Architecture & DAG Traversal

Deterministic Subgraph Traversal

When an LLM agent requires external context, CirIAco does not execute a naive cosine similarity search across isolated chunks. Instead, it extracts causally complete subgraphs.

       [ Atmo A: Foundational Theorem ]
                  ▲
                  │  (Cryptographic Lineage)
       [ Atmo B: Algorithmic Variant ]
                  ▲
                  │
       [ Atmo C: Empirical Benchmark ]

Context Packing Algorithm

1. Seed Identification: Identify entry Atmo nodes via dense vector similarity over propositions.
2. Upward Traversal: Walk parent hashes to gather all necessary axiomatic foundations until the token budget is reached.
3. Downward Expansion: Include immediate validated corollaries if context window space permits.
4. Serialization: Emit a compact machine-readable serialization (YAML/JSON) directly into the model’s context window.

Zero-Trust Privacy Boundary

CirIAco enforces strict boundaries when operating with third-party LLMs:

• Proprietary entities and PII are obfuscated via deterministic SHA-256 hashes (devlacruz-hashlib).
• The model reasons over opaque symbols (<NODE_7f8a9...>), preserving intellectual property and privacy while keeping relational semantics intact.