Define a hard constraint on the models Atlas builds

Constraints carry domain knowledge into the build: the description is put in front of the agent as a hard requirement, and if check is supplied the constraint is verified - every final model is tested after the build, and violations are sent back to the agent to fix (see max_fix_rounds in atlas()). Plain strings passed to constraints are shorthand for constraint(<string>): enforced via instructions only.

Usage

constraint(description, check = NULL)

Arguments

description

What must hold, in plain language. The agent reads this.

check

Optional verification: function(model, data) returning TRUE if the constraint holds, or a string describing the violation (FALSE also counts as a violation). Errors count as violations too, so a check may simply stop() on bad models.

Value

An atlas_constraint.

See also

con_uses(), con_monotone() for ready-made checks.

Examples

constraint("never use the `id` column; it is a row identifier")
#> [prompt-only]     never use the `id` column; it is a row identifier

constraint(
  "predictions must never be negative",
  check = function(model, data) {
    p <- predict(model, newdata = data)
    if (all(p >= 0)) TRUE else "some predictions are negative"
  }
)
#> [machine-checked] predictions must never be negative