jevbooks

← All projects

jev-compaction

Waxmell114514/jev-compaction · Homepage

Score-only context compactor: kept lines verbatim, elided lines recoverable via expand().

Scoring & Ranking90%Runner-up: Data Labeling & CurationQuestionnaireFlight recorder

What it is

A standalone context compactor for agents. It scores transcript segments with Jev; kept lines stay byte-for-byte original, low scorers move to a store behind an expand() pointer, and the append-only frozen prefix keeps the prompt cache valid.

How it uses Jev

Jev scores each transcript segment in one request. High scores enter the frozen prefix; low scores are relocated to a store with an expand() pointer. The agent can call expand() to retrieve original bytes. Decisions are logged for later threshold replay.

Primitives:score

Technique worth stealing

Score-only gating: Jev never writes text, so kept lines are always verbatim originals and elided content is recoverable.

Try it

git clone, uv venv, uv pip install -e '.[dev]', then .venv/bin/python demo.py (offline, no key).

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricyes0.87
  • Shows a System One patternno0.25
  • Handles uncertaintyunclear0.44
  • Measuredyes0.96
  • Runnableyes0.92
  • Worth recommendingyes0.62
  • Model replicaunclear0.44
  • Problem scopescore on a 0–2 scale1.54
  • About Jevyes0.86

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 22 Sept 2026.