Thesis
Once a system has been identified, its fundamental behavior becomes visible.
For language models, the central property is stochasticity.
Language models do not produce deterministic outputs. Sampling, temperature, and latent structure introduce variance even when inputs remain identical.
This randomness is not an implementation detail.
It is the defining characteristic of the system.
Reliable architectures therefore treat language models as stochastic cores surrounded by deterministic infrastructure.
Formal Definition
A stochastic operator T maps an input domain Din to a distribution over output domain Dout. For a given input x, the output T(x)is not a point but a probability measure over possible completions.
The system is reliable when the support of that distribution lies within an admissible region — not when the output is deterministic.
Mathematical Framing
T: D_in → P(D_out)
Reliable iff supp(T(x)) ⊆ A(x) ∀x ∈ D_in
T = stochastic operator | P(D_out) = probability measures over output domain | A(x) = admissible region
Noise, Disturbance, and Bounded Regions
In control systems, noise and disturbance are modeled explicitly — not ignored. The same discipline applies here. LLM output variance is a measurable quantity. Bounded variance within an admissible region is the definition of reliable stochastic behavior.
Unbounded variance is not "creativity." It is loss of control.
Tool-Calling as Structural Separation
Tool-calling separates the stochastic core (LLM reasoning) from the deterministic periphery (execution). The model proposes; the tool executes. This is not a convenience pattern — it is a control architecture. The stochastic component is bounded by the deterministic one.
Why It Matters
Every system that treats LLM output as deterministic will eventually fail in production. Modeling the stochastic core explicitly is prerequisite to building anything reliable on top of language models.