Jev scores retrieval chunks: 1.02M tokens cut to 86K in context
Jev Score rates 1,146 retrieved chunks from 0 to 3, dropping 1,009 and keeping 41 whole, cutting the tokens the model reads by 92% with the same answer.
Jev Engineering is what stops your agent from paying a frontier model to read 1,009 files it will never cite. one turn pulled 1,146 chunks. 1.02 M tokens of grep hits, file reads and web pages for a single question. a Score runs on every chunk before the context window is built: → 3 names the symbol, keep it verbatim → 2 on topic, keep it → 1 adjacent, compress to one line → 0 unrelated, never send it 1,009 scored 0. 41 went in whole. the model read 86 K instead of 1.02 M and gave the same answer. so instead of: retrieve → stuff everything → summarise → hope you get: retrieve → score → keep, compress or drop → one small context → cite 92 percent of the raw tokens never arrive, and nothing is deleted: a chunk that scored 0 is still addressable, it just did not earn a seat this turn. that is the point of Jev Engineering: the decision about what the model reads is an engineering layer, not a vibe. full breakdown in the article below ↓
This post quotes another post: x.com/hanakoxbt/status/2101709924828934222