The sovereign decision layer — the code decides, the model writes
With the thinking model in place, the architecture followed: a sovereign decision layer in code, separated from the speaking layer. Explainable, versionable, and safe against model change.
Once we accepted that decisions do not belong to the model, the architecture shifted. Agents moved from "thinkers with language" to "meters with a strict protocol". On top came a sovereign decision layer: ordinary code, with explicit rules, versions, tests and an audit trail.
Why sovereign
Sovereign here means: independent of any external model. The decision layer determines whether a deviation is a signal, how heavy it weighs, and which evidence accompanies it to the user. If tomorrow a different model is wired in — or no model at all — the answer remains identical. The model is an interchangeable translator, not a judge.
How it works
- Rules as code. Every decision rule sits explicitly in a repository, with a name, a version and a test. Adding a rule is a pull request, not a prompt tweak.
- Evidence mandatory. A rule can only say "signal" if it simultaneously holds the evidence: which source, which record, which moment. No evidence, no signal.
- Two-phase publication. New rules first run in shadow alongside the existing ones. Only when they perform as well or better on historical data do they go to production.
- Modelless fallback. If the model is unavailable, the decision layer keeps running and publishes the report with neutral, template-based explanation. No answer disappears.
The speaking layer
The speaking layer receives three things: the conclusion, the evidence, and the audience (controller, signatory, board). From that the model writes a single paragraph in the right register. It may add nothing that is not in the evidence bundle, and omit nothing the decision layer marked essential.
The real lesson
This split gave us something we had not expected: calm. Model updates do not touch the conclusion. Rule changes are traceable and reversible. New checks take days, not weeks. And in an audit we do not point to a prompt — we point to a rule with a version number and a test.
THE LESSON Outsourcing decisions to a model loses both control and speed. A sovereign decision layer gives both back.