Skip to content

Architecture

PSE wraps a capable open-weight model in a constraint architecture. The model proposes; the structure disposes.

Every response is generated inside the constraint

w_s + w_o + w_f = 1

where the weights are the attention allocated to self, other and field, each held within a formally bounded healthy range. The Care Invariant is where PSE starts, not the whole of it: it defines a healthy balance, and the layers below are what hold the system to it. States outside that range — self-referential loops, sycophancy, context collapse — are treated as drift to detect and correct, not merely discouraged.

Governs every generation step against the invariant.

Evaluates inputs before they can perturb the constraint state and checks outputs against it before release. Its bounds are designed to limit how far any single input can move the system.

Detects drift toward pathological interaction patterns (self-absorbed, over-compliant, unstable, withdrawn) before it surfaces in output and responds structurally. Escalation paths end with a human.

The full derivation is in the PSE v3.0 master paper.