ENGINEERING INNER ALIGNMENT A CONTROL-THEORETIC APPROACH TO AI SAFETY

CATEGORY

RESEARCH PAPER

READING TIME

6 MIN + PAPER

AUTHOR

HAN KAY

SHARE

SHARE

[X]
[LINKEDIN]
[COPY]

ALIGNMENT IS A SYSTEMS PROPERTY

Most alignment work evaluates behavior after a model has produced it. ConsciOS asks a different question: what control structure would make goals, constraints, internal state, and corrective feedback visible before action? The proposal treats alignment as an architectural property to specify and test, not a claim that good behavior has already been achieved.

A NESTED CONTROL ARCHITECTURE

The model separates short-horizon perception and action, mid-level policy and frame selection, and long-horizon priors and governance constraints. These are described as the Embodied Controller, Supervisory Controller, and Meta-Controller. The separation is intended to make conflicts between immediate utility and longer-term viability inspectable.

SELECTION BY UTILITY, COHERENCE, AND COST

At the supervisory layer, a proposed Resonance Engine compares candidate frames using expected utility, coherence with higher-order priors, and execution cost. The aim is not to eliminate optimization, but to stop a single short-term objective from silently becoming the whole system’s purpose.

ICS AND TIME-INTEGRATED COHERENCE

The paper proposes an Interoceptive Control Signal (ICS) for fast feedback and Time-Integrated Coherence (TIC) as a state variable that can gate policy complexity and option availability. Both are formal constructs to be instrumented and tested; neither is presented as a finished safety guarantee.

WHAT THE CURRENT EVIDENCE DOES—AND DOES NOT—SHOW

Version 5 provides definitions, algorithmic sketches, falsifiable hypotheses, proposed simulation and human-subject protocols, governance considerations, and a reproducible toy instrumentation demo. The demo shows how selector variables can be logged and visualized. It is illustrative, not empirical validation of the architecture.

READ THE CURRENT PAPER

ConsciOS v5 is available as an open preprint on Zenodo under CC BY 4.0. Read the manuscript, inspect the supporting code, and evaluate the proposal on its formal claims rather than the older marketing language around it.

SEE THE WHOLE.
FIND THE CONSTRAINT.
MOVE WHAT MATTERS.

YOUR NEXT DECISION IS ALREADY OMITTING SOMETHING. SEE IT FIRST.

SYSDOM AI
SYSTEMS INTELLIGENT AGENTIC VENTURE BUILDER

SEE THE WHOLE. FIND THE CONSTRAINT. MOVE WHAT MATTERS.

BEFORE YOU ADD ANOTHER AGENT, FIND WHAT’S ACTUALLY STUCK.

Sysdom.ai