Read With Jev

The project calls itself a reading instrument for inspecting model output and explicitly not a validated emotion map. Nothing in it is a benchmark submission. Scores are whole-sentence association suggestions, not extracted evidence spans, character attributions or intensity, and colour is never the only carrier of meaning.

A key is only needed for highlighting. Set TYPESAFE_API_KEY in .env.local and the reader requests paid emotion analysis for sentences near the viewport as you scroll; without one it still loads, the threshold dial is disabled and no sentence is highlighted. The key never reaches the browser.

The paste-in reader at /your-text takes up to 20,000 characters in CodeMirror and keeps analysing nearby sentences as you scroll, with an overview that counts only completed sentences. Guests keep a draft in that browser tab; with an account configured, texts autosave privately and raw guest text never enters the disk cache.

It was removed so the deploy would not need a build-time download of a third-party lexicon, or the licensing questions that come with redistributing one. NRC is still used as a baseline comparator in the research pipeline, reachable through the affect:prepare script, and no lexicon is deployed.

library/reader-selection.json is an explicit editorial allow-list rather than a scrape. Collections, plays, poetry, nonfiction and incomplete editions stay archived and are never published or sent for cast extraction, and only one edition per work is published: a second Gutenberg edition of a title already in the catalogue is excluded as a duplicate, npm run library:prepare enforces both rules, and it fails on a duplicate work. Rebuilding never restores an excluded title.

Node.js 22 or newer and the GitHub CLI, since GitHub downloads go through gh while Hugging Face downloads use Node fetch. No Python is needed for the reader. Then npm ci, npm run library:prepare, npm run data:fetch, npm run data:prepare and npm run dev, and open http://127.0.0.1:3000. Downloads are pinned to exact revisions, checksummed and kept in the ignored data/raw/, and preparation verifies sources before writing data/processed/.

A guest allowance and a shared cache. The server signs a persistent guest identity cookie and atomically reserves one of 1,000 calls for each new JEV request, and cache hits are free, so the shared reader cache is what keeps a popular chapter from billing repeatedly. If the quota store cannot be reached, new guest analysis pauses rather than running unbounded. Clearing the cookie resets the identity, so it is a practical usage limit rather than a hard one, and the cookie secret should be a random value of at least 32 characters.