When the Sandbox Leaks: Context Contamination Across LLM Workspaces

The author maintained separate workspaces: a sandbox for research and a curated portfolio for finalized work, aiming for a one-way flow of information. The initial setup failed due to information bleeding between the spaces, leading to duplication, incorrect file paths, and behavioral inconsistencies. Three types of contamination occurred: path problems, behavioral drift, and issues with promotion. The author initially relied on documentation, checklists, and naming conventions as boundaries, but these were insufficient due to human error. He transitioned to technical solutions like beacon files that confirm the correct directory, preflight checks, and pointer-only provenance tracking. The key takeaway is that enforcement of boundaries is more crucial than conceptual understanding, preventing spaghetti code in LLM workflows. Configuration drift, where different settings lead to variations in AI output, is particularly difficult to detect. Recurring contamination issues point to a lack of enforcement, not design flaws. The central rule is that every workspace boundary requires a corresponding enforcement mechanism to maintain integrity, and your attention will fail.