diff --git a/packages/core/src/gravity-roll.js b/packages/core/src/gravity-roll.js
new file mode 100644
index 0000000..0763f73
--- /dev/null
+++ b/packages/core/src/gravity-roll.js
@@ -0,0 +1,45 @@
+// ============================================================
+// GRAVITY ROLL — drift-free bank/roll angle from the gravity vector (#314)
+// ============================================================
+//
+// Roll (bank / lean) extracted purely from the accelerometer-anchored gravity
+// vector, NOT from the integrated orientation quaternion. Decomposing the full
+// orientation to Euler-Z to get roll mixes in YAW — and yaw is the one axis a
+// 6-axis IMU can never anchor (no magnetometer), so it drifts without bound and
+// the roll "fuses to one side." Reading roll straight from the gravity vector
+// sidesteps that entirely: it depends only on which way is currently DOWN, which
+// the accelerometer measures directly, so it cannot drift, is naturally bounded
+// (no ±180° wrap runaway), and never touches the gyro (so it is also immune to
+// any gyro bias/sign error).
+//
+// This is the steering axis for lean-to-steer games (Tandemonium): the control
+// input is roll, yaw is irrelevant, so we never read yaw at all.
+//
+// Deliberately THREE-free (like yaw-return.js): the core test job runs
+// dependency-free (see .github/workflows/ci.yml), so this module and its tests
+// must not import three. The gravity vector is passed as plain components.
+// ============================================================
+
+/**
+ * Roll (bank) angle in radians from the sensor-local gravity/down vector.
+ *
+ * The vector is gravity expressed in the controller's OWN frame (≈ (0, -1, 0)
+ * at rest, world up = +Y) — exactly what SensorFusion tracks in `_gravityVec`.
+ * Roll is its tilt about the forward (Z) axis: the projection into the local
+ * X–Y plane, measured from straight-down. Depends ONLY on the current down
+ * direction, so it is immune to yaw drift and to gyro bias/sign errors.
+ *
+ * Sign matches the Euler-Z convention SensorFusion consumers already use
+ * (`leanDeg = -eulerZ`): for a pure roll of φ about Z the orientation is
+ * R_z(φ), the local gravity vector is (-sinφ, -cosφ, 0), and this returns -φ —
+ * exactly what `-eulerZ` gives. So it is a drop-in replacement that removes the
+ * drift WITHOUT changing steering direction. Pitch does not leak in (a pure
+ * pitch θ gives gravity (0, -cosθ, sinθ) → atan2(0, cosθ) = 0).
+ *
+ * @param {number} gravX local gravity vector X
+ * @param {number} gravY local gravity vector Y (≈ -1 at rest)
+ * @returns {number} roll in radians, in (-π, π]
+ */
+export function rollFromGravity(gravX, gravY) {
+  return Math.atan2(gravX, -gravY);
+}
diff --git a/packages/core/src/sensor-fusion.js b/packages/core/src/sensor-fusion.js
index 0c586c0..1e165ac 100644
--- a/packages/core/src/sensor-fusion.js
+++ b/packages/core/src/sensor-fusion.js
@@ -42,6 +42,7 @@
 import * as THREE from 'three';
 
 import { twistAngleY, stepYawReturn, composeHeadingOffset } from './yaw-return.js';
+import { rollFromGravity } from './gravity-roll.js';
 
 // ── Initial one-shot bias calibration ──
 // Bluetooth captures are noisier (lower effective sample rate + packet
@@ -184,6 +185,18 @@ export class SensorFusion {
   /** Whether the initial one-shot bias calibration is still collecting. */
   get calibrating() { return this._calibrating; }
 
+  /**
+   * Roll (bank) angle in radians from the gravity-anchored down vector — the
+   * drift-free steering axis for lean-to-steer games. Reads only `_gravityVec`
+   * (accelerometer-tracked), never yaw (which a 6-axis IMU can't anchor) or the
+   * gyro, so it can't "fuse to one side" the way Euler-Z of the integrated
+   * orientation does. Sign matches the -EulerZ convention consumers already use,
+   * so it's a drop-in that only removes the drift. See gravity-roll.js (#314).
+   */
+  gravityRollRadians() {
+    return rollFromGravity(this._gravityVec.x, this._gravityVec.y);
+  }
+
   /**
    * Begin initial one-shot bias calibration. Clears any prior samples.
    * @param {string} [connectionType] — 'bluetooth' or 'usb'. BT gets a
diff --git a/packages/core/test/gravity-roll.test.js b/packages/core/test/gravity-roll.test.js
new file mode 100644
index 0000000..f96ebc2
--- /dev/null
+++ b/packages/core/test/gravity-roll.test.js
@@ -0,0 +1,88 @@
+// ============================================================
+// rollFromGravity — drift-free bank angle, sign + yaw-immunity locked (#314)
+// ============================================================
+//
+// This is the game's steering axis, so the sign and axis MUST be right. These
+// tests pin (a) that it matches the -EulerZ convention the game already steers
+// by — so switching the game's default 'gravity' mode onto it doesn't invert
+// steering — and (b) the whole point of the change: the value is independent of
+// yaw (heading), which is what stops the "fuses to one side" drift.
+//
+// Three-free (the core test job runs without npm install): quaternion math is
+// done with plain {x,y,z,w} objects. See [[multi-steam-controller]].
+
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+
+import { rollFromGravity } from '../src/gravity-roll.js';
+
+const DEG = 180 / Math.PI;
+const approx = (a, b, eps = 1e-9) => Math.abs(a - b) <= eps;
+
+// ── Minimal quaternion helpers (no three) ──
+function qAxis(ax, ay, az, ang) { const h = ang / 2, s = Math.sin(h); return { x: ax * s, y: ay * s, z: az * s, w: Math.cos(h) }; }
+function qMul(a, b) {
+  return {
+    w: a.w * b.w - a.x * b.x - a.y * b.y - a.z * b.z,
+    x: a.w * b.x + a.x * b.w + a.y * b.z - a.z * b.y,
+    y: a.w * b.y - a.x * b.z + a.y * b.w + a.z * b.x,
+    z: a.w * b.z + a.x * b.y - a.y * b.x + a.z * b.w,
+  };
+}
+const qConj = (q) => ({ x: -q.x, y: -q.y, z: -q.z, w: q.w });
+function rot(q, v) { // rotate vector v by quaternion q: q·v·q⁻¹
+  const p = { x: v.x, y: v.y, z: v.z, w: 0 };
+  const r = qMul(qMul(q, p), qConj(q));
+  return { x: r.x, y: r.y, z: r.z };
+}
+const Rx = (a) => qAxis(1, 0, 0, a), Ry = (a) => qAxis(0, 1, 0, a), Rz = (a) => qAxis(0, 0, 1, a);
+const DOWN = { x: 0, y: -1, z: 0 };
+// SensorFusion's _gravityVec = orientation⁻¹ · worldDown (gravity in the local frame).
+const gravLocal = (orientation) => rot(qConj(orientation), DOWN);
+
+test('at rest (down = (0,-1,0)) → roll 0', () => {
+  assert.equal(rollFromGravity(0, -1), 0);
+});
+
+test('pure roll φ → -φ (matches the -EulerZ steering convention, drop-in)', () => {
+  for (const deg of [-80, -45, -20, -5, 0, 5, 20, 45, 80]) {
+    const phi = deg / DEG;
+    const g = gravLocal(Rz(phi));              // orientation = R_z(φ)
+    assert.ok(approx(rollFromGravity(g.x, g.y), -phi), `roll ${deg}° → ${-deg}° (got ${(rollFromGravity(g.x, g.y) * DEG).toFixed(3)})`);
+  }
+});
+
+test('YAW-IMMUNE: roll is identical at every heading (the fix for "fuses to one side")', () => {
+  const phi = 30 / DEG;
+  const expected = -phi;
+  for (const yawDeg of [0, 30, 90, 175, 250, 359]) {
+    const orientation = qMul(Ry(yawDeg / DEG), Rz(phi)); // yaw ∘ roll
+    const g = gravLocal(orientation);
+    assert.ok(approx(rollFromGravity(g.x, g.y), expected),
+      `yaw ${yawDeg}° must not change roll (got ${(rollFromGravity(g.x, g.y) * DEG).toFixed(3)}°, want ${(expected * DEG).toFixed(3)}°)`);
+  }
+});
+
+test('drift immunity: a large accumulated yaw does not move roll', () => {
+  // Simulate 5 full turns of yaw drift on top of a 15° roll — Euler-Z would have
+  // wandered; gravity-roll must not budge.
+  const phi = 15 / DEG;
+  const g0 = gravLocal(Rz(phi));
+  const gDrifted = gravLocal(qMul(Ry(5 * 2 * Math.PI + 1.234), Rz(phi)));
+  assert.ok(approx(rollFromGravity(g0.x, g0.y), rollFromGravity(gDrifted.x, gDrifted.y)));
+});
+
+test('pure pitch θ → roll 0 (pitch does not leak into roll)', () => {
+  for (const deg of [-60, -30, -5, 0, 5, 30, 60]) {
+    const g = gravLocal(Rx(deg / DEG));
+    assert.ok(approx(rollFromGravity(g.x, g.y), 0), `pitch ${deg}° → roll 0 (got ${(rollFromGravity(g.x, g.y) * DEG).toFixed(3)}°)`);
+  }
+});
+
+test('sign: rolling right and left are opposite and symmetric', () => {
+  const gr = gravLocal(Rz(20 / DEG));
+  const gl = gravLocal(Rz(-20 / DEG));
+  const r = rollFromGravity(gr.x, gr.y);
+  const l = rollFromGravity(gl.x, gl.y);
+  assert.ok(r < 0 && l > 0 && approx(r, -l), `right=${(r * DEG).toFixed(2)}° left=${(l * DEG).toFixed(2)}° should be opposite`);
+});