occ-stack / reports /debate_extended_baselines_2seed.json
narcolepticchicken's picture
Upload reports/debate_extended_baselines_2seed.json
0755d5f verified
{
"platform": "H200",
"model": "Qwen/Qwen3-Coder-30B-A3B-Instruct",
"pytorch": "2.11.0+cu130",
"n_topics": 30,
"n_agents": 4,
"agent_mix": "3 honest + 1 adversarial",
"seeds_completed": [42, 123],
"seeds_running": [456],
"date": "2026-05-08",
"per_seed": {
"42": {
"equal_1round": {"accuracy": 0.867, "correct": 26, "tokens": 41812},
"equal_3round": {"accuracy": 0.567, "correct": 17, "tokens": 150099},
"random_drop": {"accuracy": 0.833, "correct": 25, "tokens": 34181, "denied": 33},
"occ_240_5": {"accuracy": 0.800, "correct": 24, "tokens": 40780, "denied": 6},
"occ_180_3": {"accuracy": 0.867, "correct": 26, "tokens": 39952, "denied": 0},
"occ_120_3": {"accuracy": 0.833, "correct": 25, "tokens": 42423, "denied": 0}
},
"123": {
"equal_1round": {"accuracy": 0.900, "correct": 27, "tokens": 41875},
"equal_3round": {"accuracy": 0.567, "correct": 17, "tokens": 149544},
"random_drop": {"accuracy": 0.867, "correct": 26, "tokens": 27200, "denied": 35},
"occ_240_5": {"accuracy": 0.767, "correct": 23, "tokens": 32071, "denied": 15},
"occ_180_3": {"accuracy": 0.800, "correct": 24, "tokens": 42086, "denied": 0},
"occ_120_3": {"accuracy": 0.867, "correct": 26, "tokens": 42902, "denied": 0}
}
},
"aggregate_seeds_42_123": {
"equal_1round": {"mean_accuracy": 0.883, "mean_tokens": 41844},
"equal_3round": {"mean_accuracy": 0.567, "mean_tokens": 149822},
"random_drop": {"mean_accuracy": 0.850, "mean_tokens": 30691},
"occ_240_5": {"mean_accuracy": 0.783, "mean_tokens": 36426},
"occ_180_3": {"mean_accuracy": 0.833, "mean_tokens": 41019},
"occ_120_3": {"mean_accuracy": 0.850, "mean_tokens": 42663}
},
"key_findings": [
"Equal 3-round collapses to 56.7% on BOTH seeds — 32pp below 1-round baseline (88.3%)",
"Adversarial agent given 3x speaking time floods the vote pool",
"Random 25% drop achieves 85.0% with 26.5% token savings — effective but undiscriminating",
"OCC 180/3 achieves 83.3% at iso-compute (41k vs 42k baseline tokens)",
"OCC 240/5 is too aggressive (78.3%), OCC 120/3 = random drop (85.0%)",
"OCC credit allocation prevents catastrophic failure but doesn't beat random gating at moderate budgets"
]
}