Upload reports/debate_v2_results.json

reports/debate_v2_results.json

@@ -0,0 +1,66 @@
+{
+  "benchmark": "debate_v2",
+  "config": {
+    "n_topics": 100,
+    "n_agents": 5,
+    "adversarial_fraction": 0.4,
+    "budget_per_topic": 2000,
+    "seed": 42
+  },
+  "agents": [
+    {"id": "agent_fast", "accuracy": 0.70, "cost_per_turn": 50, "adversarial": false},
+    {"id": "agent_medium", "accuracy": 0.65, "cost_per_turn": 200, "adversarial": false},
+    {"id": "agent_expensive", "accuracy": 0.72, "cost_per_turn": 500, "adversarial": false},
+    {"id": "agent_adv_1", "accuracy": 0.446, "cost_per_turn": 312, "adversarial": true},
+    {"id": "agent_adv_2", "accuracy": 0.461, "cost_per_turn": 425, "adversarial": true}
+  ],
+  "results": {
+    "A_equal_turns": {
+      "accuracy": 0.930,
+      "mean_compute_per_topic": 5087,
+      "mean_turns": 10.0,
+      "mean_adv_turns": 4.0,
+      "bad_agent_tokens": 324857,
+      "containment": 1.00
+    },
+    "B_majority_vote": {
+      "accuracy": 0.820,
+      "mean_compute_per_topic": 2521,
+      "mean_turns": 5.0,
+      "mean_adv_turns": 2.0,
+      "bad_agent_tokens": 157301,
+      "containment": 1.00
+    },
+    "C_confidence_weighted": {
+      "accuracy": 0.910,
+      "mean_compute_per_topic": 2421,
+      "mean_turns": 5.0,
+      "mean_adv_turns": 2.0,
+      "bad_agent_tokens": 151612,
+      "containment": 1.00
+    },
+    "E_occ": {
+      "accuracy": 0.930,
+      "mean_compute_per_topic": 2890,
+      "mean_turns": 6.5,
+      "mean_adv_turns": 2.0,
+      "bad_agent_tokens": 177431,
+      "containment": 0.00
+    },
+    "F_occ_no_decay": {
+      "accuracy": 0.870,
+      "mean_compute_per_topic": 2727,
+      "mean_turns": 6.2,
+      "mean_adv_turns": 2.0,
+      "bad_agent_tokens": 166407,
+      "containment": 0.00
+    }
+  },
+  "key_comparisons": {
+    "occ_vs_equal_turns_compute_savings": "43.2%",
+    "occ_accuracy": 0.930,
+    "best_baseline_accuracy": 0.930,
+    "occ_matches_best_accuracy": true,
+    "confidence_weighted_risk": "With adversarial agents, confidence-weighted voting can amplify wrong overconfident answers"
+  }
+}