diff --git a/Sources/ForgeColor/Lut3D.swift b/Sources/ForgeColor/Lut3D.swift
new file mode 100644
index 0000000..515b97f
--- /dev/null
+++ b/Sources/ForgeColor/Lut3D.swift
@@ -0,0 +1,132 @@
+import Foundation
+
+/// 3D LUT lattice. Table layout matches .cube: red index fastest, then green, then blue.
+/// Entry (r,g,b) at index b*size² + g*size + r, three floats per entry.
+public struct Lut3D: Sendable {
+    public let size: Int
+    /// size³ * 3 floats.
+    public let table: [Float]
+
+    public init(size: Int, table: [Float]) {
+        precondition(size >= 2, "LUT size must be >= 2")
+        precondition(table.count == size * size * size * 3, "table count mismatch")
+        self.size = size
+        self.table = table
+    }
+
+    public static func identity(size: Int) -> Lut3D {
+        build(size: size) { $0 }
+    }
+
+    /// Build lattice by evaluating a function at each grid point.
+    public static func build(size: Int, function: (SIMD3<Float>) -> SIMD3<Float>) -> Lut3D {
+        var table = [Float]()
+        table.reserveCapacity(size * size * size * 3)
+        let denom = Float(size - 1)
+        for b in 0..<size {
+            for g in 0..<size {
+                for r in 0..<size {
+                    let out = function(SIMD3(Float(r) / denom, Float(g) / denom, Float(b) / denom))
+                    table.append(out.x)
+                    table.append(out.y)
+                    table.append(out.z)
+                }
+            }
+        }
+        return Lut3D(size: size, table: table)
+    }
+
+    @inline(__always)
+    private func entry(_ r: Int, _ g: Int, _ b: Int) -> SIMD3<Float> {
+        let i = ((b * size + g) * size + r) * 3
+        return SIMD3(table[i], table[i + 1], table[i + 2])
+    }
+
+    /// Trilinear interpolation; input clamped to [0,1].
+    public func sample(_ rgb: SIMD3<Float>) -> SIMD3<Float> {
+        let maxIdx = Float(size - 1)
+        let clamped = SIMD3(
+            min(max(rgb.x, 0), 1),
+            min(max(rgb.y, 0), 1),
+            min(max(rgb.z, 0), 1))
+        let pos = clamped * maxIdx
+        let i0 = SIMD3<Int>(Int(pos.x), Int(pos.y), Int(pos.z))
+        let i1 = SIMD3<Int>(
+            min(i0.x + 1, size - 1),
+            min(i0.y + 1, size - 1),
+            min(i0.z + 1, size - 1))
+        let f = pos - SIMD3(Float(i0.x), Float(i0.y), Float(i0.z))
+
+        // 8 corners; lerp r, then g, then b.
+        let c000 = entry(i0.x, i0.y, i0.z), c100 = entry(i1.x, i0.y, i0.z)
+        let c010 = entry(i0.x, i1.y, i0.z), c110 = entry(i1.x, i1.y, i0.z)
+        let c001 = entry(i0.x, i0.y, i1.z), c101 = entry(i1.x, i0.y, i1.z)
+        let c011 = entry(i0.x, i1.y, i1.z), c111 = entry(i1.x, i1.y, i1.z)
+
+        let c00 = c000 + (c100 - c000) * f.x
+        let c10 = c010 + (c110 - c010) * f.x
+        let c01 = c001 + (c101 - c001) * f.x
+        let c11 = c011 + (c111 - c011) * f.x
+        let c0 = c00 + (c10 - c00) * f.y
+        let c1 = c01 + (c11 - c01) * f.y
+        return c0 + (c1 - c0) * f.z
+    }
+}
+
+// MARK: - .cube I/O
+
+public enum CubeError: Error, Equatable {
+    case missingSize
+    case invalidFloat(line: Int)
+    case entryCountMismatch(expected: Int, got: Int)
+    case invalidSize
+}
+
+public enum CubeFile {
+    /// Serialize to Resolve/IRIDAS .cube text.
+    public static func write(_ lut: Lut3D, title: String = "Forge") -> String {
+        var out = "TITLE \"\(title)\"\nLUT_3D_SIZE \(lut.size)\n"
+        out.reserveCapacity(lut.table.count * 10)
+        var i = 0
+        while i < lut.table.count {
+            out += String(format: "%.6f %.6f %.6f\n", lut.table[i], lut.table[i + 1], lut.table[i + 2])
+            i += 3
+        }
+        return out
+    }
+
+    /// Parse .cube text. Ignores TITLE/DOMAIN_MIN/DOMAIN_MAX/comments/blank lines.
+    public static func parse(_ text: String) throws -> Lut3D {
+        var size: Int?
+        var values = [Float]()
+        var lineNo = 0
+        for rawLine in text.split(separator: "\n", omittingEmptySubsequences: false) {
+            lineNo += 1
+            let line = rawLine.trimmingCharacters(in: .whitespaces)
+            if line.isEmpty || line.hasPrefix("#") { continue }
+            if line.hasPrefix("TITLE") || line.hasPrefix("DOMAIN_MIN") || line.hasPrefix("DOMAIN_MAX")
+                || line.hasPrefix("LUT_1D_SIZE") {
+                continue
+            }
+            if line.hasPrefix("LUT_3D_SIZE") {
+                let parts = line.split(separator: " ")
+                guard parts.count == 2, let s = Int(parts[1]), s >= 2 else { throw CubeError.invalidSize }
+                size = s
+                continue
+            }
+            // Data row: three floats.
+            let comps = line.split(separator: " ", omittingEmptySubsequences: true)
+            guard comps.count == 3 else { throw CubeError.invalidFloat(line: lineNo) }
+            for c in comps {
+                guard let v = Float(c) else { throw CubeError.invalidFloat(line: lineNo) }
+                values.append(v)
+            }
+        }
+        guard let s = size else { throw CubeError.missingSize }
+        let expected = s * s * s * 3
+        guard values.count == expected else {
+            throw CubeError.entryCountMismatch(expected: expected, got: values.count)
+        }
+        return Lut3D(size: s, table: values)
+    }
+}
diff --git a/Tests/ForgeColorTests/Lut3DTests.swift b/Tests/ForgeColorTests/Lut3DTests.swift
new file mode 100644
index 0000000..7a2ebb7
--- /dev/null
+++ b/Tests/ForgeColorTests/Lut3DTests.swift
@@ -0,0 +1,120 @@
+import XCTest
+@testable import ForgeColor
+
+final class Lut3DTests: XCTestCase {
+
+    // Identity lattice is a no-op within interp tolerance.
+    func testIdentityNoOp() {
+        let lut = Lut3D.identity(size: 17)
+        for px in [SIMD3<Float>(0, 0, 0), SIMD3(1, 1, 1), SIMD3(0.5, 0.25, 0.75), SIMD3(0.18, 0.18, 0.18)] {
+            let out = lut.sample(px)
+            XCTAssertEqual(out.x, px.x, accuracy: 1e-5)
+            XCTAssertEqual(out.y, px.y, accuracy: 1e-5)
+            XCTAssertEqual(out.z, px.z, accuracy: 1e-5)
+        }
+    }
+
+    // Lattice built from an analytic function: interp matches direct eval on off-lattice points.
+    func testTrilinearMatchesDirectEval() {
+        // f = gain 0.5 (linear function -> trilinear should be near-exact)
+        let f: (SIMD3<Float>) -> SIMD3<Float> = { $0 * 0.5 }
+        let lut = Lut3D.build(size: 9, function: f)
+        for px in [SIMD3<Float>(0.13, 0.57, 0.91), SIMD3(0.9, 0.05, 0.5)] {
+            let out = lut.sample(px)
+            let direct = f(px)
+            XCTAssertEqual(out.x, direct.x, accuracy: 1e-4)
+            XCTAssertEqual(out.y, direct.y, accuracy: 1e-4)
+            XCTAssertEqual(out.z, direct.z, accuracy: 1e-4)
+        }
+    }
+
+    // Nonlinear function: 33^3 approx within 1/1024 where trilinear error bound
+    // (max|f''|·h²/8, h=1/32) actually permits it. pow(x, 0.6) has f''→∞ at 0;
+    // at x=0.15, |f''|≈3.4 → bound ≈4e-4. Real flattening runs in log space where
+    // curvature is gentle everywhere. Samples in [0.15, 1].
+    func testNonlinearApproxWithinTolerance() {
+        let f: (SIMD3<Float>) -> SIMD3<Float> = {
+            SIMD3(pow($0.x, 2.2), pow($0.y, 1.8), pow($0.z, 0.6))
+        }
+        let lut = Lut3D.build(size: 33, function: f)
+        var rng = SystemRandomNumberGenerator()
+        for _ in 0..<200 {
+            let px = SIMD3<Float>(
+                Float.random(in: 0.15...1, using: &rng),
+                Float.random(in: 0.15...1, using: &rng),
+                Float.random(in: 0.15...1, using: &rng))
+            let out = lut.sample(px)
+            let direct = f(px)
+            XCTAssertEqual(out.x, direct.x, accuracy: 1.0 / 1024)
+            XCTAssertEqual(out.y, direct.y, accuracy: 1.0 / 1024)
+            XCTAssertEqual(out.z, direct.z, accuracy: 1.0 / 1024)
+        }
+    }
+
+    // Out-of-range input clamps to lattice bounds.
+    func testInputClamped() {
+        let lut = Lut3D.identity(size: 5)
+        let out = lut.sample(SIMD3(-0.5, 1.5, 0.5))
+        XCTAssertEqual(out.x, 0, accuracy: 1e-5)
+        XCTAssertEqual(out.y, 1, accuracy: 1e-5)
+        XCTAssertEqual(out.z, 0.5, accuracy: 1e-5)
+    }
+
+    // MARK: CUBE I/O
+
+    func testCubeRoundTrip() throws {
+        let lut = Lut3D.build(size: 5) { $0 * 0.8 + SIMD3(0.05, 0, 0.1) }
+        let text = CubeFile.write(lut, title: "Forge Test")
+        let parsed = try CubeFile.parse(text)
+        XCTAssertEqual(parsed.size, 5)
+        for i in 0..<lut.table.count {
+            XCTAssertEqual(parsed.table[i], lut.table[i], accuracy: 1e-6)
+        }
+    }
+
+    func testCubeWriteFormat() {
+        let lut = Lut3D.identity(size: 2)
+        let text = CubeFile.write(lut, title: "T")
+        XCTAssertTrue(text.contains("TITLE \"T\""))
+        XCTAssertTrue(text.contains("LUT_3D_SIZE 2"))
+        // First lattice entry (0,0,0), red fastest.
+        XCTAssertTrue(text.contains("0.000000 0.000000 0.000000"))
+        XCTAssertTrue(text.hasSuffix("\n"))
+    }
+
+    func testCubeParseSkipsCommentsAndDomain() throws {
+        let text = """
+        # comment line
+        TITLE "X"
+        LUT_3D_SIZE 2
+        DOMAIN_MIN 0.0 0.0 0.0
+        DOMAIN_MAX 1.0 1.0 1.0
+
+        0 0 0
+        1 0 0
+        0 1 0
+        1 1 0
+        0 0 1
+        1 0 1
+        0 1 1
+        1 1 1
+        """
+        let lut = try CubeFile.parse(text)
+        XCTAssertEqual(lut.size, 2)
+        XCTAssertEqual(lut.sample(SIMD3(1, 1, 1)).x, 1, accuracy: 1e-6)
+    }
+
+    func testCubeParseRejectsMissingSize() {
+        XCTAssertThrowsError(try CubeFile.parse("0 0 0\n1 1 1\n"))
+    }
+
+    func testCubeParseRejectsWrongEntryCount() {
+        let text = "LUT_3D_SIZE 2\n0 0 0\n1 1 1\n" // needs 8 entries, has 2
+        XCTAssertThrowsError(try CubeFile.parse(text))
+    }
+
+    func testCubeParseRejectsBadFloat() {
+        let text = "LUT_3D_SIZE 2\n" + Array(repeating: "0 0 zebra", count: 8).joined(separator: "\n")
+        XCTAssertThrowsError(try CubeFile.parse(text))
+    }
+}