dataset overlap audit derived

{
  "input": {
    "body": {
      "key": [
        "customer_id"
      ],
      "datasets": [
        {
          "csv": "customer_id,label\nC-001,1\nC-002,0\nC-003,1\nC-004,0\n",
          "name": "train"
        },
        {
          "csv": "customer_id,label\nC-003,1\nC-005,0\n",
          "name": "evaluation"
        },
        {
          "csv": "customer_id,label\nC-006,1\n",
          "name": "holdout"
        }
      ]
    },
    "type": "http",
    "method": "POST",
    "bodyType": "json"
  },
  "output": {
    "type": "json",
    "example": {
      "pairs": [
        {
          "first": "train",
          "second": "evaluation",
          "examples": [
            {
              "key": "C-003",
              "row_in_first": 4,
              "row_in_second": 2
            }
          ],
          "share_of_smaller": 0.5,
          "shared_key_count": 1
        },
        {
          "first": "train",
          "second": "holdout",
          "examples": [],
          "share_of_smaller": 0,
          "shared_key_count": 0
        },
        {
          "first": "evaluation",
          "second": "holdout",
          "examples": [],
          "share_of_smaller": 0,
          "shared_key_count": 0
        }
      ],
      "limits": {
        "maxRows": 5000,
        "maxListed": 200,
        "maxColumns": 64,
        "maxCsvBytes": 262144,
        "maxDatasets": 10,
        "maxFindings": 500,
        "maxCategories": 1000,
        "maxKeyColumns": 4
      },
      "passed": false,
      "complete": true,
      "datasets": [
        {
          "name": "train",
          "rows": 4,
          "distinct_keys": 4
        },
        {
          "name": "evaluation",
          "rows": 2,
          "distinct_keys": 2
        },
        {
          "name": "holdout",
          "rows": 1,
          "distinct_keys": 1
        }
      ],
      "disjoint": false,
      "findings": [
        {
          "code": "DATASETS_OVERLAP",
          "detail": "train and evaluation share 1 key(s), which is 50.00% of the smaller set. If these are meant to be disjoint splits, anything measured across them is optimistic.",
          "dataset": "train",
          "other_dataset": "evaluation"
        }
      ],
      "request_id": "example-request",
      "service_id": "dataset_overlap_audit",
      "assumptions": [
        "Keys are canonicalised exactly, so an identifier written as 42 in one split and 42.0 in the other is the same key and the overlap is found.",
        "Overlap is reported per pair as well as across all datasets, because two splits leaking into each other is a different problem from one record appearing everywhere.",
        "Share is expressed against the smaller set, since that is the one whose measurements are compromised."
      ],
      "limitations": [
        "Key overlap between the datasets supplied. Whether near-identical records with different keys are duplicates is not assessed here; duplicate_key_audit does that within one dataset.",
        "disjoint is true only when every dataset was fully read and no pair shared a key."
      ],
      "observed_at": "2026-10-02T00:00:00.000Z",
      "not_examined": [],
      "dataset_count": 3,
      "finding_count": 1,
      "service_version": "1.0.0",
      "findings_by_code": {
        "DATASETS_OVERLAP": 1
      },
      "keys_in_every_dataset": 0
    }
  }
}