diff --git a/lib/codeqa/engine/analyzer.ex b/lib/codeqa/engine/analyzer.ex
index d1f0acf..99fc5d2 100644
--- a/lib/codeqa/engine/analyzer.ex
+++ b/lib/codeqa/engine/analyzer.ex
@@ -230,14 +230,15 @@ defmodule CodeQA.Engine.Analyzer do
           "mean_#{key}" => stats.mean,
           "std_#{key}" => stats.std,
           "min_#{key}" => stats.min,
-          "max_#{key}" => stats.max
+          "max_#{key}" => stats.max,
+          "p90_#{key}" => stats.p90
         })
 
       Map.put(acc, metric, updated)
     end)
   end
 
-  defp compute_stats([]), do: %{max: 0.0, mean: 0.0, min: 0.0, std: 0.0}
+  defp compute_stats([]), do: %{max: 0.0, mean: 0.0, min: 0.0, p90: 0.0, std: 0.0}
 
   defp compute_stats(values) do
     n = length(values)
@@ -250,7 +251,15 @@ defmodule CodeQA.Engine.Analyzer do
       max: Float.round(Enum.max(values) * 1.0, 4),
       mean: Float.round(mean * 1.0, 4),
       min: Float.round(Enum.min(values) * 1.0, 4),
+      p90: Float.round(percentile(values, 0.9) * 1.0, 4),
       std: Float.round(std * 1.0, 4)
     }
   end
+
+  # Nearest-rank percentile: the value the worst (1 - q) share of files stays under.
+  defp percentile(values, q) do
+    sorted = Enum.sort(values)
+    index = (length(sorted) * q) |> Float.floor() |> trunc() |> min(length(sorted) - 1)
+    Enum.at(sorted, index)
+  end
 end
diff --git a/lib/codeqa/health_report/categories.ex b/lib/codeqa/health_report/categories.ex
index 7146cf5..fedf87a 100644
--- a/lib/codeqa/health_report/categories.ex
+++ b/lib/codeqa/health_report/categories.ex
@@ -68,6 +68,10 @@ defmodule CodeQA.HealthReport.Categories do
         ],
         name: "Readability"
       },
+      # Halstead runs per FILE here, not per function, so these are anchored on measured
+      # file sizes rather than the function-level figures from the literature: a = a
+      # 100-LOC file (the common case), d = a 500-LOC file (the largest still tolerated).
+      # Past d the score falls off a cliff, which is where files over 500 LOC land.
       %{
         key: :complexity,
         metrics: [
@@ -77,7 +81,7 @@ defmodule CodeQA.HealthReport.Categories do
             good: :low,
             name: "difficulty",
             source: "halstead",
-            thresholds: %{a: 10, b: 20, c: 35, d: 50},
+            thresholds: %{a: 35, b: 50, c: 60, d: 70},
             weight: 0.35
           },
           %{
@@ -86,7 +90,7 @@ defmodule CodeQA.HealthReport.Categories do
             good: :low,
             name: "effort",
             source: "halstead",
-            thresholds: %{a: 5000, b: 20_000, c: 50_000, d: 100_000},
+            thresholds: %{a: 130_000, b: 440_000, c: 920_000, d: 1_850_000},
             weight: 0.3
           },
           %{
@@ -95,7 +99,7 @@ defmodule CodeQA.HealthReport.Categories do
             good: :low,
             name: "volume",
             source: "halstead",
-            thresholds: %{a: 300, b: 1000, c: 3000, d: 8000},
+            thresholds: %{a: 4_000, b: 8_700, c: 16_600, d: 25_000},
             weight: 0.2
           },
           %{
@@ -103,12 +107,48 @@ defmodule CodeQA.HealthReport.Categories do
             good: :low,
             name: "estimated_bugs",
             source: "halstead",
-            thresholds: %{a: 0.1, b: 0.5, c: 1.0, d: 3.0},
+            thresholds: %{a: 1.35, b: 2.9, c: 5.5, d: 8.4},
             weight: 0.15
           }
         ],
         name: "Complexity"
       },
+      # The mean says how big a typical file is; this says how bad the tail is. p90 is
+      # the value the worst 10% of files stay under, so a repo of uniformly sized files
+      # scores the same here as on Complexity, while one with a handful of monsters does
+      # not. Anchored on the same LOC bands: a = 200 LOC, d = 800 LOC.
+      %{
+        key: :complexity_outliers,
+        metrics: [
+          %{
+            fix_hint:
+              "The largest 10% of files carry far more logic than the rest — split the worst offenders",
+            good: :low,
+            name: "p90_volume",
+            source: "halstead",
+            thresholds: %{a: 8_700, b: 16_600, c: 25_000, d: 38_000},
+            weight: 0.4
+          },
+          %{
+            fix_hint:
+              "The largest files repeat the same operands heavily — extract named intermediates there",
+            good: :low,
+            name: "p90_difficulty",
+            source: "halstead",
+            thresholds: %{a: 50, b: 60, c: 70, d: 85},
+            weight: 0.35
+          },
+          %{
+            fix_hint: "The largest files dominate total implementation effort — decompose them",
+            good: :low,
+            name: "p90_effort",
+            source: "halstead",
+            thresholds: %{a: 440_000, b: 920_000, c: 1_850_000, d: 3_200_000},
+            weight: 0.25
+          }
+        ],
+        name: "Complexity Outliers"
+      },
       %{
         key: :structure,
         metrics: [
diff --git a/lib/codeqa/health_report/grader.ex b/lib/codeqa/health_report/grader.ex
index 4216d60..693c1be 100644
--- a/lib/codeqa/health_report/grader.ex
+++ b/lib/codeqa/health_report/grader.ex
@@ -174,13 +174,19 @@ defmodule CodeQA.HealthReport.Grader do
         aggregate,
         scale \\ Categories.default_grade_scale()
       ) do
-    # Convert aggregate format (mean_X keys) to file-metric-like format
+    # Convert aggregate format to file-metric-like format. mean_X is exposed as X so
+    # categories read it as a plain metric; p90_X keeps its prefix so a category can
+    # ask for the outlier tail instead of the average.
     file_like =
       Map.new(aggregate, fn {source, stats} ->
         values =
           stats
-          |> Enum.filter(fn {k, _v} -> String.starts_with?(k, "mean_") end)
-          |> Map.new(fn {"mean_" <> key, v} -> {key, v} end)
+          |> Enum.flat_map(fn
+            {"mean_" <> key, v} -> [{key, v}]
+            {"p90_" <> _ = key, v} -> [{key, v}]
+            _other -> []
+          end)
+          |> Map.new()
 
         {source, values}
       end)
diff --git a/test/codeqa/engine/analyzer_test.exs b/test/codeqa/engine/analyzer_test.exs
index ce655f1..daacfb1 100644
--- a/test/codeqa/engine/analyzer_test.exs
+++ b/test/codeqa/engine/analyzer_test.exs
@@ -33,7 +33,8 @@ defmodule CodeQA.Engine.AnalyzerTest do
         keys
         |> Enum.each(fn {key, val} ->
           assert String.starts_with?(key, "mean_") or String.starts_with?(key, "std_") or
-                   String.starts_with?(key, "min_") or String.starts_with?(key, "max_")
+                   String.starts_with?(key, "min_") or String.starts_with?(key, "max_") or
+                   String.starts_with?(key, "p90_")
 
           assert is_float(val) or is_integer(val)
         end)
diff --git a/test/codeqa/health_report/categories_test.exs b/test/codeqa/health_report/categories_test.exs
index 6b4e282..504b0ee 100644
--- a/test/codeqa/health_report/categories_test.exs
+++ b/test/codeqa/health_report/categories_test.exs
@@ -62,14 +62,14 @@ defmodule CodeQA.HealthReport.CategoriesTest do
       end)
     end
 
-    test "has exactly 24 metrics across 6 categories" do
+    test "has exactly 27 metrics across 7 categories" do
       categories = Categories.defaults()
 
-      assert length(categories) == 6
+      assert length(categories) == 7
 
       metrics = categories |> Enum.flat_map(& &1.metrics)
 
-      assert length(metrics) == 24
+      assert length(metrics) == 27
     end
   end
 end