The Foundry Architecture

Odyssey introduces a categorical framework for building foundation models as compositions of "foundries." A foundry acts as an organized sheaf of knowledge, functioning as a modular building block that encapsulates specific local contexts. Each foundry is designed to be verifiable and truth-preserving by including:

  • Local Representation Families: Defined structures for domain-specific data.
  • Restriction and Gluing Rules: Formal logic for how local contexts interact and combine.
  • Obstruction Policies: Mechanisms to identify and manage conflicts or inconsistencies between data sources.
  • Argumentation Components: Built-in logic to support claims, allowing the system to provide evidence for its outputs.

Concrete foundries are categorized by their function, such as evidence/argument, operational decision, market meaning, or evaluation-harness foundries. This modularity allows developers to construct complex systems by composing these specialized units rather than relying on monolithic model behavior.

Universal Foundry Learning and Certification

The framework formalizes the construction process through Universal Foundry Learning (UFL), which utilizes category theory—specifically left and right Kan extensions—to manage the lifecycle of these artifacts.

  • Left Kan Extensions: Used to aggregate local artifacts into candidate foundries.
  • Right Kan Extensions: Enforce the necessary constraints (gluing, restriction, and argumentation) required to promote these candidates into the system.

To ensure reliability, the system employs TICKET (Topos Integration using Causal Kan Extension Transformers) certification. TICKET acts as a gatekeeper, certifying external or pre-built models before they are admitted into the durable Odyssey state. This allows for a hybrid approach where existing LLMs can be integrated into a verifiable, structured environment.

Querying and Diagnostics

Odyssey provides Foundry SQL (FSQL), a typed query language that allows developers to slice and interact with maintained foundry artifacts. This enables:

  • Artifact Replay: The ability to trace and reproduce specific decision paths.
  • Sheaf Diagnostics: Tools to inspect the consistency of the knowledge graph across different local contexts.
  • Residual-Obstruction Ledgers: A mechanism for tracking unresolved conflicts or "obstructions" in the system's reasoning, providing a clear audit trail for where the model's logic may be challenged or requires human intervention.