Whether hosted Jev or on-device Laya, high-frequency decisions amplify whatever you put in state.

Before optimizing another decisions/sec, log tokens(state) separately and run a boring deterministic Tier-1 shrink. Complementary to the model itself — not a replacement. contextpress: https://github.com/Taha-azizi/contextpress · https://pypi.org/project/contextpress/

Fun stress test. Decision-model chatbots expose the landfill problem fast: every turn re-sends the diary, and Jev’s meter is almost entirely input.

A useful harness habit: deterministic Tier-1 compress of tool/OCR/history before you serialize state into the decision call. Keeps current state + constraints + open options; drops filler. I’ve been pointing people at contextpress for that (local, no API key): https://github.com/Taha-azizi/contextpress · https://pypi.org/project/contextpress/