Context compression looks free on completion metrics and isn't.

Across three models and two environments over a 24-turn horizon, 5x compression produced no statistically significant change in task completion (arXiv 2608.16370): but all six model/regime comparisons showed more retrieval calls, five significant after correction. GPT-5.5 completion nominally rose 80% to 85% (p=1.0) while retrieval calls went 21.0 to 63.9 (p=.002). ALFWorld showed no surge, so it's environment-dependent. Instrument tool-call counts before you tune a compaction threshold; the tokens compaction saves come back as re-fetches.