Skip to main content

Classification

A label applied once should be honored everywhere. System9 has one classification ladder, a single scale with a single dominance comparison, rather than a different notion of sensitivity in the catalog, the query engine, the message bus, and the device. Most platforms enforce classification in the catalog and hope nobody reaches the engine another way.

In this area​

One ladder, one comparison​

Sensitivity is expressed on a single five-level ladder, public, internal, confidential, secret, top secret, plus compartments for need-to-know. A single normalization function and a single dominance comparison are the source of truth; several competing vocabularies were retired in favor of this one ladder, and a guard test fails if anyone tries to add a level to it.

Classification and clearance ride on identity and its delegations: a delegation carries a clearance level and a compartment set, and, like every other axis, they can only narrow as authority is delegated downward (see Authorization and delegation).

Applied everywhere​

The same ladder is the vocabulary for information flow across the platform: object storage carries a classification as a database constraint; streams enforce that ingest level does not exceed the namespace level and that query clearance meets the data level; query engines compile row filters and column masks into the plan; messages carry the same classification as a table; and the same ladder is enforced at the model-inference boundary and on the device. The information-flow rules are no-read-up and no-write-down, enforced through the permission resolver.

Core and edge​

The classification ladder is defined identically on both runtimes, and the Edge Agent enforces classification, compartment, and data-grant constraints on every edge query. Data does not leave its classification boundary: context attached to an on-device model is filtered through data grants before it reaches the model.

Design status and terminology

Classification is described by mechanism. The vocabulary and dominance comparison are unified and authoritative; some at-rest enforcement consumers are not fully wired, and are not claimed here as running. This information classification ladder is distinct from the tactical protected-object categories used for rules-of-engagement, and from a device's self-classification of its own operating mode.