Concept Atlas

Distillation is the active extraction of the essential nature, causal relationships, and structure of information, and its integration into a persistent mental model. The process is irreversible — the result is understanding, not a recoverable encoding. Distillation may or may not reduce volume; what matters is that essence is captured. It is related to but distinct from compression, which optimizes for faithful reconstruction rather than understanding.

Knowledge is stored not as a flat list or neural weights, but as a structured, multi-dimensional space. Models connect to other models through explicit interfaces, forming a relational topology. This structure allows for non-hierarchical, multi-way relationships.

Learning is the act of integrating new distilled models into the existing world model. It requires finding the "connection points" or interfaces where the new concept links to what is already known. Integration operates within the context of the active working model, which pre-loads guidance about relevance and likely connection points. New knowledge is classified by type (confirmatory, extending, dissonant, alien), and the complementary question of what it does to the model is characterized by its integration effect (see #10).

The World Model is the persistent, comprehensive truth (too big to hold in active focus). A Working Model is distilled out of the world model for a specific context; this generation process is itself a form of distillation, extracting what's relevant from the larger structure. Working models come in two flavors: operational (stable, efficient, discarded when done) and learning (living, expanding, restructuring as understanding deepens). In humans, working model assembly is instinctual: you walk into a situation, spend a few minutes ramping up, and the relevant pieces pull themselves together from whatever regions of your understanding apply. Insights from working models flow back through the integration pipeline.

Knowledge should be accessible at any level of granularity, from high-level principles to deep mechanical details. The structure is self-similar across scales.

The ability to distill is bounded by the maturity of the current world model. As the model grows, the system must return to source materials to extract deeper layers of meaning that were previously invisible.

The validity of incomplete understanding. When distilling complex information, unknown parts are preserved as "dangling endpoints"—interfaces that don't connect to anything yet, waiting for future knowledge.

A structured holding area for information of uncertain validity. Quarantine allows systems to work with unverified information without corrupting validated knowledge. It operates at two levels:

• Session Quarantine (lower bar, local to individual interactions)
• Central Quarantine (higher bar, gateway to the persistent world model)

Quarantine is triggered primarily by source provenance, not just content coherence—information from unverifiable sources enters quarantine regardless of how plausible it appears.

Grounding is the foundational connection between abstract knowledge structures and external reality. Without grounding, a model, no matter how internally coherent, is "floating" with no reference to what it represents. Grounding operates at three levels:

• Reality Grounding (connection to the physical/observable world, the most concrete and tractable level)
• Identity Grounding (a persistent operational frame of reference: the system's awareness of its own state, capabilities, and boundaries)
• Social Grounding (understanding through interaction with other agents, building on reality and identity grounding)

This addresses the classical symbol grounding problem: how do symbols acquire meaning beyond their relationships to other symbols? For embodied intelligences, grounding emerges naturally through sensorimotor interaction. For disembodied systems, grounding is the fundamental challenge. In practice, grounding likely spans a spectrum: simulated environments providing structured grounding during training, passive real-world access through sensors and data feeds, and full embodiment providing the richest connection. These approaches are not mutually exclusive.

Integration Effects describe what distilled knowledge does when it meets the world model — the nature of the interaction, not just what the knowledge is. Six primary effects are identified:

• Illuminate (reveal existing but invisible connections)
• Catalyze (trigger latent reorganization)
• Reinforce (strengthen existing structure)
• Ground (anchor abstract knowledge to reality)
• Agitate (create productive tension)
• Nucleate (seed entirely new understanding where none existed)

The effect is a property of the interaction between distillate and model state — the same knowledge might illuminate one model and nucleate in another. Effects propagate through the model's topology via resonance, where the primary effect at the point of contact triggers secondary effects of potentially different types at structurally connected nodes.