DATAGOATDocsQuick startConnectKeysBilling

Examples

Four worked examples on the free samples, each run live on 2026-09-24 against engine core d65816fb…. The samples are synthetic, so each example shows how Datagoat behaves, not how your data will. Every call here is free and deterministic: the same call returns the same numbers.

In each, "on unseen rows" means top_decile_lift: the top tenth of cases by chance, among rows the model never saw, had the outcome that many times as often as average (Quality).

Deal risk: refused, then answered

Question: will deal deal_0001@demo be won?

  • From the deal table alone (sample:deal_stages: each deal's stage and amount), the answer is refused (no_finding_cleared). Stage and amount hold no pattern that survives on unseen deals.
  • From the activity log read as of each stage (sample:deal_activity with the snapshots shape), it is answered: p 0.70, confidence high, 2.73× on unseen rows, measured on whole deals held out (held_out.whole_cases_by: "deal_id", validation_scheme: out_of_time).
  • The pattern: inbound_count_30d over 6, meeting_count_30d over 0, amount 12000 or less, inbound_count_7d over 1; this deal meets all four.

The calls: Shapes: snapshots. A longer write-up of the same sample: datagoat.io/gde.

What it does not claim: that meetings cause wins (reasons are associations), or that a real pipeline shows the same pattern.

Machine faults

Question: which of 40 machines will fault in the next three days (sample:sensor_stream, signals shape, daily snapshots)?

  • answered, confidence high, 3.36× on unseen rows (out_of_time).
  • The top machine, asset_030 (p 0.59), meets all four pattern conditions: pressure_psi_min_7 38.359 or less, pressure_psi_min_3 over 39.7188, pressure_psi_max_7 over 59.017, vibration_max_7 over 1.468.

What it does not claim: anything about machines other than these 40 synthetic ones, or a fault date. Asked with only the first three weeks of readings (as_of 2026-04-22), it answers not_yet: 840 snapshots are too few to hold out a reliable check.

Agent runs

Question: which of 800 agent runs will fail (sample:agent_traces, traces shape)?

  • answered, 2.5× on unseen rows (out_of_time), confidence low.
  • The riskiest run, run_0392 (p 0.73): the browser tool, which had failed on 21% of earlier runs, on a refactor task. The pattern names the same tool and task.
  • With band: true, the top four runs are act / L2, the fifth escalate / L1.

What it does not claim: much certainty. Its own confidence is low: say so when you show it.

Churn and the best offer

Questions: will acct_0001 churn, and which contract makes it least likely (sample:telco_churn)? Both in one call.

  • yesno: p 0.198, confidence high, 6.19× on unseen rows (holdout); with band: true, escalate / L1.
  • choice over contract: one_year (0.091), against two_year (0.198) and month-to-month (0.701).
  • Scored from the fitted model with no fit, a new customer with the same values but tech_support: yes drops from 0.63 to 0.12.

What it does not claim: that switching contract or adding tech support would change what a customer does. That is what Did acting work? measures, once you act and report outcomes.

Next

Take one of these to your own data: Ask your data.