2603.16021v2

Core Extraction (Thesis & Synthesis)

The Interpretable Context Methodology (ICM) replaces code-level multi-agent orchestration frameworks with structured filesystem hierarchies to coordinate sequential, human-in-the-loop AI workflows. Rather than managing agent context passing, memory, and error handling through complex application code, ICM represents workflow stages as numbered directories containing declarative markdown contracts (CONTEXT.md). These contracts explicitly define inputs, processes, and outputs, allowing a single AI agent or compiler-like orchestrator to execute the pipeline by reading and writing files sequentially.

By organizing the workspace into a five-layer context hierarchy, ICM scopes context delivery to the model at each stage, preventing the cognitive degradation and token waste associated with monolithic prompts. The architecture separates persistent reference constraints (Layer 3 “recipes”) from per-run working artifacts (Layer 4 “ingredients”), ensuring the model receives pre-sorted context. Surfacing intermediate outputs as plain text files creates natural review gates, making human course-correction cheap and establishing an inherently observable and version-controllable (Git-compatible) system.

Source Grounding

• Source Document: 2603.16021v2.pdf
• Exact Citation: Local Reference

Semantic Connections

• Links:
  • Concept Anchors:
    • Interpretable Context Methodology: A filesystem-based context orchestration protocol replacing programmatic multi-agent frameworks.
    • Context Engineering: The software discipline of structuring and delivering optimal inputs to a model’s context window.
    • Mixed-Initiative Systems: Human-AI collaboration models prioritizing shared state, transparency, and explicit intervention gates.
  • Relational Connections:
    • Interpretable Context Methodology optimizes Context Engineering by isolating stage-specific inputs to keep context windows under 8,000 tokens.
    • Interpretable Context Methodology supports Mixed-Initiative Systems by rendering all intermediate model outputs as editable markdown files on disk.

Inquiry & Speculation

• How does the filesystem latency and synchronization overhead of ICM compare to memory-based state managers in highly concurrent environments?
• What are the limitations when attempting to translate non-linear, dynamic branching agent logic into folder-based execution namespaces?
• Can we build semantic compiler tools that trace outputs back through folder paths to pinpoint exactly where an instruction diverged from source data?

References

• Interpretable Context Methodology (ICM) GitHub Repo
• Anthropic Model Context Protocol (MCP)