CAGE 4 Released Policies - Failure Tracker Check Off: leaderboard

CAGE 4 Released Policies - Failure Tracker Check Off: Mean episode reward (higher better; ceiling 0; no finite floor). CAGE 4 defender audit arXiv 2611.14068v1: released KEEP graph-PPO five-agent team (cybermonic/cage-4-submission 2afd652d80ce, aggregate checkpoint MD5 55fb3e7f46540eeea63230b149e1744e) and H-MARL Expert (adityavs14/Hierarchical-MARL 6fe960f5931f, fifteen policies+algorithm state, MD5 fd9e9172fd56b723fedca72ce0e78a16). These are trained released cyber-defense specialists, not generic LLMs; Meta is a different artifact and is not aliased to Expert. One hundred common environment seeds, 500-step episodes; shipped KEEP wrapper records 499 actions. Actions sampled per shipped scripts. Mean episode reward higher better, ceiling 0, no finite theoretical floor. Action-validity check on/off and adversary stay separate native configurations. Re-trained/internal continuation arms, human supervision and deterministic reference rules are held; duplicate corpus baselines are not recounted.

Metric: Mean episode reward (higher better; ceiling 0; no finite floor). Source: arxiv.org. 1 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=cage-4-released-policies-failure-tracker-check-off · How It Works · Data refreshed daily, snapshot 2026-11-12.