{
  "paper": "2609.23718",
  "dataset": "KuaiRand-Pure",
  "diagnostic_only": true,
  "schema_version": 2,
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": false,
    "claim_policy": "scaled public-data mechanism diagnostic, not paper-fidelity or production evidence"
  },
  "provenance": {
    "dataset_fingerprint": "sha256:c814bf6f3624c0cfae83c57de3df26b2ed206e5c57bab4c4dcbfabbabe20cbf0",
    "artifact_path": "docs/reproductions/2609.23718-unique/metrics/kuairand-pure-seeds42-44.json"
  },
  "setup": {
    "source_rows_considered": 500000,
    "retained_exposures": 107850,
    "items": 7583,
    "train": 16279,
    "validation": 1709,
    "test": 1917,
    "steps_per_variant": 300,
    "evaluation_impressions_per_seed": 400,
    "code_retrieval_positives_per_seed": 100,
    "split": "per-user last two chronological impressions held out"
  },
  "seeds": {
    "42": {
      "unique_scaled": {
        "test_ctr_auc": 0.5184403578,
        "test_code_recall_at_4": 0.12,
        "test_reranked_recall_at_10": 0.01,
        "code_perplexity": 59.88076019,
        "mean_candidate_pool": 681.9
      },
      "without_joint_generation": {
        "test_ctr_auc": 0.5144960305,
        "code_perplexity": 60.01227951
      },
      "without_balance": {
        "test_ctr_auc": 0.5159280474,
        "test_code_recall_at_4": 0.33,
        "test_reranked_recall_at_10": 0.01,
        "code_perplexity": 24.56116676,
        "mean_candidate_pool": 2218.26
      }
    },
    "43": {
      "unique_scaled": {
        "test_ctr_auc": 0.5388431417,
        "test_code_recall_at_4": 0.07,
        "test_reranked_recall_at_10": 0.0,
        "code_perplexity": 60.51030731,
        "mean_candidate_pool": 704.5
      },
      "without_joint_generation": {
        "test_ctr_auc": 0.5469265443,
        "code_perplexity": 60.78816986
      },
      "without_balance": {
        "test_ctr_auc": 0.5200070509,
        "test_code_recall_at_4": 0.44,
        "test_reranked_recall_at_10": 0.0,
        "code_perplexity": 20.44973373,
        "mean_candidate_pool": 2933.11
      }
    },
    "44": {
      "unique_scaled": {
        "test_ctr_auc": 0.5237961855,
        "test_code_recall_at_4": 0.06,
        "test_reranked_recall_at_10": 0.0,
        "code_perplexity": 60.48347855,
        "mean_candidate_pool": 561.48
      },
      "without_joint_generation": {
        "test_ctr_auc": 0.5232869038,
        "code_perplexity": 60.42392349
      },
      "without_balance": {
        "test_ctr_auc": 0.523337832,
        "test_code_recall_at_4": 0.22,
        "test_reranked_recall_at_10": 0.02,
        "code_perplexity": 24.7651844,
        "mean_candidate_pool": 1081.57
      }
    }
  },
  "interpretation": "No robust CTR or reranking improvement; balanced code usage reduces candidate-pool size but also lowers four-code recall on this slice.",
  "manifest_ref": "reproduction:unique"
}
