diff --git a/.github/workflows/check.yml b/.github/workflows/check.yml
index 6bedf380..c463033b 100644
--- a/.github/workflows/check.yml
+++ b/.github/workflows/check.yml
@@ -51,9 +51,9 @@ jobs:
         run: >-
           uv run --frozen --no-build pytest --no-header -q
           test_bench_stats.py test_bench_measure.py test_bench_runner.py
-          test_bench_arms.py test_bench_config.py test_sdk_commands.py
-          test_tdfs_units.py test_encryption_units.py test_sizes_units.py
-          test_zip64_units.py
+          test_bench_arms.py test_bench_config.py test_bench_report.py
+          test_sdk_commands.py test_tdfs_units.py test_encryption_units.py
+          test_sizes_units.py test_zip64_units.py
         working-directory: xtest
       - name: Lint and test otdf-local
         run: |
diff --git a/xtest/perf/report.py b/xtest/perf/report.py
index f5e4e8d6..71179440 100644
--- a/xtest/perf/report.py
+++ b/xtest/perf/report.py
@@ -226,7 +226,8 @@ def markdown(
         "",
         f"<sub>seed {config.seed}; warm-up {config.warmup} rounds; "
         f"{config.min_rounds}-{config.max_rounds} measured rounds per cell; "
-        "stopping on attained CI width, never on significance.</sub>",
+        "stopping on attained CI width, never on significance; "
+        "p is one-sided, in the direction of the observed ratio.</sub>",
     ]
     return "\n".join(lines) + "\n"
 
@@ -240,7 +241,14 @@ def _ratio_cell(c: stats.PairedComparison) -> str:
 
 
 def _p_cell(c: stats.PairedComparison) -> str:
-    p = c.p_adjusted if c.p_adjusted is not None else c.p_value
+    # Show the one-sided tail matching the observed effect. Below one that is
+    # the explicit faster-tail test: a BH-adjusted slower-tail p-value cannot
+    # be read backwards as evidence of improvement, and printing it beside an
+    # IMPROVED verdict shows ~1.000 next to a significant result.
+    if c.ratio < 1:
+        p = c.p_adjusted_faster if c.p_adjusted_faster is not None else c.p_value_faster
+    else:
+        p = c.p_adjusted if c.p_adjusted is not None else c.p_value
     if p is None or not math.isfinite(p):
         return "-"
     return f"{p:.3f}" if p >= 0.001 else "<0.001"
diff --git a/xtest/test_bench_report.py b/xtest/test_bench_report.py
new file mode 100644
index 00000000..bba99f7e
--- /dev/null
+++ b/xtest/test_bench_report.py
@@ -0,0 +1,84 @@
+"""Unit tests for the benchmark report's rendering helpers.
+
+Offline: no SDKs, no services. Builds `PairedComparison` values directly.
+"""
+
+import math
+
+import pytest
+
+from perf import stats
+from perf.report import _p_cell
+from perf.stats import Verdict
+
+
+def _comparison(
+    ratio: float,
+    *,
+    p_value: float,
+    p_value_faster: float,
+    p_adjusted: float | None = None,
+    p_adjusted_faster: float | None = None,
+    verdict: Verdict = Verdict.INCONCLUSIVE,
+) -> stats.PairedComparison:
+    return stats.PairedComparison(
+        n_rounds=20,
+        baseline_median=1.0,
+        candidate_median=ratio,
+        ratio=ratio,
+        ci_low=ratio * 0.9,
+        ci_high=ratio * 1.1,
+        p_value=p_value,
+        p_value_faster=p_value_faster,
+        p_adjusted=p_adjusted,
+        p_adjusted_faster=p_adjusted_faster,
+        verdict=verdict,
+    )
+
+
+class TestPCell:
+    def test_an_improvement_shows_the_faster_tail(self):
+        # The slower tail of a 2x speedup is ~1; printing it beside IMPROVED
+        # contradicts the verdict on the same row.
+        c = _comparison(
+            0.5,
+            p_value=1.0,
+            p_value_faster=0.0004,
+            p_adjusted=1.0,
+            p_adjusted_faster=0.0011,
+            verdict=Verdict.IMPROVED,
+        )
+        assert _p_cell(c) == "0.001"
+
+    def test_a_slowdown_shows_the_slower_tail(self):
+        c = _comparison(
+            1.5,
+            p_value=0.0004,
+            p_value_faster=1.0,
+            p_adjusted=0.002,
+            p_adjusted_faster=1.0,
+            verdict=Verdict.REGRESSION,
+        )
+        assert _p_cell(c) == "0.002"
+
+    def test_no_change_shows_the_slower_tail(self):
+        c = _comparison(1.0, p_value=0.5, p_value_faster=0.5, p_adjusted=0.6)
+        assert _p_cell(c) == "0.600"
+
+    @pytest.mark.parametrize(
+        ("ratio", "expected"),
+        [(0.8, "0.030"), (1.2, "0.020")],
+        ids=["faster", "slower"],
+    )
+    def test_unadjusted_tail_is_used_when_not_corrected(self, ratio, expected):
+        # Controls are reported unadjusted: they stay out of the BH families.
+        c = _comparison(ratio, p_value=0.02, p_value_faster=0.03)
+        assert _p_cell(c) == expected
+
+    def test_tiny_p_is_floored(self):
+        c = _comparison(0.5, p_value=1.0, p_value_faster=1e-6, p_adjusted_faster=1e-5)
+        assert _p_cell(c) == "<0.001"
+
+    def test_an_empty_comparison_shows_a_dash(self):
+        c = _comparison(math.nan, p_value=math.nan, p_value_faster=math.nan)
+        assert _p_cell(c) == "-"