyDigital, Digital Sovereignty Passport, Family History, Tractatus, Village Village AI, Agent Lightning, Community, Platform Admin - **Word Count:** ~4,500 diff --git a/docs/markdown/steering-vectors-mechanical-bias-sovereign-ai.md b/docs/markdown/steering-vectors-mechanical-bias-sovereign-ai.md index 54ab4cf2..15132172 100644 --- a/docs/markdown/steering-vectors-mechanical-bias-sovereign-ai.md +++ b/docs/markdown/steering-vectors-mechanical-bias-sovereign-ai.md @@ -14,7 +14,7 @@ This paper investigates whether a class of biases in large language models operates at a sub-reasoning, representational level analogous to motor automaticity in human cognition, and whether steering vector techniques can intervene at this level during inference. We distinguish between *mechanical bias* (statistical patterns that fire at the embedding and early-layer representation level before deliberative processing begins) and *reasoning bias* (distortions that emerge through multi-step chain-of-thought reasoning). Drawing on empirical work in Contrastive Activation Addition (CAA), Representation Engineering (RepE), FairSteer, Direct Steering Optimization (DSO), and Anthropic's sparse autoencoder feature steering, we assess the maturity of each technique and its applicability to sovereign small language models (SLMs) trained and served locally.[^sll] -[^sll]: We use "sovereign small language model" (SLM) for continuity with the technical literature. In the Tractatus framework (STO-INN-0003, v2.1; Stroh & Claude, 2026), these systems are designated "Sovereign Locally-trained Language Models" (SLLs) to emphasise that their distinguishing property is architectural sovereignty — governance authority over training, deployment, and inference — not parameter count. The SLL designation is the more precise term within the framework. We find that sovereign SLM deployments, specifically the Village Home AI platform using QLoRA-fine-tuned Llama 3.1/3.2 models, possess a structural advantage over API-mediated deployments: full access to model weights and activations enables steering vector extraction, injection, and evaluation that is architecturally impossible through commercial API endpoints. We propose a four-phase implementation path integrating steering vectors into the existing two-tier training architecture and Tractatus governance framework. +[^sll]: We use "sovereign small language model" (SLM) for continuity with the technical literature. In the Tractatus framework (STO-INN-0003, v2.1; Stroh & Claude, 2026), these systems are designated "Sovereign Locally-trained Language Models" (SLLs) to emphasise that their distinguishing property is architectural sovereignty — governance authority over training, deployment, and inference — not parameter count. The SLL designation is the more precise term within the framework. We find that sovereign SLM deployments, specifically the Village Village AI platform using QLoRA-fine-tuned Llama 3.1/3.2 models, possess a structural advantage over API-mediated deployments: full access to model weights and activations enables steering vector extraction, injection, and evaluation that is architecturally impossible through commercial API endpoints. We propose a four-phase implementation path integrating steering vectors into the existing two-tier training architecture and Tractatus governance framework. --- @@ -174,9 +174,9 @@ This table reveals that **none of the steering vector techniques described in Se > > **Added reference:** Radhakrishnan, A., Beaglehole, D., Belkin, M., & Boix-Adserà, E. (2026). Exposing biases, moods, personalities, and abstract concepts hidden in large language models. *Science.* Published 19 February 2026. -### 4.2 The Village Home AI Platform +### 4.2 The Village Village AI Platform -The Village platform's Home AI system (Stroh, 2025-2026) is designed as a sovereign small language model (SLM) deployment with the following architecture: +The Village platform's Village AI system (Stroh, 2025-2026) is designed as a sovereign small language model (SLM) deployment with the following architecture: - **Base model:** Llama 3.1 8B (Tier 1 platform base) / Llama 3.2 3B (Tier 2 per-tenant adapters) - **Fine-tuning method:** QLoRA (4-bit quantised Low-Rank Adaptation) @@ -317,7 +317,7 @@ The indicator-wiper analogy suggests a useful distinction between biases that op Steering vector techniques (CAA, RepE, FairSteer, DSO, sparse autoencoder feature steering) provide the theoretical and practical toolkit for such intervention. Critically, these techniques require full access to model weights and activations -- access that is available exclusively in sovereign local deployments and architecturally unavailable through commercial API endpoints. -The Village Home AI platform, with its QLoRA-fine-tuned Llama models, two-tier training architecture, and Tractatus governance integration, is structurally positioned to pioneer the application of steering vectors to cultural bias mitigation in community-serving AI. The proposed four-phase implementation path is conservative, empirically grounded, and designed to produce measurable results within a 16-week timeline. +The Village Village AI platform, with its QLoRA-fine-tuned Llama models, two-tier training architecture, and Tractatus governance integration, is structurally positioned to pioneer the application of steering vectors to cultural bias mitigation in community-serving AI. The proposed four-phase implementation path is conservative, empirically grounded, and designed to produce measurable results within a 16-week timeline. The indicator-wiper problem is solvable. The driver eventually recalibrates. The question for sovereign AI is whether we can accelerate that recalibration -- not by telling the model to "be less biased" (the equivalent of verbal instruction), but by directly adjusting the representations that encode the bias (the equivalent of physical relocation of the indicator stalk). diff --git a/docs/markdown/taonga-centred-steering-governance-polycentric-ai.md b/docs/markdown/taonga-centred-steering-governance-polycentric-ai.md index f4c9a6ca..75068607 100644 --- a/docs/markdown/taonga-centred-steering-governance-polycentric-ai.md +++ b/docs/markdown/taonga-centred-steering-governance-polycentric-ai.md @@ -153,11 +153,11 @@ In this model: | Actor | Role | Governance Source | Example | | --- | --- | --- | --- | -| Platform operator | Technical infrastructure, safety baselines, general debiasing | Tractatus framework, platform constitution | Village / Home AI team | +| Platform operator | Technical infrastructure, safety baselines, general debiasing | Tractatus framework, platform constitution | Village / Village AI team | | Iwi steering authority | Cultural steering for iwi-specific domains | Tikanga, iwi governance structures | Iwi data governance board | | Community trust | Domain-specific or locality-specific steering | Trust charter, community deliberation | Regional health trust, marae committee | | Application operator | Selects and composes steering packs for a specific deployment | Contractual, regulatory, relational obligations | School running a local AI assistant | -| Affected community | Contests outputs, flags bias, triggers review | Rights of participation and appeal | Whanau using a Home AI deployment | +| Affected community | Contests outputs, flags bias, triggers review | Rights of participation and appeal | Whanau using a Village AI deployment | ### 3.4 Steering Registries and Taonga Services @@ -255,11 +255,11 @@ These rights structurally prevent the platform from becoming the default locus o --- -## 5. Case Study: Marae-Based Home AI Deployment +## 5. Case Study: Marae-Based Village AI Deployment ### 5.1 Scenario -A marae in Aotearoa operates a Home AI deployment for its whānau community. The system helps members write stories, summarise kōrero, and triage content for moderation. It runs a Llama 3.2 3B model, Quantised Low-Rank Adaptation (QLoRA) fine-tuned with community-contributed data, on local hardware. +A marae in Aotearoa operates a Village AI deployment for its whānau community. The system helps members write stories, summarise kōrero, and triage content for moderation. It runs a Llama 3.2 3B model, Quantised Low-Rank Adaptation (QLoRA) fine-tuned with community-contributed data, on local hardware. ### 5.2 Steering Configuration @@ -282,7 +282,7 @@ The deployment composes three steering packs: ### 5.3 Steering Provenance in Action -A community member asks the Home AI to summarise a kōrero about a recently deceased kuia. The steering provenance for this inference: +A community member asks the Village AI to summarise a kōrero about a recently deceased kuia. The steering provenance for this inference: ``` Steering Provenance: diff --git a/public/architectural-alignment-policymakers.html b/public/architectural-alignment-policymakers.html index 2f3ce3be..50395d57 100644 --- a/public/architectural-alignment-policymakers.html +++ b/public/architectural-alignment-policymakers.html @@ -372,7 +372,7 @@

7. From Existential Stakes to Everyday Governance

7.1 Why Existential Risk Framing Matters for Policy

-

The existential risk literature may seem remote from practical policy concerns about home AI assistants. The connection is essential:

+

The existential risk literature may seem remote from practical policy concerns about Village AI assistants. The connection is essential:

Containment architectures cannot be developed after the systems that need them exist. If advanced AI systems eventually pose existential risks—a possibility serious researchers take seriously—the governance infrastructure, institutional capacity, and cultural expectations required to contain them must be developed in advance.

Current deployment is the development ground. The patterns that work at village scale become the patterns available when stakes are higher. Constitutional gating implemented for home SLLs creates:

diff --git a/public/architectural-alignment.html b/public/architectural-alignment.html index 0d88744b..ccaf3870 100644 --- a/public/architectural-alignment.html +++ b/public/architectural-alignment.html @@ -225,7 +225,7 @@

4.4 From Existential Stakes to Everyday Deployment

-

Why apply frameworks designed for existential risk to home AI assistants? The answer lies in temporal structure:

+

Why apply frameworks designed for existential risk to Village AI assistants? The answer lies in temporal structure:

Containment architectures cannot be developed after the systems that need them exist. The tooling, governance patterns, cultural expectations, and institutional capacity for AI containment must be built in advance.

diff --git a/public/architecture.html b/public/architecture.html index 3c3127db..9a405002 100644 --- a/public/architecture.html +++ b/public/architecture.html @@ -560,8 +560,8 @@
- - Read the Full Home AI Story → + + Read the Full Village AI Story →

Two-model architecture, three training tiers, thirteen wisdom traditions, indigenous data sovereignty @@ -969,13 +969,13 @@

- +
2
-

Home AI: Sovereign Language Model

+

Village AI: Sovereign Language Model

The current research direction: applying all five architectural principles to model training, not just inference. BoundaryEnforcer operates inside the training loop. Three training tiers (platform, tenant, individual) with governance at each level. @@ -1000,7 +1000,7 @@

Status: inference in production; training pipeline designed, hardware ordered.

- Full Home AI Story → + Full Village AI Story →
@@ -1078,7 +1078,7 @@

- Home AI → + Village AI →
diff --git a/public/docs.html b/public/docs.html index 99154431..0435b6a3 100644 --- a/public/docs.html +++ b/public/docs.html @@ -662,7 +662,7 @@

Case Studies

Research Papers