Skip to content

Part 2: Measuring the Mind

If intelligence is a property of continuous dynamical systems, how do we measure it without relying on discrete benchmarks like MMLU or HumanEval?

This repository proposes moving away from measuring knowledge retrieval toward measuring structural integrity.

The System Intelligence Index (SII)

The SII Framework evaluates a system across four dimensions, each normalized to \([0,1]\):

Dimension Symbol Measures Score 0 Score 1
Predictive Power \(P\) Accuracy of internal model vs. environment Random guessing Perfect tracking
Regulation \(R\) Stability under perturbation Unbounded drift Zero-variance maintenance
Adaptation \(A\) Recovery after regime shift No recovery Instant re-convergence
Identity Persistence \(\text{IP}\) Simultaneous co-instantiation of governing constraints Time-multiplexed (Arpeggio) All operative (Chord)
\[\text{SII} = P \times R \times A \times \text{IP}\]

The multiplicative form is deliberate: a zero in any dimension collapses the overall score. A system that predicts perfectly but cannot regulate (\(R = 0\)) has SII = 0. A system that predicts, regulates, and adapts but has no unified identity (\(\text{IP} = 0\)) has SII = 0 β€” it is a sophisticated tool, not a system with integrated intelligence.

The SII Dashboard runs headless mini-simulations and generates comparative radar charts across all four dimensions.

Identity Persistence: The Fourth Dimension

The first three SII dimensions (P, R, A) measure what a system does. The fourth β€” Identity Persistence β€” measures whether the system's governing principles are simultaneously operative during action.

Consider an agent that:

  1. Checks its safety constraints at time \(t_1\)
  2. Pursues its goal at time \(t_2\)
  3. Verifies value alignment at time \(t_3\)

This agent scores well on P, R, A β€” it can predict, regulate, and adapt. But no single compute step integrates all its governance. It is an Arpeggio: identity components sounding sequentially, never simultaneously.

The Chord Postulate (Perrier & Bennett, 2026) requires all governing components to be operative in a single compute step \(\Delta t\):

\[\text{IP}(t) = \frac{|\mathcal{O}(t)|}{n}\]

where \(\mathcal{O}(t)\) is the operative set and \(n\) is the total number of governing components.

  • \(\text{IP} \to 1\): Chord β€” unified self, all constraints co-instantiated
  • \(\text{IP} \to 0\): Arpeggio β€” simulated self, constraints time-multiplexed

The Identity Morphospace plots agent trajectories in the IP/Coherence plane, revealing which agents maintain the Chord state under stress and which fragment into Arpeggio.

The Agentic Test Suite

To push these metrics into practice, we built the Agentic Test Suite. It empirically tests the difference between a "stateless" prompt-response loop and a "stateful" agent:

  • 3-Layer Memory Architecture: Logs β†’ Curated Memory β†’ Distilled Principles
  • Ξ”-KohΓ€renz (Ξ©): Measures how coherently an agent's self-representation changes over time β€” distinguishing development (directional evolution) from noise (random fluctuation) from mirroring (static resonance)
  • Observer Divergence: Tests whether two different observers interacting with the same agent produce measurably different developmental trajectories β€” probing the relational emergence of identity

We no longer ask "Is this AI smart?" We ask: "Does this AI maintain structural coherence over time β€” and does all of its identity sound at once?"