diff --git a/tests/testthat/test-live-corpus.R b/tests/testthat/test-live-corpus.R
new file mode 100644
index 0000000..ca9bd86
--- /dev/null
+++ b/tests/testthat/test-live-corpus.R
@@ -0,0 +1,62 @@
+# Network-gated. Set USCOGDATA_LIVE_TEST=true to run.
+#
+# This file exists because the defect fixed for 0.3.0 -- no remote corpus was
+# readable at all, because DuckDB cannot expand a glob over generic HTTP --
+# survived precisely because every other test path used a LOCAL corpus (the
+# bundled fixture), and so did the API in production (a host mount). Nothing
+# ever exercised the package the way a new user does.
+skip_live <- function() {
+  testthat::skip_if_not(
+    identical(tolower(Sys.getenv("USCOGDATA_LIVE_TEST", "")), "true"),
+    "live-corpus test: set USCOGDATA_LIVE_TEST=true to run"
+  )
+}
+
+# The suite's setup.R pins USCOGDATA_URL to the bundled fixture, so reaching
+# the default requires clearing both the env var and the option.
+with_default_corpus <- function(code) {
+  withr::local_envvar(
+    USCOGDATA_URL = NA, USCOGDATA_FIXTURE_URL = NA,
+    .local_envir = parent.frame()
+  )
+  withr::local_options(uscogdata.url = NULL, .local_envir = parent.frame())
+  cog_close()
+  withr::defer(cog_close(), envir = parent.frame())
+  force(code)
+}
+
+test_that("the package reads the public corpus with no configuration at all", {
+  skip_live()
+  with_default_corpus({
+    g <- cog_gov_search(name = "Madison", state = "WI", type = 2)
+    expect_gt(nrow(g), 0)
+
+    s <- cog_spending(g$canonical_govid[1], years = 2022)
+    expect_gt(nrow(s), 0)
+    expect_true(all(c("amt_nominal", "year", "category") %in% names(s)))
+
+    # Amounts are full dollars, already x1000. A city's annual spending is
+    # millions, not thousands -- this catches a regression that dropped or
+    # doubled the conversion.
+    expect_gt(sum(s$amt_nominal, na.rm = TRUE), 1e6)
+
+    p <- attr(s, "provenance")
+    expect_true(isTRUE(p$transformations$units_conversion$applied))
+    expect_equal(p$transformations$units_conversion$multiplier, 1000)
+  })
+})
+
+test_that("a multi-decade query reads across many partitions", {
+  skip_live()
+  with_default_corpus({
+    g <- cog_gov_search(name = "Madison", state = "WI", type = 2)
+    # `years` is required on cog_spending() -- there is no full-history
+    # default at the reader level (the API's /profile route supplies one).
+    s <- cog_spending(g$canonical_govid[1], years = 2000:2022)
+    # Enumeration builds one read_parquet() path per requested partition. If
+    # the list were truncated, or silently collapsed to a single file, the
+    # returned span is what catches it.
+    expect_gt(diff(range(s$year)), 10)
+    expect_gt(length(unique(s$year)), 5)
+  })
+})