Not a chatbot — an AI that decides

JK Move AI

JK Move AI turns “what next?” into structured questions answered in one pass, returning calibrated probabilities — judgement, not generated text. The game's decision model is JK-Move-AI-one, following the decision contract defined by JEV; inference runs in real time on Jacky Kit's own GPU server (currently on an open-source Laya decision-model checkpoint) behind a parallel worker pool.

The JEV model: the foundation

JK Move AI is built on the JEV model as its complete conceptual foundation. Its purpose is real-time decision-making: explainable, calibrated judgements within milliseconds to seconds, while the situation is still moving — not after-the-fact analysis.

At JEV's core is “decisions, not conversations”: give it a situation and a question template, and it reads once and answers every question in parallel — each answer typed, bounded and carrying its own confidence; when confidence falls short, the verdict turns from auto to review.

Response time depends on how many questions and answers are asked: a single yes/no question is not the same as a game council of a dozen-plus questions, each with several options. In JK's in-depth testing the self-built decision model answers in under 100 ms at its fastest — it judges instead of generating text token by token, an order of magnitude more efficient than a large language model.

This Three Kingdoms game is JEV's high-speed closed-loop lab. Every season runs a full lap: observe the board → turn it into questions → decide → act → resolve simultaneously → log every probability and outcome → feed learning.

• Real time: questions fan out to parallel inference workers, and the AI finishes its council while you plan.
• Predictable latency: total ≈ questions × cost per question ÷ parallel workers. The game currently runs on CPU with 6 parallel workers, so a lord's dozen-plus questions take a few seconds a season — the live numbers are in the AI lab.
• Closed loop: each decision is logged with its outcome (odds vs result) for calibration and self-play training.
• Observable: the AI lab shows latency, confidence, raw AI and rule prior live — every step of the loop is visible.
• Controllable: personality thresholds, difficulty temperature and rule fallback keep real-time decisions robust under uncertainty — the game's version of JEV's auto / review verdict.

Learn about the JEV decision model (JK AI Hub) →

How it plays Three Kingdoms

1. 1
   Read the board

   A rule layer condenses the board into a short brief and generates candidates: reachable fronts, every army's legal orders.

2. 2
   Strategy questions

   Each AI lord asks several questions per season: posture and main front (choice), aggression (score), truce / alliance / betrayal / stratagem (yes-no).

3. 3
   Operations questions

   Given that posture, each key army gets its own question — “best order for Zhang Liao's army?” — over its candidates.

4. 4
   Confidence blend

   The model returns calibrated probabilities and a confidence. Final = confidence × AI + (1 − confidence) × rule prior: a sure AI leads, an unsure one defers.

5. 5
   Personality & difficulty

   The same probability means different things to different lords: benevolent Liu Bei betrays only above 85%, cunning Cao Cao at 45%. Higher difficulty samples more decisively.

6. 6
   Thinking ahead

   AI orders depend only on the start-of-season board, so they deliberate while you plan — ending a season barely waits.

7. 7
   Your advisor

   The same model answers from your side of the board: suggestions with confidence, adoptable in one click.

8. 8
   The storyteller

   Rules rank the season's key events and an on-site language model writes them up as a Romance-style chronicle (template text when busy).

9. 9
   The AI lab

   Every question, probability, latency and source, live. Enemy councils stay scrambled until the season resolves — unless your spies read them.

Three question types

• Choice: a probability per option — posture, front, orders.
• Score: a graded level — aggression from passive to all-in.
• Yes / no: one probability compared with the ruler's threshold — betray or not.

Every decision is logged as data for future self-play training, so the AI can keep getting stronger.