plan-experiment

To turn a mechanistic interpretability research idea into a structured experiment plan, you crystallize the objective into a RESEARCH_OBJECTIVE.md and build a PLAN.md defining the analysis dependency DAG. This establishes pre-flight gates and per-node I/O contracts for downstream execution.

An experiment DAG structures causal abstraction research by mapping the analysis chain from baseline to locate, subspace, and activation geometry checks. It defines explicit dependencies between nodes, measurable pre-flight gates, and I/O contracts to ensure downstream runners execute reliably.

You plan experiment sweeps and cache reuse by defining runner configuration strategies within the session plan. This approach materializes configs for downstream execution while reducing redundant compute by explicitly mapping cache reuse dependencies across the analysis DAG.

Yes, you need an active session to generate a research objective and dependency DAG. The process reads and writes session plan artifacts in the plan/ directory, resolving the active session to produce RESEARCH_OBJECTIVE.md and PLAN.md without overwriting existing plans.

The best way to avoid overwriting existing session plans is to resolve the active session before generating new artifacts. The planning process specifically targets the plan/ directory and avoids overwriting existing RESEARCH_OBJECTIVE.md and PLAN.md files within that session.