## ---- brier_score ----------------------------------------------------

test_that("Brier score is zero for perfect predictions", {
  outcome    <- c(0, 1, 0, 1)
  confidence <- c(0, 1, 0, 1)
  expect_equal(brier_score(outcome, confidence), 0)
})

test_that("Brier score is one for maximally wrong predictions", {
  outcome    <- c(0, 1, 0, 1)
  confidence <- c(1, 0, 1, 0)
  expect_equal(brier_score(outcome, confidence), 1)
})

test_that("Brier score ignores incomplete cases", {
  outcome    <- c(0, 1, NA, 1)
  confidence <- c(0, 1, 0.5, 1)
  expect_equal(brier_score(outcome, confidence), 0)
})

## ---- ece and calibration_curve -------------------------------------

test_that("ECE is zero for perfectly calibrated predictions", {
  # within each bin, predicted probability equals empirical frequency
  set.seed(10)
  confidence <- rep(c(0.05, 0.95), each = 100)
  outcome <- c(rbinom(100, 1, 0.05), rbinom(100, 1, 0.95))
  val <- ece(outcome, confidence, n_bins = 10)
  expect_lt(val, 0.05)   # near zero up to sampling noise
})

test_that("ECE is large for badly miscalibrated predictions", {
  # model claims high confidence but is wrong half the time
  confidence <- rep(0.95, 100)
  outcome <- rep(c(0, 1), 50)  # only 50% actually positive
  val <- ece(outcome, confidence, n_bins = 10)
  expect_gt(val, 0.4)
})

test_that("calibration_curve omits empty bins rather than treating them as zero", {
  confidence <- c(0.1, 0.1, 0.9, 0.9)
  outcome <- c(0, 0, 1, 1)
  cc <- calibration_curve(outcome, confidence, n_bins = 10)
  expect_true(all(cc$n > 0))
  expect_lte(nrow(cc), 10)
})

## ---- input validation ----------------------------------------------

test_that("calibration rejects probabilities outside [0, 1]", {
  expect_error(ai_calibration(c(0, 1, 1), c(-0.1, 0.5, 1.1)),
               "\\[0, 1\\]")
})

test_that("calibration rejects non-binary outcomes", {
  expect_error(ai_calibration(c(0, 1, 2), c(0.1, 0.5, 0.9)),
               "binary")
})

test_that("calibration rejects mismatched lengths", {
  expect_error(ai_calibration(c(0, 1), c(0.5)), "same length")
})

test_that("calibration rejects n_bins < 2", {
  expect_error(ai_calibration(c(0, 1), c(0.2, 0.8), n_bins = 1), "n_bins")
})

test_that("ai_calibration returns ece, brier, and curve", {
  set.seed(11)
  outcome <- rbinom(100, 1, 0.4)
  confidence <- pmin(pmax(outcome * 0.5 + runif(100, 0, 0.5), 0), 1)
  out <- ai_calibration(outcome, confidence)
  expect_s3_class(out, "aiEvalR_calibration")
  expect_true(all(c("ece", "brier", "curve") %in% names(out)))
})

## ---- empirical_output_interval -------------------------------------

test_that("empirical_output_interval brackets the median", {
  set.seed(12)
  samples <- rnorm(1000, mean = 5, sd = 1)
  out <- empirical_output_interval(samples, conf = 0.90)
  expect_lt(out$lower, out$median)
  expect_lt(out$median, out$upper)
  expect_equal(out$median, median(samples), tolerance = 0.1)
})