diff --git a/tests/testthat/test-ts-tbr-dirty-rescore.R b/tests/testthat/test-ts-tbr-dirty-rescore.R
new file mode 100644
index 000000000..cda2f874b
--- /dev/null
+++ b/tests/testthat/test-ts-tbr-dirty-rescore.R
@@ -0,0 +1,121 @@
+# Tier 2: skipped on CRAN; see tests/testing-strategy.md
+skip_on_cran()
+# T-304: enduring regression test for the T-300 dirty-set incremental
+# rescore wired into the tbr_search SPR accept path (src/ts_tbr.cpp,
+# ~lines 1138-1180).
+#
+# On an accepted SPR move the search does NOT call full_rescore; instead it
+# updates only the nodes on the union of paths nz->root and nx->root via
+# fitch_dirty_downpass / fitch_dirty_uppass (EW) or the NA-block variants
+# (fitch_na_dirty_*), then derives the score incrementally.  Four code paths
+# exist: EW, IW, NA, and NA-IW (is_spr && !has_na | is_spr && has_na, each
+# crossed with use_iw).
+#
+# The DEBUG_RESCORE / DEBUG_NA_RESCORE / DEBUG_NNI_RESCORE cross-checks that
+# originally guarded this were removed (commits 5b210fdd, 44a4ebeb, 2be8228d),
+# and an earlier incremental attempt regressed with a systematic delta = -3
+# and had to be reverted (b7303ee5).  This test is the permanent guard: it
+# drives MANY accepted SPR moves (small tips, weak signal, high maxHits) and
+# asserts that the score the search reports equals an independent full
+# recomputation.  If the dirty-set rescore ever drifts from the authoritative
+# score, result$score != ts_score(result_tree, ds) and these fail.
+
+ts_tbr <- function(tree, ds, maxHits = 20L, concavity = -1.0,
+                   min_steps = integer(0)) {
+  TreeSearch:::ts_tbr_search(tree$edge, ds$contrast, ds$tip_data,
+                             ds$weight, ds$levels,
+                             maxHits = maxHits, min_steps = min_steps,
+                             concavity = concavity)
+}
+
+result_tree <- function(result, ref_tree) {
+  rt <- ref_tree
+  rt$edge <- result$edge
+  rt
+}
+
+test_that("TBR dirty-set rescore matches full rescore (EW, many accepts)", {
+  # 12 tips, 6 random multistate characters -> weak signal, so the search
+  # accepts a long chain of SPR moves, each exercising the EW dirty-set path.
+  set.seed(4471)
+  mat <- matrix(sample(0:3, 12 * 6, replace = TRUE),
+                nrow = 12, dimnames = list(paste0("t", 1:12), NULL))
+  dataset <- MatrixToPhyDat(mat)
+  ds <- make_ts_data(dataset)
+
+  for (start in c(1, 17, 88, 256, 777)) {
+    tree <- as.phylo(start, 12)
+    set.seed(1000 + start)
+    result <- ts_tbr(tree, ds, maxHits = 50L)
+
+    rt <- result_tree(result, tree)
+    independent_score <- ts_score(rt, ds)
+    expect_equal(result$score, independent_score,
+                 info = paste("EW start =", start))
+    validate_result(result, 12L)
+  }
+})
+
+test_that("TBR dirty-set rescore matches full rescore (IW, many accepts)", {
+  set.seed(5529)
+  mat <- matrix(sample(0:2, 12 * 6, replace = TRUE),
+                nrow = 12, dimnames = list(paste0("t", 1:12), NULL))
+  dataset <- MatrixToPhyDat(mat)
+  ds <- make_ts_data(dataset)
+  minSteps <- as.integer(MinimumLength(dataset, compress = TRUE))
+
+  for (start in c(1, 17, 88, 256, 777)) {
+    tree <- as.phylo(start, 12)
+    set.seed(2000 + start)
+    result <- ts_tbr(tree, ds, maxHits = 50L, concavity = 10,
+                     min_steps = minSteps)
+
+    rt <- result_tree(result, tree)
+    independent_score <- ts_score(rt, ds, concavity = 10, min_steps = minSteps)
+    expect_equal(result$score, independent_score, tolerance = 1e-10,
+                 info = paste("IW start =", start))
+    validate_result(result, 12L)
+  }
+})
+
+test_that("TBR dirty-set rescore matches full rescore (NA dataset, many accepts)", {
+  skip_if_not_installed("TreeSearch")
+  data("inapplicable.phyData", package = "TreeSearch")
+  dataset <- inapplicable.phyData[["Vinther2008"]]
+  ds <- make_ts_data(dataset)
+  n_tip <- length(dataset)
+
+  for (start in c(1, 42, 100, 314)) {
+    tree <- as.phylo(start, n_tip)
+    set.seed(3000 + start)
+    result <- ts_tbr(tree, ds, maxHits = 20L)
+
+    rt <- result_tree(result, tree)
+    independent_score <- ts_score(rt, ds)
+    expect_equal(result$score, independent_score,
+                 info = paste("NA start =", start))
+    validate_result(result, n_tip)
+  }
+})
+
+test_that("TBR dirty-set rescore matches full rescore (NA-IW dataset, many accepts)", {
+  skip_if_not_installed("TreeSearch")
+  data("inapplicable.phyData", package = "TreeSearch")
+  dataset <- inapplicable.phyData[["Vinther2008"]]
+  ds <- make_ts_data(dataset)
+  n_tip <- length(dataset)
+  minSteps <- as.integer(MinimumLength(dataset, compress = TRUE))
+
+  for (start in c(1, 42, 100, 314)) {
+    tree <- as.phylo(start, n_tip)
+    set.seed(4000 + start)
+    result <- ts_tbr(tree, ds, maxHits = 20L, concavity = 10,
+                     min_steps = minSteps)
+
+    rt <- result_tree(result, tree)
+    independent_score <- ts_score(rt, ds, concavity = 10, min_steps = minSteps)
+    expect_equal(result$score, independent_score, tolerance = 1e-10,
+                 info = paste("NA-IW start =", start))
+    validate_result(result, n_tip)
+  }
+})