FFN Gate Sparsity as a Query-Level Correlate of Knowledge Availability: A Length-Controlled Re-analysis Across Eight Language Models

This paper replaces six earlier preprints (TES-1 to TES-6, March 2026). A re-analysis of all archived per-prompt data found that the Gate Sparseness Index (GSI) depends strongly on prompt length, which the original analyses did not control, and several pipeline errors affected results. With length controlled, the difference between answerable and unanswerable queries remains in 7 of 8 models with smaller effects; the main claims of TES-1 and TES-2 survive in revised and narrowed form, and the main claims of TES-3 to TES-6 are retracted. A later pre-registered replication on a new, length-matched question set in 18 models found GSI worse than first-token entropy, or without signal beyond length, on plain prompts in every model. GSI is described as a query-level correlate of knowledge availability, not as a detector. Every changed claim is listed with its data source.