diff --git a/docs/eager-execution-backends-and-kernels.md b/docs/eager-execution-backends-and-kernels.md
index 0544dadf..8f2aa33a 100644
--- a/docs/eager-execution-backends-and-kernels.md
+++ b/docs/eager-execution-backends-and-kernels.md
@@ -88,5 +88,7 @@ those formats were JVM-only and broke on Native.
 - ❌ **Other GGML quant formats** (Q5_K, Q2_K, Q3_K, Q8_K, IQ4_NL/XS) — loadable via dequant-to-FP32, but no packed matmul kernel.
 - ❌ **Non-CPU eager backends** (IREE, Metal, GPU) — the `KernelProvider` SPI anticipates them, but none are implemented for the eager path today.
 
-> Generated as a hand-authored overview. A machine-generated kernel × platform matrix
-> (derived from the registered `KernelProvider`s) is a planned follow-up so this stays in sync.
+> This mindmap is a hand-authored overview. Its companion
+> [kernel × platform support matrix](kernel-support-matrix.md) is **machine-generated** from
+> the registered `KernelProvider`s (`KernelSupportMatrixTest`) and CI-gated against drift in
+> the scalar floor, so the per-platform coverage stays in sync with the code.
diff --git a/docs/kernel-support-matrix.md b/docs/kernel-support-matrix.md
new file mode 100644
index 00000000..c1094a63
--- /dev/null
+++ b/docs/kernel-support-matrix.md
@@ -0,0 +1,20 @@
+# Kernel × platform support matrix
+
+> Generated by `KernelSupportMatrixTest`. The scalar (all-platform) coverage is
+> auto-derived from `KernelProvider.supports(...)`; re-run the test to refresh.
+> Cell = best available provider for `FP32 × format` on that platform.
+
+| Weight format | JVM | Android | Native·linux | Native·apple | JS/WASM |
+|---|:--:|:--:|:--:|:--:|:--:|
+| `Float32` | native-ffm | panama-vector | scalar | scalar | scalar |
+| `BFloat16` | native-ffm | panama-vector | scalar | scalar | scalar |
+| `Q8_0` | native-ffm | panama-vector | scalar | scalar | scalar |
+| `Q4_0` | native-ffm | panama-vector | scalar | scalar | scalar |
+| `Q4_K` | native-ffm | panama-vector | scalar | scalar | scalar |
+| `Q6_K` | scalar | scalar | scalar | scalar | scalar |
+| `Q5_1` | panama-vector | panama-vector | scalar | scalar | scalar |
+| `Q5_0` | panama-vector | panama-vector | scalar | scalar | scalar |
+
+Priority: native-ffm (100) → panama-vector (50) → scalar (0). Formats without any cell (e.g. Q5_K/Q2_K/Q3_K/IQ4) are dequant-to-FP32 only.
+
+See also the [eager backends & kernels mindmap](eager-execution-backends-and-kernels.md).
diff --git a/skainet-backends/skainet-backend-native-cpu/src/jvmTest/kotlin/sk/ainet/exec/kernel/KernelSupportMatrixTest.kt b/skainet-backends/skainet-backend-native-cpu/src/jvmTest/kotlin/sk/ainet/exec/kernel/KernelSupportMatrixTest.kt
new file mode 100644
index 00000000..c42ff611
--- /dev/null
+++ b/skainet-backends/skainet-backend-native-cpu/src/jvmTest/kotlin/sk/ainet/exec/kernel/KernelSupportMatrixTest.kt
@@ -0,0 +1,97 @@
+package sk.ainet.exec.kernel
+
+import java.io.File
+import kotlin.test.Test
+import kotlin.test.assertEquals
+import sk.ainet.backend.api.kernel.KernelProvider
+
+/**
+ * Generates the kernel × platform support matrix (rendered to
+ * `build/kernel-support-matrix.md`) and **gates drift in the scalar floor**: the
+ * all-platform baseline coverage is auto-derived from `ScalarKernelProvider.supports(...)`,
+ * so adding/removing a scalar packed kernel without updating the docs fails this test (it
+ * runs under `java-tests` in CI).
+ *
+ * The SIMD/native tiers (Panama, native-FFM) are env-availability-gated (`isAvailable()`
+ * probes the JDK incubator module / the loaded `.so`), so their *capability* is declared
+ * here (the single place to edit when a provider gains a kernel) rather than probed.
+ */
+class KernelSupportMatrixTest {
+
+    private val formats = listOf("Float32", "BFloat16", "Q8_0", "Q4_0", "Q4_K", "Q6_K", "Q5_1", "Q5_0")
+
+    private data class Tier(val name: String, val priority: Int, val targets: Set<String>, val formats: Set<String>)
+
+    // Source-set -> targets. commonMain reaches all; backend-cpu jvmMain -> {jvm,android};
+    // backend-native-cpu jvmMain -> {jvm} (the native module declares only jvm()).
+    private val allTargets = setOf("jvm", "android", "native-linux", "native-apple", "js-wasm")
+
+    private fun scalarFormats(): Set<String> =
+        formats.filter { ScalarKernelProvider.supports("matmul", listOf("Float32", it)) }.toSet()
+
+    private fun tiers(): List<Tier> = listOf(
+        Tier("scalar", 0, allTargets, scalarFormats()),
+        Tier("panama-vector", 50, setOf("jvm", "android"),
+            setOf("Float32", "BFloat16", "Q8_0", "Q4_0", "Q4_K", "Q5_1", "Q5_0")),
+        Tier("native-ffm", 100, setOf("jvm"),
+            setOf("Float32", "BFloat16", "Q8_0", "Q4_0", "Q4_K")),
+    )
+
+    private fun bestTier(fmt: String, target: String, tiers: List<Tier>): Tier? =
+        tiers.filter { target in it.targets && fmt in it.formats }.maxByOrNull { it.priority }
+
+    private fun render(tiers: List<Tier>): String {
+        val cols = listOf("jvm" to "JVM", "android" to "Android", "native-linux" to "Native·linux",
+            "native-apple" to "Native·apple", "js-wasm" to "JS/WASM")
+        val sb = StringBuilder()
+        sb.appendLine("# Kernel × platform support matrix")
+        sb.appendLine()
+        sb.appendLine("> Generated by `KernelSupportMatrixTest`. The scalar (all-platform) coverage is")
+        sb.appendLine("> auto-derived from `KernelProvider.supports(...)`; re-run the test to refresh.")
+        sb.appendLine("> Cell = best available provider for `FP32 × format` on that platform.")
+        sb.appendLine()
+        sb.append("| Weight format |")
+        cols.forEach { sb.append(" ${it.second} |") }
+        sb.appendLine()
+        sb.append("|---|")
+        cols.forEach { _ -> sb.append(":--:|") }
+        sb.appendLine()
+        for (fmt in formats) {
+            sb.append("| `$fmt` |")
+            for ((target, _) in cols) {
+                val t = bestTier(fmt, target, tiers)
+                sb.append(" ${t?.name ?: "—"} |")
+            }
+            sb.appendLine()
+        }
+        sb.appendLine()
+        sb.appendLine("Priority: native-ffm (100) → panama-vector (50) → scalar (0). " +
+            "Formats without any cell (e.g. Q5_K/Q2_K/Q3_K/IQ4) are dequant-to-FP32 only.")
+        sb.appendLine()
+        sb.appendLine("See also the [eager backends & kernels mindmap](eager-execution-backends-and-kernels.md).")
+        return sb.toString()
+    }
+
+    @Test
+    fun generate_and_gate_support_matrix() {
+        val tiers = tiers()
+
+        // Drift gate on the scalar floor (the all-platform baseline): the documented set
+        // below must equal what the scalar provider actually carries. Update both together.
+        assertEquals(
+            setOf("Float32", "BFloat16", "Q8_0", "Q4_0", "Q4_K", "Q6_K", "Q5_1", "Q5_0"),
+            scalarFormats(),
+            "ScalarKernelProvider coverage changed — update the matrix doc + this expected set",
+        )
+
+        // Sanity: every provider singleton is a KernelProvider (compile-time anchor).
+        val providers: List<KernelProvider> = listOf(ScalarKernelProvider, PanamaVectorKernelProvider, NativeKernelProvider)
+        assertEquals(3, providers.size)
+
+        val md = render(tiers)
+        File("build").mkdirs()
+        File("build/kernel-support-matrix.md").writeText(md)
+        // Echo so CI logs carry the current matrix (easy to copy into docs/).
+        println("KERNEL_SUPPORT_MATRIX_BEGIN\n$md\nKERNEL_SUPPORT_MATRIX_END")
+    }
+}