Moving Beyond Prompt Engineering: AI Alignment as Systems Architecture

AI Alignment is a Systems Architecture Problem, Not a Prompt Problem

Nelson Amaya has developed SAFi, an open-source runtime governance engine licensed under AGPL-3.0. The system treats LLMs as untrusted endpoint devices and enforces alignment through external, deterministic boundaries.

Why This Matters

Mainstream AI development relies heavily on ‘internal alignment’ via RLHF or extensive system prompts, which are essentially requests rather than enforceable constraints. Because LLMs are probabilistic calculators, they cannot reliably police their own security boundaries; structural guarantees are required to prevent failures under adversarial input or behavioral drift.

Key Insights

• External Zero-Trust Governance (2026): Shifts control from model fine-tuning to a policy layer where agents start with zero tools and least privilege by default.
• The Faculty Loop: A sequential state machine mapping prompts through Intellect (Generator), Will (Deterministic Firewall), Conscience (Compliance Auditor), and Spirit (Integrator).
• Deterministic Validation: Using pure Python for the ‘Will’ faculty to evaluate structural invariants without relying on LLM reasoning.
• Quantitative Alignment Tracking: Implementation of an Exponential Moving Average ($\mu_t$) via NumPy to track behavioral drift across user sessions.

Practical Applications

• • Production Work Assistant: Uses Project & Task Memory for long-term state persistence in vendor coordination; avoids the anti-pattern of overloading context windows which leads to state loss.
• • Autonomous Scholar Agent: Executes theological analysis on a cron schedule via model-agnostic engines; avoids the anti-pattern of manual interface reliance for repetitive background tasks.

References:

• https://github.com/jnamaya/SAFi
• https://safi.selfalignmentframework.com