ci: publish the benchmark table as a reproducible artifact (

docs/explanation/benchmark_results.txt

@@ -0,0 +1,11 @@
+Synthetic donor pool: 4000 rows, positive rate 0.378 (major donors), features=['total_gift_amount', 'years_active', 'event_attendance_count']
+Stratified 75/25 split, test n=1000, seeds=[42, 43, 44, 45, 46]
+Each cell is mean (min-max) across the seeds.
+
+model                                      precision                recall                    f1               roc_auc
+----------------------------------------------------------------------------------------------------------------------
+PropensityScorer (baseline)        0.000 (0.000-0.000)     0.000 (0.000-0.000)     0.000 (0.000-0.000)     0.500 (0.500-0.500)
+DonorPropensityModel               0.669 (0.625-0.707)     0.607 (0.588-0.632)     0.636 (0.623-0.645)     0.810 (0.802-0.815)
+MajorGiftClassifier                0.732 (0.693-0.752)     0.565 (0.544-0.593)     0.638 (0.614-0.663)     0.828 (0.819-0.832)
+LapsePredictor                     0.669 (0.625-0.707)     0.607 (0.588-0.632)     0.636 (0.623-0.645)     0.810 (0.802-0.815)
+PlannedGivingIntentScorer          0.730 (0.717-0.740)     0.605 (0.576-0.636)     0.662 (0.643-0.683)     0.840 (0.833-0.844)

docs/explanation/benchmarks.md

@@ -70,6 +70,10 @@ Five seeds rather than one on purpose. A single three-decimal score reads as a
 claim about the method when it is mostly a claim about the split; the spread
 below is the honest resolution of these numbers.
 
+CI runs `scripts/benchmark_models.py` on every push and diffs the output against
+the committed golden file `docs/explanation/benchmark_results.txt` (generated on
+Linux so BLAS rounding matches the runner).
+
 ## Results
 
 Synthetic pool: 4,000 rows, positive rate 0.378; test split 1,000 rows.

scripts/benchmark_models.py

@@ -36,7 +36,6 @@
     PropensityScorer,
 )
 
-RANDOM_STATE = 42
 # Five seeds, not one: a three-decimal score from a single split reads as a
 # claim about the method when it is mostly a claim about the split.
 SEEDS = (42, 43, 44, 45, 46)

tests/test_benchmark_models.py

@@ -0,0 +1,54 @@
+"""Lock the benchmark table to scripts/benchmark_models.py output."""
+
+import re
+import subprocess
+import sys
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+GOLDEN = REPO_ROOT / "docs" / "explanation" / "benchmark_results.txt"
+SCRIPT = REPO_ROOT / "scripts" / "benchmark_models.py"
+
+# The metrics come from fitted sklearn models, so they move with the
+# BLAS/sklearn build. CI showed means drifting up to ~0.006 and seed range
+# endpoints up to ~0.015 across Python versions on the same commit; anything
+# past these bands means the benchmark itself broke.
+MEAN_TOLERANCE = 0.02
+RANGE_TOLERANCE = 0.05
+DECIMAL = re.compile(r"\d+\.\d+")
+CELL = re.compile(r"(\d+\.\d+) \((\d+\.\d+)-(\d+\.\d+)\)")
+
+
+def test_benchmark_table_is_reproducible():
+    result = subprocess.run(
+        [sys.executable, str(SCRIPT)],
+        cwd=REPO_ROOT,
+        capture_output=True,
+        text=True,
+        check=True,
+    )
+    golden_lines = GOLDEN.read_text().splitlines()
+    actual_lines = result.stdout.splitlines()
+    assert len(actual_lines) == len(golden_lines), (
+        "benchmark output changed shape; regenerate "
+        "docs/explanation/benchmark_results.txt and commit verbatim"
+    )
+    for actual, golden in zip(actual_lines, golden_lines):
+        assert DECIMAL.sub("#", actual) == DECIMAL.sub("#", golden), (
+            f"non-metric part of the benchmark table drifted:\n"
+            f"expected: {golden}\nactual:   {actual}"
+        )
+        actual_cells = CELL.findall(actual)
+        golden_cells = CELL.findall(golden)
+        assert len(actual_cells) == len(golden_cells), (
+            f"metric cell count drifted:\nexpected: {golden}\nactual:   {actual}"
+        )
+        for actual_cell, golden_cell in zip(actual_cells, golden_cells):
+            for actual_value, golden_value, tolerance in zip(
+                actual_cell, golden_cell, (MEAN_TOLERANCE,) + (RANGE_TOLERANCE,) * 2
+            ):
+                assert abs(float(actual_value) - float(golden_value)) <= tolerance, (
+                    f"benchmark metric drifted beyond {tolerance}: "
+                    f"expected {golden_value}, got {actual_value}\n"
+                    f"line: {actual}"
+                )