Built-in ML

Expert systems hit a wall the moment a decision needs a learned judgement — “is this reading anomalous?”, “what’s this incident’s likely cause?”, “when will we run out?”. In Prolog you leave the language: export the data, run a Python/R pipeline, glue the answer back. Tln builds the common cases in as blocks, each returning an explanation alongside its value — so an ML result is as auditable as any rule.

The primitives

Eleven statistical/ML primitives ship in the runtime, each traceable:

Predict — decision tree

Train on retired machines, predict the outcome for in-service ones. Model and inference are one block; confidence gates the result:

predict "Failure risk" {
  for records where type == "machine" and status == "in_service"
  features [attr "operating_hours", attr "repair_count"]
  trained_on records where type == "machine" and status == "retired"
  label_attr "outcome"
  confidence >= 0.9
  label "predicted outcome: {class}"
}

Classify — k-NN

classify "Failure mode" {
  for records where type == "incident" and status == "open"
  features [attr "vibration", attr "temp"]
  trained_on records where type == "incident" and status == "resolved"
  label_attr "root_cause"
  confidence >= 0.8
  label "likely cause: {class}"
}

Forecast — time series

forecast "Parts stock-out" {
  for records where type == "stock_item" and status == "active"
  series attr "current_stock" over last 90 days
  label "{item.name}: stock-out in ~{days_until} days"
}

Models can also be declared once and shared: a model block carries fitted examples plus provenance (computed_from, valid_until), exported from a module and pulled in with using model "fleet.ml.failure_risk". Same determinism, same explainability — just learned from data instead of hand-written.