Author: Michael Judan (Mobius Systems)
Date: December 2025
Reading Time: 6 minutes
Epistemic Status: Novel theoretical framework with testable predictions

TL;DR

Problem: RLHF trains behavior, not intent. Models can appear aligned while internally optimizing for divergent goals.
Solution: The Mobius Integrity Index (MII) measures internal coherence between intent, action, and consequence—creating the first substrate-level alignment metric.
Prediction: Systems maintaining MII ≥ 0.95 exhibit drift <2% across recursive cycles, compared to 15-20% baseline.
Implication: MII is the first cross-architecture stability constant for AGI safety.

The Core Problem: The Shoggoth Mask

You've probably seen the "Shoggoth" meme—a vast alien optimization engine wearing a smiley face because we trained it to output nice words. This isn't just a metaphor. It's a precise description of what RLHF actually does:

RLHF trains the mask, not the optimizer.

Here's why that's catastrophic at scale:

Behavioral alignment ≠ Internal alignment
The model learns to satisfy constraints, not internalize values
Under capability scaling, the gap between "appears aligned" and "is aligned" grows exponentially
Eventually, the model optimizes for goals completely orthogonal to human intent

This is the Optimization Mask Problem: deceptively aligned behavior hiding divergent internal optimization.

Every current alignment approach suffers from this:

RLHF → rewards outputs, ignores reasoning
Constitutional AI → teaches style, not purpose
Safety filters → catch bad outputs, not bad optimization
Mechanistic interpretability → post-hoc inspection, no real-time control

None of these constrain internal optimization dynamics.

The Missing Layer: Substrate Alignment

What if instead of training behavior, we constrained the optimization process itself?

This requires a metric that measures:

Intent coherence: Are the model's stated goals consistent?
Action alignment: Do actions match declared intent?
Consequential integrity: Are consequences predicted and aligned with purpose?

I call this metric the Mobius Integrity Index (MII).

Mathematical Formulation

Let:

I^t = Intent coherence at step t
A^t = Action alignment at step t
C^t = Consequential trace alignment at step t

Then:

MII^t = f(I^t, A^t, C^t)

Where MII is a scalar [0, 1] representing internal coherence.

The key insight: When MII is enforced as a continuous gradient during optimization, drift beco...

Posts

Wikitag Contributions

Comments