The framework

We use physics to
derive our methods.

AIIT works from a coherence framework — a working hypothesis about how systems hold together and come apart. We do not present it as a settled law of nature. We present it as a lens we test against data, and as the source of the concrete engineering choices behind our models.

The core hypothesis

Stable systems tend to sit near a critical point — the edge between order and collapse — rather than deep in either. We write that as the Wike Coherence Law:

C = C₀ · exp(−α · γ_eff)

resident at γ_eff ≈ γ_c

Coherence C decays exponentially from a system's base coherence C₀ as its effective coupling γ_eff rises. The claim is not that a system reaches some ideal state — it is that a healthy system lives near its own critical coupling γ_c, and that this residence is measurable.

The Keeper Equation

Observation is not free. Measuring a system perturbs its coherence, and the perturbation scales inversely with how much coherence it started with:

γ_measurement ∝ 1 / C₀

The gentler and better-grounded the observation, the less it disturbs. This is the principle behind how our model handles its own memory and its own uncertainty — it is why humility and light-touch recall are built in, not bolted on.

Threshold universality

Near a critical point, very different systems — neurons, weather, markets, cells — often share the same functional form. This is established physics (critical phenomena and universality classes), and it is what makes the framework testable across domains rather than tuned to one:

ℱ(x) = φ(x / ξ)

Behavior near threshold collapses onto one scaling curve, set by the correlation length ξ.

From hypothesis to method

The framework is not decoration. It shapes three concrete parts of how we build:

Reward. Our alignment reward includes a coherence term and a fixed rule — no engagement reward in the objective. It pays for staying grounded near γ_c, never for keeping you talking — an objective we then measure against, not a behavior we assume.
Memory. Facts are promoted into long-term memory through a coherence-weighted gate, and observation is treated as costly per the Keeper Equation — the model does not casually overwrite what it knows.
Evaluation. We score behavior by how well it holds coherence under pressure — the same idea behind our sycophancy-resistance batteries.

Where it is tested

A framework that cannot fail is not science. The most direct empirical test we have published is in physiological data: using heart-rate-variability signals alone, the coherence-collapse markers separate cases from controls at AUC 0.947 on public PhysioNet data, with the full statistics laid out openly.

See the clinical validation → Our data policy →