Breakout Transfer Validation v1
policy
{
"name": "breakout_policy_v1",
"logic": "RAM/Go heuristic tracks ball x against paddle x and fires on reset."
}
go_shadow_eval
{
"episodes": 2,
"score_sum": 18.0,
"score_mean": 9.0,
"score_min": 5.0,
"score_max": 13.0,
"steps_sum": 948,
"video_urls": [
"https://file.aimusic.win/gemma/breakout/20260510T105850Z_ep000_seed0.mp4",
"https://file.aimusic.win/gemma/breakout/20260510T105850Z_ep001_seed1.mp4"
],
"rows": [
{
"timestamp": "20260510T105850Z",
"trial_index": 1,
"policy_name": "breakout_policy_v1",
"seed": 0,
"episode": 0,
"score": 13,
"steps": 508,
"done": true,
"replay_path": "breakout/runs-rom-fidelity-v2/replays/20260510T105850Z_ep000_seed0.jsonl",
"video_path": "breakout/runs-rom-fidelity-v2/videos/20260510T105850Z_ep000_seed0.mp4",
"video_url": "https://file.aimusic.win/gemma/breakout/20260510T105850Z_ep000_seed0.mp4"
},
{
"timestamp": "20260510T105850Z",
"trial_index": 1,
"policy_name": "breakout_policy_v1",
"seed": 1,
"episode": 1,
"score": 5,
"steps": 440,
"done": true,
"replay_path": "breakout/runs-rom-fidelity-v2/replays/20260510T105850Z_ep001_seed1.jsonl",
"video_path": "breakout/runs-rom-fidelity-v2/videos/20260510T105850Z_ep001_seed1.mp4",
"video_url": "https://file.aimusic.win/gemma/breakout/20260510T105850Z_ep001_seed1.mp4"
}
]
}
atari_backtest
{
"episodes": 2,
"score_sum": 9.0,
"score_mean": 4.5,
"score_min": 2.0,
"score_max": 7.0,
"steps_sum": 3600,
"video_urls": [
"https://file.aimusic.win/gemma/breakout/rom_fidelity_v2_atari.mp4"
],
"rows": [
{
"episode": 0,
"score": 7.0,
"steps": 1800,
"first_reward_step": 40,
"reward_count": 7,
"done": false
},
{
"episode": 1,
"score": 2.0,
"steps": 1800,
"first_reward_step": 40,
"reward_count": 2,
"done": false
}
]
}
judgment
{
"verdict": "partial_transfer",
"go_score_mean": 9.0,
"atari_score_mean": 4.5,
"atari_to_go_score_ratio": 0.5,
"interpretation": "The policy earns positive Atari reward, so the shadow-env optimization transfers partially.",
"next_recommendation": "Run forced-loss trace to confirm lives/serve RAM, then use transfer results to tune survival and late-rally dynamics."
}