Sati Is Not Inside the Model

2026-06-03 · Source: AI Advances - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Advanced, long

Summary

The article proposes an external, Abhidhamma-inspired "sati" (mindfulness) gate for Large Language Model (LLM) workflows, arguing that true awareness cannot reside within the model itself. It posits that erroneous LLM outputs originate early in the generation process, not at the visible error stage, necessitating pre-output intervention. The proposed runtime outlines a seven-step gate: Object determination, Surface classification, Memory/evidence check, Retrieval before output, Sati gate (separating fact from inference), Response trajectory, and Exit audit. This method relies on prompt structure and retrieval discipline, enabling users to shape model behavior without fine-tuning or accessing internal weights. It reframes alignment as a human-side responsibility, where the user's configuration, instructions, and external gates are crucial for guiding the model's response before probability becomes speech.

Key takeaway

For AI Engineers designing LLM applications, recognize that model "awareness" is an external workflow concern, not an internal model feature. Implement pre-output gates using prompt structure and retrieval discipline to prevent erroneous trajectories before commitment. You should integrate steps like object determination, surface classification, and memory checks early in the generation process. This proactive approach ensures human-side alignment, stopping wrong responses before they become conditioning for subsequent output, enhancing reliability and safety.

Key insights

LLM "mindfulness" must be an external pre-output gate, not an internal model capability, to prevent bad outputs.

Principles

• Bad LLM outputs start pre-visible error.
• Shape behavior before output, not after.
• External gates enable human-side alignment.

Method

A seven-step pre-output gate is implemented via prompt structure and retrieval discipline: Object determination, Surface classification, Memory/evidence check, Retrieval before output, Sati gate, Response trajectory, and Exit audit.

In practice

• Restate user's query for object determination.
• Classify query surface (e.g., legal, casual).
• Retrieve relevant material before generation.

Topics

• LLM Alignment
• Prompt Engineering
• Pre-output Gates
• Retrieval-Augmented Generation
• Human-side Alignment
• Abhidhamma Cognition

Best for: AI Engineer, Machine Learning Engineer, Prompt Engineer

Related on AIssential

• Activation Steering for Aligned Open-ended Generation without Sacrificing Coherence — Activation steering offers a lightweight runtime defense against LLM misalignment by continuously correcting activations.
• Engineering a human-aligned LLM evaluation workflow with Prodigy and DSPy — Human-aligned LLM evaluation requires iterative workflows that integrate granular human feedback to engineer context-specific metrics.
• My Workflow for Understanding LLM Architectures — Inspect open-weight LLM code and config files to understand architectures when papers lack detail.
• The Must-Know Topics for an LLM Engineer — LLM engineering requires understanding an interdependent stack, from data representation and model architecture to training, optimization, and evaluation.
• Human Mind Reconstruction w/ AI (ToM, UserHarness) — Decoupling epistemic structure from LLM generation significantly boosts reasoning accuracy and efficiency, even for smaller models.
• Research on LLM alignment as latent discourse-level regimes vs. token-level filtering? — LLM alignment may stem from global latent space geometry and discourse-level regimes, not just token-level filtering.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Advances - Medium.