Predictive Ledger

What this is.Each row is the answer to "if you held out a sample of this customer's history and asked Aito to predict it, how often would Aito be right?". Accuracy is the share of held-out predictions Aito got right; baseline is what an always-predict-majority classifier would get; gain is the difference; geomP is the geometric mean of Aito's predicted probability for the correct answer (higher = more confident calibration).

Running 4 cross-validations in parallel against Aito… this takes 15–30s on first load.

The Predictive DB

_evaluateOperator

LivePredictions

0Records

IndexedModel

_evaluate (cross-validation)

Aito's _evaluate runs leave-one-out cross-validation on a held-out test sample. One row per prediction task: GL code, approver, payment matching, help-article click. Each task's row shows accuracy, the baseline (always-predict-majority), and the gain.

Example query

{
  "testSource": {
    "from": "invoices",
    "where": {
      "customer_id": "CUST-0000"
    },
    "limit": 50
  },
  "evaluate": {
    "from": "invoices",
    "where": {
      "customer_id": "CUST-0000",
      "vendor": {
        "$get": "vendor"
      },
      "amount": {
        "$get": "amount"
      }
    },
    "predict": "gl_code"
  }
}

Verify yourself

Every claim on this page traces back to an Aito query. No separate model file — Aito predicts directly from its index.

View live schema (JSON) Query API reference

Note: GL codes (4100, 5300, 6200…) are illustrative — real Finnish Liikekirjuri uses different numbering. Chosen for demo readability.

Learn more

API reference: _evaluate

Start free trial →