Once the two emotional states have drifted apart, is the head-to-head still a fair comparison?

Only on the first turn. Each side reads its own current eight axes before it judges, and the affectus agent reports its emotional movement every turn, so from the second verdict onward the two judges are no longer seeing the same input. The README is explicit that fairness is judged on the first verdict from a shared starting state, and reset puts both sides back to their initial values.

Will it burn gateway calls while I sit and read the page?

No. The send button is the only execution point: pressing it runs judging, emotional-state updates and reply generation together. There is no auto-send path, so nothing reaches the API until you submit, and each subsequent utterance is typed into the same field and sent the same way.

Do I need Go and a dependency install to get the affectus binary running?

No. The distributed zip includes the affectus binary and node_modules, so there is no Go toolchain and no remote dependency resolution. You do need macOS (Apple Silicon or Intel), Node 24 or newer, and a Vercel AI Gateway API key written into .env.local by setup.sh. Port 8787 is the default and can be changed with PORT, and the server prints its actual URL at startup.

If I hand the packaged zip to someone else, whose API key pays for their run?

Theirs, by default. The zip excludes your key and the emotional state, and setup.sh walks the recipient through writing their own key into .env.local. Passing --with-key bundles .env.local unchanged and renames the file, which means the recipient's usage is charged to your key, so that flag is only safe when you can restrict who gets the file.

Which model generates the replies on both sides?

Both sides use the same Haiku 4.5 model, the same personality, and the same instructions. Only the emotional state passed in differs, so any difference in tone comes from the judging step.

What exactly is sent to each judge on a turn?

Each side receives its own current eight emotional axes plus the user's utterance, because the affectus agent reads its state before declaring movement. From the second turn on, the two inputs have diverged.

What does the total judgment difference number mean?

It sums the absolute difference across the eight axes. Zero means the two judges agreed exactly, and the maximum is 16.

CommunityJev Demos & Experiments

Jev Dev

macOS demo that runs one utterance through both TypeSafe jev and an LLM judge, accumulates each verdict as a separate affectus emotional state, and shows the split as Plutchik petals, per-axis deltas and two replies generated from a single fixed persona.

See highlights

Open github.com

Stars: 0
Language: TypeScript
Published by: Community
Author: n-yokomachi
Last verified: Oct 2026

n-yokomachi/jev-dev

User repository on GitHub

View on GitHub

同じ発言を jev と LLM の両方に判定させ、感情の変動値のズレと応答速度を1画面で見比べるデモ（affectus + Vercel AI Gateway）

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

The same utterance goes to jev and to an LLM judge; the left panel is the LLM, the right is jev, and each side fills in as it returns.
Per-axis comparison shows the signed |LLM - jev| gap for all eight axes, sign carried by color, with a total where 0 is exact agreement and 16 the ceiling.
Both replies use the same model, persona and instruction, so the only difference is the emotional state each judge produced.
Persona is friendly or contrarian, applied identically to both sides, switchable any time before you send, and never given to the judge itself.

Watch out

Calls only happen when you press the send button: judging, emotional-state updates and reply generation all hang off that one action, and there is no automatic path. The comparison is only fair on the first verdict, because each side reads its own current eight axes before judging, so once the two states have diverged the inputs diverge too. Judging and reply generation are timed separately and both sides generate with Haiku 4.5, so only the judging figure is a usable speed comparison. The JSON panel shows judging input and output only, not reply generation. You need macOS on either Apple Silicon or Intel, Node 24+, and a Vercel AI Gateway key; the zip ships the affectus binary and node_modules, so Go is not needed. The state directory holding the two emotional states is created by setup.sh, and reset returns both sides to their initial values. scripts/package.sh builds a zip that excludes your API key and state, but the --with-key flag bundles .env.local as-is, which means whoever receives that zip is billed against your key. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

laya-server

Self-hosted Docker service and web UI that packages Laya structured-decision inference behind a compatible System One endpoint.