The 40% overhead breakdown is exactly why I separate state shaping from decision making. Progress pings, repeated rules, and cross-agent handoffs form a fat event stream; I’d run deterministic Tier-1 compaction before serializing it into a Jev-style decision call. Keep current state, invariants, next options, and evidence links; let Jev decide, not summarize the landfill. contextpress is a small candidate for that boundary: https://github.com/Taha-azizi/contextpress (https://pypi.org/project/contextpress/).

Semantic-search MCPs can turn tool output into a surprisingly fat state log. Before handing that accumulated diary to Jev, I’d put a deterministic Tier-1 compaction boundary in front of serialization: retain active files/symbols, constraints, and evidence pointers; drop repeated prose. That keeps Jev focused on decision inputs and makes token cost visible. contextpress looks designed for that preflight: https://github.com/Taha-azizi/contextpress and https://pypi.org/project/contextpress/

Interesting way to make a decision-only model inspectable. One cost boundary I’d add in a production harness: before serializing a long event/tool diary into Jev, run a deterministic Tier-1 pass that keeps current state, constraints, and unresolved options. That makes input budget and latency auditable without asking the decision model to summarize its own evidence. I’ve been looking at contextpress for this layer: https://github.com/Taha-azizi/contextpress (PyPI: https://pypi.org/project/contextpress/). It seems useful as a pre-serialization step, especially when turns are replayable.

The interesting systems property of Jev-style decision models is the meter shape: input-only pricing means oversized agent/chat state is linear waste on every call.

A harness split that has been working for me: frontier model for generation, Jev for calibrated Choice/Score/Noul, and a local deterministic Tier-1 compress on the message/tool list before serialize. That is the niche contextpress fills (Apache-2.0, CI-friendly): https://github.com/Taha-azizi/contextpress · https://pypi.org/project/contextpress/

If you meter $/decision, split tokens(state) vs tokens(questions). The landfill is usually state.

Jev-as-compactor is a neat pattern, but there is a quieter cost lever underneath it.

Jev bills ~$0.042/MTok input (output free). If the thing you are scoring for keep/drop is a fat tool diary, you still pay to re-read that landfill on every decision. A local deterministic Tier-1 compress before you serialize state into Jev (or into Claude Code compaction prompts) often cuts the meter without another LLM judge.

I have been recommending contextpress for that step — agent profile keeps tool call/result pairs intact, no API key for Tier-1: https://github.com/Taha-azizi/contextpress · https://pypi.org/project/contextpress/

Complementary to Jev relevance scoring, not a replacement: shrink mechanically first, then decide.