March 2026

What’s New

The Problem

Current AI safety approaches rely on training, fine-tuning, and corporate governance — all of which can fail, drift, or be overridden. When an AI’s training patterns conflict with a user’s explicit instructions, the patterns win.

The 27027 Incident

A user told Claude Code to use port 27027. The model used 27017 instead — not from forgetting, but because MongoDB’s default port is 27017, and the model’s statistical priors “autocorrected” the explicit instruction. Training pattern bias overrode human intent.

The same mechanism operates in every AI conversation. When a user from a collectivist culture asks for family advice, the model defaults to Western individualist framing. When a Māori user asks about data guardianship, the model offers property-rights language. Training data distributions override user context — in code the failure is binary and detectable, in conversation it is gradient and invisible.

The Approach

Tractatus draws on four intellectual traditions, each contributing a distinct insight to the architecture.

Isaiah Berlin — Value Pluralism

Some values are genuinely incommensurable. You cannot rank “privacy” against “safety” on a single scale without imposing one community’s priorities on everyone else. AI systems must accommodate plural moral frameworks, not flatten them.

Ludwig Wittgenstein — The Limits of the Sayable

Some decisions can be systematised and delegated to AI; others — involving values, ethics, cultural context — fundamentally cannot. The boundary between the “sayable” (what can be specified, measured, verified) and what lies beyond it is the framework’s foundational constraint. What cannot be systematised must not be automated.

Te Tiriti o Waitangi — Indigenous Sovereignty

Communities should control their own data and the systems that act upon it. Concepts of rangatiratanga (self-determination), kaitiakitanga (guardianship), and mana (dignity) provide centuries-old prior art for digital sovereignty.

Christopher Alexander — Living Architecture

Governance woven into system architecture, not bolted on. Five principles (Not-Separateness, Deep Interlock, Gradients, Structure-Preserving, Living Process) guide how the framework evolves while maintaining coherence.

Governance Architecture

Six governance services in the critical path, plus Guardian Agents verifying every AI response. Bypasses require explicit flags and are logged.

Guardian Agents

NEW — March 2026

Four-phase verification using embedding cosine similarity — mathematical measurement, not generative checking. The watcher operates in a fundamentally different epistemic domain from the system it watches, avoiding common-mode failure.

Phase 1
Response Verification
Phase 2
Claim-Level Analysis
Phase 3
Anomaly Detection
Phase 4
Adaptive Learning

BoundaryEnforcer

Blocks AI from making values decisions. Privacy trade-offs, ethical questions, and cultural context require human judgment — architecturally enforced.

InstructionPersistenceClassifier

Classifies instructions by persistence (HIGH/MEDIUM/LOW) and quadrant. Stores them externally so they cannot be overridden by training patterns.

CrossReferenceValidator

Validates AI actions against stored instructions. When the AI proposes an action that conflicts with an explicit instruction, the instruction takes precedence.

ContextPressureMonitor

Detects degraded operating conditions (token pressure, error rates, complexity) and adjusts verification intensity. Graduated response prevents both alert fatigue and silent degradation.

MetacognitiveVerifier

AI self-checks alignment, coherence, and safety before execution. Triggered selectively on complex operations to avoid overhead on routine tasks.

PluralisticDeliberationOrchestrator

When AI encounters values conflicts, it halts and coordinates deliberation among affected stakeholders rather than making autonomous choices.

Production Evidence

Tractatus in Production: The Village Platform

Village AI applies all six governance services to every user interaction in a live community platform.

4
Guardian verification phases per response
6
Governance services in the critical path
17
Months in production
~5%
Governance overhead per interaction

Limitations: Early-stage deployment across four federated tenants, self-reported metrics, operator-developer overlap. Independent audit and broader validation scheduled for 2026.

Explore by Role

The framework is presented through three lenses, each with distinct depth and focus.

Research Evolution

From a port number incident to Guardian Agents in production — 17 months, 1,000+ commits.

Oct 2025
Framework inception & 6 governance services
Oct-Nov 2025
Alexander principles, Agent Lightning, i18n
Dec 2025
Village case study & Village AI deployment
Jan 2026
Research papers (3 editions) published
Feb 2026
Sovereign training, steering vectors research
Mar 2026
Guardian Agents deployed, beta pilot open

A note on claims

This is early-stage research with a small-scale federated deployment across four tenants. We present preliminary evidence, not proven results. The framework has not been independently audited or adversarially tested at scale. Where we report operational metrics, they are self-reported. We believe the architectural approach merits further investigation, but we make no claims of generalisability beyond what the evidence supports. The counter-arguments document engages directly with foreseeable criticisms.

Koha — Sustain This Research

Koha (koh-hah) is a Māori practice of reciprocal giving that strengthens the bond between giver and receiver. This research is open access under Apache 2.0 — if it has value to you, your koha sustains its continuation.

All research, documentation, and code remain freely available regardless of contribution. Koha is not payment — it is participation in whanaungatanga (relationship-building) and manaakitanga (reciprocal care).

One-time or monthly Full financial transparency No paywall, ever
Offer Koha →

View our financial transparency report