siege
SIEGE: 200 people vs one agent.
What it does
SIEGE: 200 people vs one agent. A typed action gate (TypeSafe System One) that learns from every breach, evaluated by W&B Weave, hardened by a defender loop. Built at CoreWeave Hacks: Agent Loops 2026. Playable proof that discrete Jev choices beat joystick LLM improv.
Fun demo, serious latency lesson. Tags in the wild: games, agents, eval. Skim the repo for setup, then open the demo when you want proof before you merge anything.
How Jev is used
TypeSafeGate sends Choice(instructions, criteria=policy) through system_one on each tool call; probabilities feed the war room UI and training loop for the defender agent.
Setup: Clone repo, follow README for WANDB_API_KEY optional tracing, open vnmoorthy.github.io/siege for the arena demo.
Games as security education
SIEGE makes fast typed decisions visible on a phone-friendly UI while crowds try to break a policy. The defender rewrites gate criteria each round, so sub-300ms Choice calls matter.
FAQ
- Which primitive powers the gate?
Choice with policy criteria in gate.py TypeSafeGate, model default jev-latest unless TYPESAFE_MODEL overrides.
- Where can I play?
https://vnmoorthy.github.io/siege/ hosts the cinematic arena linked from the README.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.