Architectural AI Governance at Community Scale — A Technical Examination of Village AI

For AI/ML Researchers and Safety Researchers

A five-part technical series examining an alternative approach to AI alignment: architectural constraint at inference time rather than alignment through training. Written for researchers familiar with RLHF, constitutional AI, mechanistic interpretability, and the broader alignment literature.

The Series

1. What AI Is — And Where the Alignment Problem Actually Sits

Emergent reasoning, scaling laws, and the distinction between capability and controllability. Why the alignment problem as experienced by deployed community systems differs from the alignment problem as studied in the laboratory — and why that gap matters.

2. Foundation Models vs. Domain-Specialised Inference — A Structural Analysis

Distribution shift, base rate bias, and the silent substitution problem. How training data composition determines default behaviour, why prompting and RLHF do not resolve distributional priors at the tail, and what a Specialised Layer strategy on an 8B parameter model offers (and sacrifices) compared to frontier systems.

3. Why Training-Time Governance Fails — Architectural Constraints as an Alternative

The 27027 incident as a case study in alignment failure: an AI system silently substituting therapeutic language for theological language despite explicit instructions. Why fine-tuning, RLHF, and constitutional AI do not solve this class of problem. Guardian Agents as epistemic separation — verification systems that are structurally independent of the model they monitor.

4. What Is Live in Production — An Unvarnished Inventory

The actual system: villageai-8b-corrected-v4, Specialised Layer fine-tuning, AMD RX 7900 XTX inference via WireGuard with CPU fallback, cosine similarity verification against source documents, four Guardian Agent layers, and an adaptive feedback loop. What works, what remains unproven, and where we are aware of limitations.

5. Beyond the Model — Platform Architecture and Governance Integration

AI as one component within a sovereign community platform. How architectural constraints extend beyond the model into data isolation, consent architecture, vocabulary-driven framing, and federated inter-community governance. What this approach sacrifices in capability and what it gains in verifiability.

Who This Is For

These articles are written for researchers working on AI alignment, safety, interpretability, and governance. The assumption is familiarity with transformer architectures, fine-tuning methods, reinforcement learning from human feedback, and the current alignment literature.

The contribution is not theoretical. It is a description of a deployed system — operational since October 2025 — that takes a different approach to the alignment problem: architectural constraint at inference time rather than alignment through training. The approach is young, operates at modest scale, and has not been independently evaluated. We present it as a case study, not as a solution.

We are candid about limitations. An 8B parameter model with domain specialisation cannot match frontier systems on general capability. The Guardian Agent architecture adds latency. The cosine similarity verification depends on the quality and coverage of the source document corpus. These are real trade-offs, and we discuss them throughout.

Landing