Jev Lab · Hundred & Shogi
Choose legal next actions for town NPCs and shogi
Implementation and evidence →JEV GUIDE
A Jev-powered NPC can start with one bounded question: which action comes next? Game code still owns maps, collisions, resources, legal moves and outcomes; the model chooses among currently permitted candidates.
Jev Lab includes the Hundred town and a shogi application, with a local Rules mode available by default and engine-generated shogi candidates. A running game does not establish that Jev was called or its strategy evaluated.
HEIST//ONE gives Jev local observations, candidate targets and tactical choices while deterministic code maintains the world and checks invalid or stale decisions. We reviewed its implementation materials without running the game to evaluate behavior.
Source: HEIST//ONE · pinned source ↗
Editorial flow: extract what the NPC can observe, generate legal actions with IDs and send a decision request with a state revision. When it returns, recheck the target, resources and revision. Discard or reconsider an action that is no longer valid.
Synthetic example: a guard requests a decision while a door is open, but it closes before “go through” returns. Recheck the path; model selection must not bypass collision handling. A timeout may trigger a predefined wait or patrol action rather than blocking the game loop indefinitely.
Editorial recommendation: compare rule-based and live Jev modes on matching maps, seeds and scenarios. Label cached replays separately. Record requests, rejected actions, fallbacks and end-to-end latency, then ask playtesters to assess behavioral coherence.
Choose request frequency around behavioral changes rather than rendering frames. Before adding NPCs, measure actual traffic and cost, including bursts and service outages. This guide makes no benchmark claim about cost, win rate or latency.
Choose legal next actions for town NPCs and shogi
Implementation and evidence →Make stealth-game guard judgments observable
Implementation and evidence →02 / JEV GUIDE
Build a Jev API request with state, typed questions and Bearer authentication. Download a JSON fixture and review response checks.
Read guide →03 / JEV GUIDE
Write focused Jev questions with original feedback-routing examples, explicit criteria and a practical edge-case checklist.
Read guide →14 / JEV GUIDE
Compare Jev and generative LLM task boundaries through rewriting and script-review cases. Typed output does not establish correctness or performance.
Read guide →