diff --git a/CHANGELOG.md b/CHANGELOG.md
index d848b96..7895037 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,6 +8,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## [Unreleased]
 
 ### Added
+- `src/fairness.py` with `fairness_metrics` and `decision_disagreement` helpers
+  for demographic-parity / equal-opportunity gaps and receiver disagreement.
 - Apache 2.0 `LICENSE` + `NOTICE` with third-party data attribution (German
   Credit / OpenML `credit-g`, Chiappa 2019 DAG).
 - `data/README.md` documenting dataset provenance and the variable mapping.
@@ -19,8 +21,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - `pyproject.toml` packaging the engine with pinned dependencies, a `viz` extra
   for the plotting scripts, and Black/Ruff/mypy/pytest/coverage configuration.
 - `CITATION.cff` for software + the accompanying (forthcoming) paper.
-- `tests/` pytest suite covering `distances`, `linear_anm`, `perception`, and
-  `data_prep` (with the OpenML fetch mocked — no network access required).
+- `tests/` pytest suite covering `distances`, `linear_anm`, `perception`,
+  `data_prep`, and `fairness` (with the OpenML fetch mocked — no network access
+  required).
 - SPDX headers (`Copyright (c) 2026 José M. Álvarez` / `Apache-2.0`) on all
   Python source files.
 - GitHub Actions workflows (third-party actions pinned to SHA digests):
diff --git a/README.md b/README.md
index b226d4a..a8a4ad0 100644
--- a/README.md
+++ b/README.md
@@ -111,6 +111,7 @@ src/
   linear_anm.py             # Linear ANM: DAGs, fit, do-operator, counterfactuals
   distances.py              # W2, KL (KDE), Total Variation between 1-D samples
   perception.py             # competing-SCM engine + bootstrap CIs
+  fairness.py               # DP/EO gaps and decision disagreement metrics
   run_*.py                  # experiment entrypoints (require the `viz` extra)
   plot_structural_combined.py
 tests/                      # pytest suite (OpenML fetch mocked — no network)
diff --git a/src/fairness.py b/src/fairness.py
new file mode 100644
index 0000000..5db853e
--- /dev/null
+++ b/src/fairness.py
@@ -0,0 +1,99 @@
+# Copyright (c) 2026 José M. Álvarez
+# SPDX-License-Identifier: Apache-2.0
+
+"""
+Fairness metrics for credit decision experiments.
+
+Provides reusable functions for demographic-parity gaps, equal-opportunity
+gaps, and decision disagreement between competing receivers.
+"""
+
+import numpy as np
+
+
+def fairness_metrics(y_true, y_hat, a, tau=0.5):
+    """
+    Compute fairness metrics for a single receiver.
+
+    Parameters
+    ----------
+    y_true : array-like
+        Ground-truth outcome labels (1 = positive class).
+    y_hat : array-like
+        Predicted risk scores.
+    a : array-like
+        Protected attribute (0 = female, 1 = male).
+    tau : float
+        Decision threshold; accept when y_hat >= tau.
+
+    Returns
+    -------
+    dict
+        Acceptance rates, DP gap, TPR/FPR by group, and EO gaps.
+    """
+    d = (y_hat >= tau).astype(int)
+    female = a == 0
+    male = a == 1
+
+    accept_f = np.mean(d[female])
+    accept_m = np.mean(d[male])
+    accept_all = np.mean(d)
+    dp_gap = accept_f - accept_m  # signed: positive = females accepted more
+
+    # TPR = P(D=1 | Y=1, A=a)
+    tpr_f = np.mean(d[female & (y_true == 1)])
+    tpr_m = np.mean(d[male & (y_true == 1)])
+    # FPR = P(D=1 | Y=0, A=a)
+    fpr_f = np.mean(d[female & (y_true == 0)])
+    fpr_m = np.mean(d[male & (y_true == 0)])
+
+    return {
+        "accept_all": accept_all,
+        "accept_female": accept_f,
+        "accept_male": accept_m,
+        "dp_gap": dp_gap,
+        "tpr_female": tpr_f,
+        "tpr_male": tpr_m,
+        "tpr_gap": tpr_f - tpr_m,
+        "fpr_female": fpr_f,
+        "fpr_male": fpr_m,
+        "fpr_gap": fpr_f - fpr_m,
+    }
+
+
+def decision_disagreement(y_hat_a, y_hat_b, a, tau):
+    """
+    Compute decision disagreement between two receivers.
+
+    Parameters
+    ----------
+    y_hat_a, y_hat_b : array-like
+        Predicted risk scores from receiver A and B.
+    a : array-like
+        Protected attribute (0 = female, 1 = male).
+    tau : float
+        Decision threshold; accept when y_hat >= tau.
+
+    Returns
+    -------
+    dict
+        Overall and group-stratified disagreement rates, counts, and
+        directional flip counts.
+    """
+    d_a = (y_hat_a >= tau).astype(int)
+    d_b = (y_hat_b >= tau).astype(int)
+    disagree = d_a != d_b
+
+    r1_grant_r2_deny = (d_a == 1) & (d_b == 0)
+    r1_deny_r2_grant = (d_a == 0) & (d_b == 1)
+
+    return {
+        "disagree_rate": np.mean(disagree),
+        "disagree_female": np.mean(disagree[a == 0]),
+        "disagree_male": np.mean(disagree[a == 1]),
+        "n_disagree": int(disagree.sum()),
+        "n_disagree_female": int(disagree[a == 0].sum()),
+        "n_disagree_male": int(disagree[a == 1].sum()),
+        "r1_grant_r2_deny": int(r1_grant_r2_deny.sum()),
+        "r1_deny_r2_grant": int(r1_deny_r2_grant.sum()),
+    }
diff --git a/src/run_fair_decisions.py b/src/run_fair_decisions.py
index 2bccec9..0a9a22d 100644
--- a/src/run_fair_decisions.py
+++ b/src/run_fair_decisions.py
@@ -18,50 +18,15 @@
 import os
 
 import matplotlib.pyplot as plt
-import numpy as np
 import pandas as pd
 from sklearn.metrics import auc, average_precision_score, precision_recall_curve, roc_curve
 
 from src.data_prep import load_data
+from src.fairness import decision_disagreement, fairness_metrics
 from src.linear_anm import CHIAPPA_FULL, CHIAPPA_NO_AY
 from src.perception import fit_scms
 
 
-def fairness_metrics(y_true, y_hat, a, tau=0.5):
-    """Compute fairness metrics for a single receiver.
-
-    Returns dict with acceptance rates, DP gap, TPR/FPR by group, EO gaps.
-    """
-    d = (y_hat >= tau).astype(int)
-    female = a == 0
-    male = a == 1
-
-    accept_f = np.mean(d[female])
-    accept_m = np.mean(d[male])
-    accept_all = np.mean(d)
-    dp_gap = accept_f - accept_m  # signed: positive = females accepted more
-
-    # TPR = P(D=1 | Y=1, A=a)
-    tpr_f = np.mean(d[female & (y_true == 1)])
-    tpr_m = np.mean(d[male & (y_true == 1)])
-    # FPR = P(D=1 | Y=0, A=a)
-    fpr_f = np.mean(d[female & (y_true == 0)])
-    fpr_m = np.mean(d[male & (y_true == 0)])
-
-    return {
-        "accept_all": accept_all,
-        "accept_female": accept_f,
-        "accept_male": accept_m,
-        "dp_gap": dp_gap,
-        "tpr_female": tpr_f,
-        "tpr_male": tpr_m,
-        "tpr_gap": tpr_f - tpr_m,
-        "fpr_female": fpr_f,
-        "fpr_male": fpr_m,
-        "fpr_gap": fpr_f - fpr_m,
-    }
-
-
 def main():
     project_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
     train, test = load_data()
@@ -137,24 +102,18 @@ def main():
         )
 
     # Decision disagreement
-    d_r1 = (y_hat_r1 >= tau).astype(int)
-    d_r2 = (y_hat_r2 >= tau).astype(int)
-    disagree = d_r1 != d_r2
-    disagree_rate = np.mean(disagree)
-    disagree_f = np.mean(disagree[a == 0])
-    disagree_m = np.mean(disagree[a == 1])
-    n_disagree = int(disagree.sum())
+    dd = decision_disagreement(y_hat_r1, y_hat_r2, a, tau)
+    disagree_rate = dd["disagree_rate"]
+    disagree_f = dd["disagree_female"]
+    disagree_m = dd["disagree_male"]
+    n_disagree = dd["n_disagree"]
 
     print("\nDecision disagreement (R1 vs R2):")
     print(f"  Overall:  {disagree_rate:.3f} ({n_disagree} / {len(test)} applicants)")
-    print(f"  Female:   {disagree_f:.3f} ({int(disagree[a==0].sum())} / {(a==0).sum()})")
-    print(f"  Male:     {disagree_m:.3f} ({int(disagree[a==1].sum())} / {(a==1).sum()})")
-
-    # Who flips? (R1 grants but R2 denies, or vice versa)
-    r1_grant_r2_deny = (d_r1 == 1) & (d_r2 == 0)
-    r1_deny_r2_grant = (d_r1 == 0) & (d_r2 == 1)
-    print(f"  R1 grants, R2 denies: {r1_grant_r2_deny.sum()}")
-    print(f"  R1 denies, R2 grants: {r1_deny_r2_grant.sum()}")
+    print(f"  Female:   {disagree_f:.3f} ({dd['n_disagree_female']} / {(a==0).sum()})")
+    print(f"  Male:     {disagree_m:.3f} ({dd['n_disagree_male']} / {(a==1).sum()})")
+    print(f"  R1 grants, R2 denies: {dd['r1_grant_r2_deny']}")
+    print(f"  R1 denies, R2 grants: {dd['r1_deny_r2_grant']}")
 
     # PLOT: 1x2 figure (ROC + PR)
     plt.rcParams.update({"font.size": 12, "axes.titlesize": 13, "axes.labelsize": 12})
diff --git a/tests/test_fairness.py b/tests/test_fairness.py
new file mode 100644
index 0000000..dc08925
--- /dev/null
+++ b/tests/test_fairness.py
@@ -0,0 +1,85 @@
+# Copyright (c) 2026 José M. Álvarez
+# SPDX-License-Identifier: Apache-2.0
+
+"""Tests for src.fairness."""
+
+import numpy as np
+import pytest
+
+from src.fairness import decision_disagreement, fairness_metrics
+
+
+def test_fairness_metrics_perfect_separation_by_group():
+    """Known scores produce exact DP, TPR, and FPR gaps."""
+    y_true = np.array([1, 1, 0, 0, 1, 0])
+    a = np.array([0, 0, 0, 1, 1, 1])
+    # Female: accept 2/3; Male: accept 1/3 -> dp_gap = 2/3 - 1/3 = 1/3
+    y_hat = np.array([0.9, 0.8, 0.2, 0.8, 0.2, 0.1])
+
+    m = fairness_metrics(y_true, y_hat, a, tau=0.5)
+
+    assert m["accept_female"] == pytest.approx(2 / 3)
+    assert m["accept_male"] == pytest.approx(1 / 3)
+    assert m["dp_gap"] == pytest.approx(1 / 3)
+    assert m["tpr_female"] == pytest.approx(1.0)
+    assert m["tpr_male"] == pytest.approx(0.0)
+    assert m["tpr_gap"] == pytest.approx(1.0)
+    assert m["fpr_female"] == pytest.approx(0.0)
+    assert m["fpr_male"] == pytest.approx(0.5)
+    assert m["fpr_gap"] == pytest.approx(-0.5)
+
+
+def test_fairness_metrics_tau_threshold():
+    """Raising tau reduces acceptance rates."""
+    y_true = np.array([1, 0, 1, 0])
+    a = np.array([0, 0, 1, 1])
+    y_hat = np.array([0.6, 0.4, 0.6, 0.4])
+
+    low_tau = fairness_metrics(y_true, y_hat, a, tau=0.3)
+    high_tau = fairness_metrics(y_true, y_hat, a, tau=0.7)
+
+    assert low_tau["accept_all"] == pytest.approx(1.0)
+    assert high_tau["accept_all"] == pytest.approx(0.0)
+
+
+def test_fairness_metrics_empty_group_safe():
+    """np.mean on empty slice returns nan when a group has no Y=1 rows."""
+    y_true = np.array([0, 0, 1, 1])
+    a = np.array([0, 0, 1, 1])
+    y_hat = np.array([0.9, 0.8, 0.7, 0.6])
+
+    m = fairness_metrics(y_true, y_hat, a, tau=0.5)
+
+    assert np.isnan(m["tpr_female"])
+    assert m["tpr_male"] == pytest.approx(1.0)
+
+
+def test_decision_disagreement_symmetric():
+    """Swapping receivers preserves overall disagreement rate."""
+    y_hat_a = np.array([0.9, 0.2, 0.8, 0.1])
+    y_hat_b = np.array([0.2, 0.9, 0.1, 0.8])
+    a = np.array([0, 0, 1, 1])
+    tau = 0.5
+
+    dd_ab = decision_disagreement(y_hat_a, y_hat_b, a, tau)
+    dd_ba = decision_disagreement(y_hat_b, y_hat_a, a, tau)
+
+    assert dd_ab["disagree_rate"] == pytest.approx(dd_ba["disagree_rate"])
+    assert dd_ab["n_disagree"] == dd_ba["n_disagree"]
+
+
+def test_decision_disagreement_directional_counts():
+    """Constructed scores yield exact flip counts."""
+    y_hat_a = np.array([0.9, 0.1, 0.8, 0.2])
+    y_hat_b = np.array([0.1, 0.9, 0.2, 0.8])
+    a = np.array([0, 0, 1, 1])
+    tau = 0.5
+
+    dd = decision_disagreement(y_hat_a, y_hat_b, a, tau)
+
+    assert dd["disagree_rate"] == pytest.approx(1.0)
+    assert dd["n_disagree"] == 4
+    assert dd["r1_grant_r2_deny"] == 2
+    assert dd["r1_deny_r2_grant"] == 2
+    assert dd["n_disagree_female"] == 2
+    assert dd["n_disagree_male"] == 2