Claims meet qualification.

XERXES is designed to test itself. Automated diagnostic frameworks, live operator notes, deterministic packaging, and preserved regression baselines turn the studio into a falsification environment rather than a demonstration environment.

The major historical signal is unambiguous: a complex-cognition baseline that once passed only 2/30 was driven to 40/40 through generalized architectural repairs—ordered instruction representation, dependency-based context inheritance, specialist ownership correction, domain collision handling, capability self-modeling, and sequential operation verification—rather than prompt patches.

Ordered multi-step cognition now compiles and executes sequential operations without the earlier silent step-loss failures. This is a necessary foundation for reliable program construction and revision under explicit cognitive control.

What has been demonstrated is listed in the evidence set. What remains experimental is also listed. What remains to be tested is the long comparative research queue. The distinction between mechanism qualification and comparative benefit is maintained throughout. Standalone cognition without an LLM is explicitly classified as RESEARCH: increasing portions of control, routing, reasoning organization, and correction are independent of an LLM, but matching the full breadth and learned world knowledge of a frontier model without one is not claimed.