Architecture and method

One identity. One answer path. Replaceable intelligence.

Lumen is being simplified around a single principle: the user owns the relationship, while the model remains an interchangeable capability.

Updated September 29, 2026

The four layers

Identity

Memory, preferences, permissions, corrections, and relationship continuity.

Brain

A model chosen for the available compute and the capability required.

Body

Mac and iPhone today; glasses and robotics are future forms, not current claims.

Constitution

Authority, evidence, safety, and receipts that persist through model changes.

The One-Door collapse

Earlier versions accumulated duplicate routes, checks, and fallback behavior. The collapse removes those competing paths in favor of one judge, one verdict, and one commit spine. The goal is not architectural elegance for its own sake; it is a more dependable answer for the person using Lumen.

Current measurements

49 / 53Representative answer-quality checks passing in the current candidate, up from 48 / 53 at baseline.

7,350 → 1,895System-prompt tokens carried into each turn, a 74% reduction.

11,298Tests in the first strict egress run, with zero violations.

≈3.5h → ≈1.5hDeployment time while retaining required checks.

What these numbers do not prove

They do not prove that every answer is correct, that every privacy risk is solved, or that Lumen is ready for broad release. They show the current candidate is becoming simpler without giving up the measured quality and enforcement properties under test.

Next acceptance evidence

After the collapse is accepted, the next product proof is an independent second-owner deployment. The website will publish that outcome only after the deployment is running and the result can be described honestly.