Compiled Knowledge

What compilation adds on top of raw observation: recovered structure, resolved identity, derived rollups and risks.

Raw observation says a file exists. Compiled knowledge says this file defines a class that extends that one, and 14 modules depend on it.

Beyond per-source recovery, compilation derives:

• Concentration risks — any object with three or more real DependsOn dependents becomes a Risk object. Observed only, with no inferred severity.
• Subsystem rollups — hierarchical summaries of modules and crates.
• Data lineage links — from sources through transformations to sinks.
• Architecture knowledge — crate topology, roles (LLM-classified, stored as reviewable claims), an evidence-coverage score, drift over time.
• Optional descriptions — LLM-written overviews (ai_overview, ai_usage) grounded in evidence and stored as properties, opt-in via [llm-description].

Anything an LLM produced is marked as such; deterministic and inferred knowledge are never mixed silently.