Jev directory
CommunityJev Models & Reimplementations
Connect AI Company v2 (RL Decider Teacher)
A PPO-trained reinforcement-learning decision brain for the CONNECT TYCOON company simulator from CONNECT AI LAB: it runs staffing, inventory, marketing, refund and security judgments and answers in the JEV request format with probabilities.
- Reading time
- 2 min
- Published by
- Community
- Author
- Charlottej23
- Last verified
- Oct 2026
- Platforms
- CLI
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- PPO policy runs a simulated company for 20 business days, judging staffing, inventory, marketing, support, refunds and security.
- Trained in two stages: supervised from a rule-based boss, then PPO with rewards from the company's profit, satisfaction and risk.
- Card's 40-month averages: profit 3.24M to 11.57M KRW, satisfaction 79% to 96%, risk incidents 7.8 to 2.6, stockouts 10.7 to 0.4 days.
- Its refund example answers full with probability 0.9775 alongside partial, deny and human probabilities and a confidence of 0.9775.
- Runs in the game via connect-tycoon play --brain hf:Charlottej23/connectailab-rl-decider-v2-teacher; Apache-2.0.
Quickstart
bash
pip install "connect-tycoon[game]"
connect-tycoon play --brain hf:Charlottej23/connectailab-rl-decider-v2-teacherWatch out
Card reviewed 2026-10-17; the improvement figures are self-reported averages over 40 unseen company months in a synthetic simulator, and the card itself states this is an educational model that does not guarantee or predict real business outcomes. It is a reinforcement-learning policy rather than a language-model readout, and the card is in Korean.
Questions about this entry
Short answers drawn from the linked project or page.
A PPO-trained decision policy that acts as the prefrontal cortex of a simulated company in CONNECT TYCOON, run by CONNECT AI LAB. It handles staffing, inventory ordering, marketing, customer support, refunds and security, and answers every judgment in the same format as TypeSafe Jev.
You send a request with a model, a state such as a refund case and typed questions with options; it returns the choice plus probabilities, confidence and a reason. In the card's refund example it picks full with probability 0.9775 and confidence 0.9775.
In two stages: first imitation of a rule-based boss through supervised learning, then PPO, with rewards derived from the company's stated values of profit 1.0, customer satisfaction 1.0 and risk 1.0. A Colab notebook for training is linked from the card.
Averages over 40 unseen months versus the rule-based boss: monthly profit 3.24M to 11.57M KRW, satisfaction 79% to 96%, risk incidents 7.8 to 2.6, human-review escalations 37.8 to 15.4 and stockout days 10.7 to 0.4. These are simulator results on a synthetic company.
Install with pip install "connect-tycoon[game]" and start the game with connect-tycoon play --brain hf:Charlottej23/connectailab-rl-decider-v2-teacher; the card also shows loading it in Python with ct.CompanyBrain.from_hub.
It judges staffing, inventory orders, marketing, customer center, refunds and security for a 20-business-day month. The card defines the company's values as profit 1.0, customer satisfaction 1.0 and risk 1.0, which drive the PPO reward.
Charlotte (JOOEUN KIM) of CONNECT AI LAB, with the CONNECT AI LAB and AI CITY BUILDERS teams; Apache-2.0. The card is in Korean and calls it an educational model on a synthetic simulator that does not guarantee or predict real business outcomes.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
X: Jev picks which Apple HIG guides go into the agent's context. NatashaTheRobot built a small AskHIG CLI and…
X: Three decision models, one game of Snake
More Jev resources like this
An Apache-2.0 decision engine with a TypeSafe-compatible endpoint, document answers with source passages, and a ViZDoom demo that picks actions from structured game state.
Jev Models & Reimplementations#community#python#local
Community
deepanwadhwa/OpenDecisionPython
deepanwadhwa/OpenDecisionPythonAn 8-bit MLX conversion by SirSahOl of chaoliangUNSW/Jev-Style-0.8B-Decision-v3 built for Apple Silicon native GPU inference: about 990 MB of active VRAM, a 262,144-token context and an average of 8.25 bits per weight.
Jev Models & Reimplementations#community#open-models#local
Community
A local yes/no decision model that speaks Jev's wire format on a laptop CPU, refusing Choice and Score until those question types ship.
Jev Models & Reimplementations#community#python#open-models
Community
feder-cr/jevC++
feder-cr/jevC++A community Q8_0 GGUF of Mapika/decider-0.8b, the smallest System One typed-decision model: one forward pass returns calibrated Choice, Score and noul answers, reporting 0.405 on JevBench hard-public at 920 MB resident.
Jev Models & Reimplementations#community#open-models#local
Community
Steerix's robo-jev v0.7 merges a robotics LoRA into Qwen3.5-2B as a 2B System 1 that answers physical-judgment questions from a robot's scene; the card reports 0.81 physical judgment against TypeSafe Jev's 0.28.
Jev Models & Reimplementations#community#open-models#robotics
Community
A Q8_0 GGUF of knowlet's Gemma-4-12B-Unified-System-One with a native System One decision runtime and an independently fitted GGUF temperature, shipped with an F16 mmproj for the model's native image and audio projections.
Jev Models & Reimplementations#community#open-models#quantized
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.