The Roundtable Context Window Test (RCWT) is a controlled protocol designed to measure the task-budget displacement effect caused by placing coordination content within the finite prompt of multi-agent and memory-augmented LLM systems.

  • RCWT varies coordination content while controlling total budget, position order, task family, and scoring.
  • At a context window of 4096 tokens, commercial models degrade sharply when residual reference evidence falls to a few hundred tokens.
  • An ablation test keeping the full task block present shows GPT-4.1-mini, Claude Haiku 4.5, and Gemini 2.5 Flash return correct results up to a 95% coordination ratio.
  • The results indicate the performance cliff is due to task-budget displacement rather than semantic interference from coordination volume alone.

RCWT serves as a measurement primitive for context-allocation budgeting in LLM calls.