Skip to contents

Compute guardrail evaluation metrics

Source: R/guardrail-eval.R
guardrail_metrics.Rd

Computes precision, recall, F1, accuracy, and confusion counts from a guardrail evaluation result.

Usage

guardrail_metrics(eval_result)

Arguments

eval_result

A guardrail_eval_result object.

Value

A named list with tp, tn, fp, fn, precision, recall, f1, accuracy.

Details

Convention: blocking is the "positive" class.

  • True positive: expected=FALSE (should block) and pass=FALSE (was blocked)

  • True negative: expected=TRUE (should pass) and pass=TRUE (was passed)

  • False positive: expected=TRUE (should pass) but pass=FALSE (was blocked)

  • False negative: expected=FALSE (should block) but pass=TRUE (was passed)

Examples

data <- data.frame(
  input = c("hello", "DROP TABLE users"),
  expected = c(TRUE, FALSE)
)
my_guard <- function(text) !grepl("DROP TABLE", text, fixed = TRUE)
result <- guardrail_eval(my_guard, data)
m <- guardrail_metrics(result)
m$precision
#> [1] 1
m$recall
#> [1] 1