Outputs

Every pass emits the full bundle: producing all outputs is one pass, so you always get all of them. A quick run returns the bundle for the tables you sent; an endpoint object carries the bundle computed over its pinned data; a serve returns it for the request data, scored by the endpoint; the held-out block is on the endpoint object. Field-by-field shapes are in the

A pass either returns the full bundle or fails as job_failed. Slices scoped by contract (the cross-table map on one table, the prediction slice under target.mode: "none") are documented null, never absent by failure.

Sector

Vertical-agnostic sector identification: the industry sector of each table, from cell values alone, any domain, no metadata required. top1 { name, confidence } and top5[], each entry with its own confidence. Blind to column names: it holds with headers stripped or coded. Per-model results on blind sector identification are in the model registry. Use it to catalog a data lake, route incoming files, or triage a dataset before anyone opens it.

Column profile

Per column: name, role, role_confidence, type, missing_pct, pii.

Use role and pii to route, mask, or gate columns in your own systems.

Cross-table map

Across any number of tables from any number of sources (two or more), the bundle relates the sources with no shared key, no lookup table, and no name matching:

• column_alignment[]: column pairs that carry the same attribute, each with attribute, confidence, and a verbatim example from the cells (555-0142 and +1 555 0142 under one phone).
• entity_matches[]: row pairs that are the same record (surfaced as entity matches), with confidence and matched_on, the aligned attributes that agreed for that pair.
• unified_schema { shared, a_only, b_only }: the proposed fold.
• unified_rows: the map materialized into joined records from the original cell values, one per matched entity, with provenance per column and shared_variants showing each source's raw value. When shared variants disagree beyond formatting, that flags a possible mis-alignment rather than silently picking one side. total is the full count; rows is a sample, paginated by cursor.

Alignments and matches are reported above your options.confidence_threshold. Every alignment and every match ships its confidence and its evidence (example, matched_on); the map is a proposal for your review, never auto-applied: act on high-confidence matches automatically and queue the rest at thresholds you choose.

On an endpoint, pin two or more datasets and the endpoint's bundle carries the map over the pinned data; a serve scores the request tables against the pinned context, and the held-out split is by matched entity. Per-model results for keyless entity matching and name-blind schema matching are in the model registry.

Missing-value imputation

filled[] lists each reconstructed cell (table, row, column, value, method); confidence is row-level, not per cell, in row_confidence[] (table, row, confidence). Numeric and categorical gaps are both filled; a categorical fill is a class label, not the most common value. Identifying columns (pii: true) are never filled.

Predictions

Three task types, all on the current base, all in-context: predictions use the labeled rows present (the request's, or the endpoint's pinned data). Nothing trains.

{
"prediction": {
  "target": "cards_db.card_brand", "task_type": "classification", "tier": "in_context",
  "classes": ["Amex", "Mastercard", "Visa"],
  "dropped_rows": { "missing_target": 0 },
  "results": [ { "row": 0, "label": "Visa", "confidence": 0.94,
                 "probabilities": { "Amex": 0.01, "Mastercard": 0.05,
                                    "Visa": 0.94 } } ]
}
}

{
"prediction": {
  "target": "credit_limit", "task_type": "regression", "tier": "in_context",
  "results": [ { "row": 0, "value": 12480.0, "median": 12010.0, "std": 3120.0,
                 "quantiles": { "0.1": 8400.0, "0.5": 12010.0, "0.9": 17900.0 } } ]
}
}

{
"prediction": {
  "task_type": "anomaly", "threshold": 0.90,
  "results": [ { "row": 0, "anomaly_score": 0.97, "is_anomaly": true },
               { "row": 1, "anomaly_score": 0.12, "is_anomaly": false } ]
}
}

Anomaly runs unsupervised with no target, on explicit request only: fraud screening, equipment monitoring, data-quality checks with no labeled history.

Missing targets. Rows whose target is missing are counted in dropped_rows.missing_target; a missing value never enters classes and is never a predictable label.

Two tiers. A quick run prediction carries held_out: null; an endpoint's current held-out score is on the endpoint object and in the report.

Target and task selection

Both are first-class on quick runs and on endpoint creation, with the same shapes; both default to auto and both explain themselves in the response.

"target": { "mode": "auto" }     // smart-target: The model picks and explains
"target": { "column": "churn" }  // you pick; "table.column" on multi-table input
"target": { "mode": "none" }     // understanding-only: no prediction slice

"task": { "mode": "auto" }       // categorical -> classification, numeric -> regression
"task": { "type": "anomaly" }    // explicit only; needs no target

The response carries target_selection { mode, column, reason, overridable, candidates[] }, where candidates lists every column with whether it was selected or ruled out and a plain-language reason, and task_selection { mode, type, reason, overridable }. You see the decision and the rationale. On an endpoint you can override the target afterwards from the endpoint page in the platform; a target change is a configuration revision and triggers a re-score. Smart-target picks categorical targets; to predict a numeric column, name it in target.column and set task.type to regression. On a text-only table, name the target in target.column; the platform prompts for it.

target.mode: "none" is the understanding-only configuration: the bundle ships sector, column profile, cross-table map, and imputation, and prediction is null.

Held-out (endpoints only)

{
"held_out": {
  "score": 0.961, "auc": 0.974, "baseline_majority": 0.710,
  "split": { "method": "in_context_holdout", "train_rows": 17920,
             "test_rows": 4480, "seed": 7 },
  "data_version": "2026-08-12 09:31", "data_snapshot": "dsv_91f2",
  "weakest_slice": "tenure<3mo: 0.84", "report_url": "/v2/reports/r_a7d0"
}
}

Present on the endpoint object. See Held-out reports.

Confidence, everywhere

Confidences are probabilities; calibration results per model are in the model card.

Large outputs and evolution

Retrieve the rest with cursor pagination on the endpoint or report resource, or direct a large run's bundle to a file or warehouse table with options.out. Adding fields to the bundle is non-breaking and announced; renaming or removing fields happens only at a new API version. Tolerate unknown fields. See Pagination.