Jev Experiments — dabit3

Throwaway by design: Devin-built latency experiments with benchmarks on seeded corpora, each with its own README and TESTING.md. Study the pattern rather than depending on the code.

turbo-rerank: 50 BM25 candidates judged by 50 score questions plus one no-good-answer noul in a single 170 ms Jev request, lifting top-1 accuracy from 50% to 100% on a 40-query benchmark.

jev-firehose: a simulated Twitch-scale chat at 300 messages per second, each judged by six Nouls and one Choice in 125 ms median, with a simulate-slow-LLM toggle showing why 2-second latency never drains the queue.

The in-monorepo launcher copy is archived. Development moved to dabit3/launcher, which is the repository to use for current code and setup instructions.

Roughly 20 latency-focused demos in one monorepo: moderation firehoses, rerankers, inbox triage, macOS launchers and closed-loop control simulators. Each lives in its own top-level directory with a README, TESTING.md and screenshots, and apps include jev-firehose, turbo-rerank and jev-launcher.

Each demo is self-contained. The turbo-rerank example is `cd turbo-rerank`, `npm ci`, `export TYPESAFE_API_KEY=...`, then `npm run dev` and `npm run bench`. Benchmarks were measured on seeded corpora, so the latency and accuracy figures are observations rather than guarantees.

The repo's own firehose run cost $37.86/hour at $0.042 per 1M input tokens, about 835 input tokens per message with free output. That is a measurement from this demo on its seeded corpus, not a general price for moderation.