test_that("ai_group_disparity computes demographic parity correctly", {
  outcome <- c(1, 1, 0, 0, 1, 0)
  group <- c("A", "A", "A", "B", "B", "B")
  res <- ai_group_disparity(outcome, group)
  expect_equal(round(res$group_means[["A"]], 3), round(2/3, 3))
  expect_equal(round(res$group_means[["B"]], 3), round(1/3, 3))
  expect_equal(round(res$demographic_parity_diff, 3), round(1/3, 3))
})

test_that("ai_group_disparity rejects probabilities passed as predicted_class", {
  outcome <- c(1, 0, 1, 0)
  group <- c("A", "A", "B", "B")
  bad_pred <- c(0.2, 0.8, 0.4, 0.9)
  expect_error(ai_group_disparity(outcome, group, predicted_class = bad_pred),
               "must be binary")
})

test_that("ai_dif_mh requires exactly two groups", {
  expect_error(
    ai_dif_mh(c(1,0,1,0,1,0), c("A","B","C","A","B","C"), c(1,2,3,4,5,6)),
    "exactly 2 groups"
  )
})

test_that("ai_dif_mh runs and returns expected fields", {
  set.seed(1)
  n <- 200
  total <- rnorm(n)
  group <- sample(c("ref", "focal"), n, replace = TRUE)
  item <- rbinom(n, 1, plogis(total))
  res <- ai_dif_mh(item, group, total, n_strata = 4)
  expect_true(all(c("mh_or", "mh_delta", "chisq", "p_value") %in% names(res)))
  expect_true(res$mh_or > 0)
})

test_that("ai_fairness warns rather than errors when aiDIF is not installed", {
  skip_if(requireNamespace("aiDIF", quietly = TRUE),
          "aiDIF is installed; see 'dispatches to aiDIF' test instead")
  outcome <- c(1,0,1,0); group <- c("A","A","B","B")
  expect_warning(
    ai_fairness(outcome, group, human_mle = list(), ai_mle = list()),
    "aiDIF"
  )
})

test_that("ai_fairness dispatches to aiDIF::fit_aidif with valid data when installed", {
  skip_if_not_installed("aiDIF")
  outcome <- c(1,0,1,0); group <- c("A","A","B","B")
  dat <- aiDIF::simulate_aidif_data(n_items = 6, seed = 1)

  res <- ai_fairness(outcome, group,
                      human_mle = dat$human, ai_mle = dat$ai)

  expect_true(!is.null(res$item_level_dasb))
  expect_s3_class(res$item_level_dasb, "aidif")
})

test_that("ai_fairness's malformed-mle error from aiDIF is informative, not swallowed", {
  skip_if_not_installed("aiDIF")
  outcome <- c(1,0,1,0); group <- c("A","A","B","B")
  # empty lists are invalid mle input -- aiDIF's own validation should
  # surface a clear error, not be silently caught by aiEvalR
  expect_error(
    ai_fairness(outcome, group, human_mle = list(), ai_mle = list()),
    "est.*var.cov|var.cov.*est"
  )
})

## ---- additional coverage: chi-square, equalized odds, dashboard ----

test_that("ai_dif_mh returns chi-square and p-value with valid structure", {
  set.seed(40)
  n <- 300
  total <- rnorm(n)
  group <- sample(c("ref", "focal"), n, replace = TRUE)
  item <- rbinom(n, 1, plogis(total))
  res <- ai_dif_mh(item, group, total, n_strata = 5)
  expect_true(all(c("mh_or", "mh_delta", "chisq", "p_value", "n_strata_used") %in% names(res)))
  expect_gte(res$chisq, 0)
  expect_gte(res$p_value, 0)
  expect_lte(res$p_value, 1)
})

test_that("ai_dif_mh detects strong planted DIF with a small p-value", {
  set.seed(41)
  n <- 600
  total <- rnorm(n)
  group <- sample(c("ref", "focal"), n, replace = TRUE)
  shift <- ifelse(group == "focal", 1.5, 0)  # strong DIF
  item <- rbinom(n, 1, plogis(total + shift))
  res <- ai_dif_mh(item, group, total, n_strata = 5)
  expect_lt(res$p_value, 0.05)
})

test_that("ai_group_disparity computes equalized-odds gaps for binary decisions", {
  set.seed(42)
  outcome <- rbinom(200, 1, 0.5)
  group <- sample(c("A", "B"), 200, replace = TRUE)
  predicted <- rbinom(200, 1, 0.5)
  out <- ai_group_disparity(outcome, group, predicted_class = predicted)
  expect_true("equalized_odds" %in% names(out))
  expect_true(all(c("tpr_gap", "fpr_gap") %in% names(out$equalized_odds)))
})

test_that("ai_fairness computes calibration-by-group when probabilities given", {
  set.seed(43)
  outcome <- rbinom(200, 1, 0.4)
  group <- sample(c("A", "B"), 200, replace = TRUE)
  prob <- pmin(pmax(outcome * 0.5 + runif(200, 0, 0.5), 0), 1)
  out <- ai_fairness(outcome, group, predicted_probability = prob)
  expect_true("calibration_by_group" %in% names(out))
})

test_that("fairness_dashboard prints a summary without error", {
  outcome <- c(1, 1, 0, 0, 1, 0)
  group <- c("A", "A", "A", "B", "B", "B")
  fobj <- ai_fairness(outcome, group)
  expect_output(fairness_dashboard(fobj), "Fairness Summary")
})

test_that("fairness_dashboard rejects non-fairness objects", {
  expect_error(fairness_dashboard(list(a = 1)), "aiEvalR_fairness")
})