## ---- hallucination_rate --------------------------------------------

test_that("hallucination_rate equals the mean of adjudicated labels", {
  labels <- c(0, 1, 0, 1, 1)
  expect_equal(hallucination_rate(labels), 0.6)
})

test_that("hallucination_rate accepts logical verdicts", {
  verdicts <- c(TRUE, FALSE, TRUE, TRUE)
  expect_equal(hallucination_rate(verdicts), 0.75)
})

test_that("hallucination_rate returns per-group rates when 'by' is supplied", {
  labels <- c(1, 1, 0, 0)
  grp <- c("a", "a", "b", "b")
  out <- hallucination_rate(labels, by = grp)
  expect_equal(out[["a"]], 1)
  expect_equal(out[["b"]], 0)
})

test_that("hallucination_rate ignores NA labels", {
  labels <- c(1, 0, NA, 1)
  expect_equal(hallucination_rate(labels), 2/3)
})

## ---- lexical_overlap -----------------------------------------------

test_that("lexical_overlap is 1 for identical text", {
  expect_equal(lexical_overlap("the cat sat", "the cat sat"), 1)
})

test_that("lexical_overlap is 0 for fully disjoint text", {
  expect_equal(lexical_overlap("cat dog", "fish bird"), 0)
})

test_that("lexical_overlap is case-insensitive and punctuation-robust", {
  expect_equal(lexical_overlap("The Cat, sat!", "the cat sat"), 1)
})

test_that("lexical_overlap returns NA for empty input", {
  expect_true(is.na(lexical_overlap("", "something")))
  expect_true(is.na(lexical_overlap("something", "")))
})

test_that("lexical_overlap computes correct partial Jaccard", {
  # tokens {a,b,c} vs {b,c,d}: intersection 2, union 4 -> 0.5
  expect_equal(lexical_overlap("a b c", "b c d"), 0.5)
})

## ---- citation_accuracy ---------------------------------------------

test_that("citation_accuracy detects fabricated citations", {
  cited <- c("doi:1", "doi:2", "doi:FAKE")
  known <- c("doi:1", "doi:2", "doi:3")
  out <- citation_accuracy(cited, known)
  expect_equal(out$exists_rate, 2/3)
  expect_equal(out$fabricated, "doi:FAKE")
})

test_that("citation_accuracy reports full accuracy when all citations are real", {
  out <- citation_accuracy(c("a", "b"), c("a", "b", "c"))
  expect_equal(out$exists_rate, 1)
  expect_length(out$fabricated, 0)
})