Jev Case Studies
Prompt libraries show you what to ask. Case studies show you what happened when someone asked. Every case here ships with the metrics, the trade-offs, the negative results — and the prompts to reproduce the pattern. Community metrics are quoted, not re-run: the source eval is always linked.
Our own self-run experiments (starting with transaction risk scoring) join this section as they are measured. Reader cases welcome — same rules apply.
Numbers, not adjectives
Every case reports concrete metrics on defined data — plus what did not work.
Methodology disclosed
Dataset shape, eval protocol, model versions and question design are written down.
Vendor claims separated
TypeSafe-reported numbers are labeled as such; case numbers come from published evals.
Cases
Agent memory reranking: Jev vs a free local cross-encoder
A community project measured Jev as a memory reranker: Recall@1 went 0.41 → 0.62 with rigorous statistics — and they publish the three graded tests where answers did NOT improve. What Jev actually bought was context compression: 2 memories ranked by Jev answered as well as 5.
Reading all of arXiv for $0.06 a day: Jev as a paper filter
One Noul question per plain-English interest, per paper. A community project validated it against real reviewers: screening 19,447 records from four unseen Cochrane reviews reproduced 96.9% of the studies they included, while removing 78% of the reading — and the misses are published too.
Have a Jev case of your own?
We publish reader-submitted cases with the same rules: real data, disclosed method, negative results included. Send yours →