diff --git a/zig/build/cuda.zig b/zig/build/cuda.zig
index 684d6c3cc9..0c97f92dbf 100644
--- a/zig/build/cuda.zig
+++ b/zig/build/cuda.zig
@@ -240,6 +240,8 @@ pub fn hostTests(b: *std.Build, draft_ids: *std.Build.Module, build_options: *st
     step.dependOn(&b.addRunArtifact(b.addTest(.{ .root_module = b.createModule(.{ .root_source_file = b.path("zig/build/check_fatbins.zig"), .target = host }) })).step);
     const cuda = runtime(b, host, .debug, &.{});
     const mods = family(b, host, .debug, cuda, draft_ids);
+    const contract_arch = b.createModule(.{ .root_source_file = b.path("zig/tests/cuda/contract_arch_test.zig"), .target = host, .optimize = .debug, .link_libc = true, .imports = &.{ .{ .name = "cuda", .module = cuda }, .{ .name = "core", .module = mods.core } } });
+    step.dependOn(&b.addRunArtifact(b.addTest(.{ .root_module = contract_arch })).step);
     const native_modules = engines(b, host, .debug, cuda, mods.core, mods.lanes, mods.nemotron, mods.heat);
     const native = native_modules.engines;
     // the HTTP server's unit tests (zig/src/server/root.zig), which the macOS build runs, on Linux hosts too
diff --git a/zig/tests/cuda/contract_arch_test.zig b/zig/tests/cuda/contract_arch_test.zig
new file mode 100644
index 0000000000..d355c8d8ba
--- /dev/null
+++ b/zig/tests/cuda/contract_arch_test.zig
@@ -0,0 +1,123 @@
+//! Host-only checks of the CUDA contract's architecture selection and its real gate-3 routing.
+const std = @import("std");
+const mm = @import("contract_mm.zig");
+const contract = @import("core").contract;
+const g = contract.gates;
+
+test "contract architecture: older SMs omit only the standalone cluster kernel" {
+    for ([_]u32{ 80, 86, 89 }) |sm| {
+        var count: usize = 0;
+        for (std.enums.values(mm.Kind)) |kind| {
+            var m: mm.Mm = .{ .rig = undefined, .kind = kind, .sm = sm };
+            const k = m.kernel();
+            try std.testing.expectEqual(kind == .qmm_group, k == null);
+            if (k != null) count += 1;
+        }
+        try std.testing.expectEqual(@as(usize, 6), count);
+    }
+}
+
+test "contract architecture: SM90 and Blackwell retain every kernel and the GEMV bit comparisons" {
+    for ([_]u32{ 90, 120, 121 }) |sm| {
+        var count: usize = 0;
+        for (std.enums.values(mm.Kind)) |kind| {
+            var m: mm.Mm = .{ .rig = undefined, .kind = kind, .sm = sm };
+            const k = m.kernel().?;
+            count += 1;
+            try std.testing.expectEqual(kind == .lane_gemv or kind == .lane_gemv_split, k.spec.replaces != null);
+        }
+        try std.testing.expectEqual(@as(usize, 7), count);
+    }
+}
+
+test "contract architecture: legacy GEMV keeps every field except the unavailable replacement" {
+    for ([_]u32{ 80, 86, 89 }) |sm| for ([_]mm.Kind{ .lane_gemv, .lane_gemv_split }) |kind| {
+        var modern: mm.Mm = .{ .rig = undefined, .kind = kind, .sm = 121 };
+        var legacy: mm.Mm = .{ .rig = undefined, .kind = kind, .sm = sm };
+        var want = modern.kernel().?.spec;
+        want.replaces = null;
+        try std.testing.expectEqualDeep(want, legacy.kernel().?.spec);
+    };
+}
+
+test "contract architecture: the legacy FP4 lane names the reduction fallback it actually selects" {
+    for ([_]u32{ 80, 86, 89 }) |sm| {
+        var m: mm.Mm = .{ .rig = undefined, .kind = .lane_f32, .sm = sm };
+        const k = m.kernel().?;
+        try std.testing.expect(std.mem.indexOf(u8, k.spec.name, "reduce_kernel") != null);
+    }
+}
+
+const Cpu = struct {
+    bytes: [2]u8 = .{ 0x80, 0x3f }, // bf16 1
+    baseline: [2]u8 = .{ 0x80, 0x3f },
+    references: usize = 0,
+    replacements: usize = 0,
+    const vt: g.VTable = .{ .prepare = prepare, .run = run, .output = output, .inputs = inputs, .intact = intact, .overlap = overlap, .count = count, .reference = reference, .replaced = replaced };
+
+    fn self(ctx: *anyopaque) *Cpu {
+        return @ptrCast(@alignCast(ctx));
+    }
+    fn prepare(_: *anyopaque, _: contract.spec.Shape, _: g.Fill, _: u64) anyerror!bool {
+        return true;
+    }
+    fn run(_: *anyopaque) anyerror!void {}
+    fn output(ctx: *anyopaque) anyerror![]const u8 {
+        return &self(ctx).bytes;
+    }
+    fn inputs(_: *anyopaque) anyerror![]const u8 {
+        return &.{};
+    }
+    fn intact(_: *anyopaque) anyerror!bool {
+        return true;
+    }
+    fn overlap(_: *anyopaque) bool {
+        return false;
+    }
+    fn count(_: *anyopaque) usize {
+        return 1;
+    }
+    fn reference(ctx: *anyopaque, want: []f64, mag: []f64) anyerror!void {
+        self(ctx).references += 1;
+        want[0] = 1;
+        mag[0] = 1;
+    }
+    fn replaced(ctx: *anyopaque) anyerror![]const u8 {
+        self(ctx).replacements += 1;
+        return &self(ctx).baseline;
+    }
+};
+
+test "contract architecture: legacy gate 3 really checks FP64 and rejects a planted bad output" {
+    const old = g.quiet;
+    g.quiet = true;
+    defer g.quiet = old;
+    for ([_]mm.Kind{ .lane_gemv, .lane_gemv_split }) |kind| {
+        var m: mm.Mm = .{ .rig = undefined, .kind = kind, .sm = 89 };
+        var cpu: Cpu = .{};
+        const k: g.Kernel = .{ .spec = m.kernel().?.spec, .ctx = &cpu, .vt = &Cpu.vt };
+        try std.testing.expectEqual(g.Verdict.pass, try g.shapes(std.testing.allocator, k, 1));
+        try std.testing.expect(cpu.references > 0);
+        try std.testing.expectEqual(@as(usize, 0), cpu.replacements);
+        cpu.bytes = .{ 0, 0 };
+        try std.testing.expectEqual(g.Verdict.fail, try g.shapes(std.testing.allocator, k, 1));
+        try std.testing.expectEqual(@as(usize, 0), cpu.replacements);
+    }
+}
+
+test "contract architecture: modern gate 3 retains bit equality and rejects a planted baseline difference" {
+    const old = g.quiet;
+    g.quiet = true;
+    defer g.quiet = old;
+    for ([_]mm.Kind{ .lane_gemv, .lane_gemv_split }) |kind| {
+        var m: mm.Mm = .{ .rig = undefined, .kind = kind, .sm = 121 };
+        var cpu: Cpu = .{};
+        const k: g.Kernel = .{ .spec = m.kernel().?.spec, .ctx = &cpu, .vt = &Cpu.vt };
+        try std.testing.expectEqual(g.Verdict.pass, try g.shapes(std.testing.allocator, k, 1));
+        try std.testing.expect(cpu.replacements > 0);
+        try std.testing.expectEqual(@as(usize, 0), cpu.references);
+        cpu.baseline = .{ 0, 0 };
+        try std.testing.expectEqual(g.Verdict.fail, try g.shapes(std.testing.allocator, k, 1));
+        try std.testing.expectEqual(@as(usize, 0), cpu.references);
+    }
+}
diff --git a/zig/tests/cuda/contract_mm.zig b/zig/tests/cuda/contract_mm.zig
index fb97d89818..89773961d5 100644
--- a/zig/tests/cuda/contract_mm.zig
+++ b/zig/tests/cuda/contract_mm.zig
@@ -55,6 +55,7 @@ pub fn spec(kind: Kind) c.Spec {
 pub const Mm = struct {
     rig: rig_.Rig,
     kind: Kind,
+    sm: u32,
     affine4: ?*const qlinear.Affine4Kernels = null,
     split: ?*const qlinear.Split = null,
     prompt: ?*const nvfp4.Prompt = null,
@@ -75,8 +76,14 @@ pub const Mm = struct {
         self.rig.deinit();
     }
 
-    pub fn kernel(self: *Mm) g.Kernel {
-        return .{ .spec = spec(self.kind), .ctx = self, .vt = &vt };
+    pub fn kernel(self: *Mm) ?g.Kernel {
+        if (self.sm < 90 and self.kind == .qmm_group) return null;
+        var declared = spec(self.kind);
+        if (self.sm < 90) {
+            declared.replaces = null;
+            if (self.kind == .lane_f32) declared.name = "cuda qmmf FP4 fp32 out (slices by reduce_kernel<true>, architecture fallback)";
+        }
+        return .{ .spec = declared, .ctx = self, .vt = &vt };
     }
 
     const vt = g.hooks(Mm);
diff --git a/zig/tests/cuda/contract_tests.zig b/zig/tests/cuda/contract_tests.zig
index e87a2e90e5..da217fb961 100644
--- a/zig/tests/cuda/contract_tests.zig
+++ b/zig/tests/cuda/contract_tests.zig
@@ -82,10 +82,18 @@ pub fn run(gpu: check.Gpu) !void {
             .lane_f32_reduce => &lane_reduce,
             else => null,
         };
-        m.* = .{ .rig = .{ .gpu = gpu, .s = s }, .kind = kind, .affine4 = &affine4, .split = &split, .prompt = &prompt, .lane = by_lane };
+        m.* = .{ .rig = .{ .gpu = gpu, .s = s }, .kind = kind, .sm = cap, .affine4 = &affine4, .split = &split, .prompt = &prompt, .lane = by_lane };
     }
     defer for (&mm) |*m| m.deinit();
-    for (&mm) |*m| try kernels.append(a, m.kernel());
+    for (&mm) |*m| {
+        const k = m.kernel() orelse {
+            std.debug.print("N/A  cuda qmm_group (cluster tile): needs sm_90 or newer, this GPU is sm_{d}\n", .{cap});
+            continue;
+        };
+        if (cap < 90 and (m.kind == .lane_gemv or m.kind == .lane_gemv_split))
+            std.debug.print("N/A  {s} replacement bits: cluster baseline needs sm_90; gate 3 uses the float64 bound\n", .{k.spec.name});
+        try kernels.append(a, k);
+    }
     var rank = rank_k.Rank{ .rig = .{ .gpu = gpu, .s = s }, .sum = &rank_sum };
     defer rank.deinit();
     try kernels.append(a, rank.kernel());