michalharakal · Oct 21, 2026 · Oct 21, 2026 · Oct 21, 2026 · Oct 21, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -11,13 +11,32 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.63.0] — 2026-10-21
+
+Against engine **0.62.0** (the JVM ↔ Kotlin/Wasm parity release: every kernel on the wasm tier, resident
+activations, fused attention / norms / softmax through FFM on the JVM and in the SIMD128 sidecar, browser IO
+parity, `skainet-data-source` on JS and wasm). Headline: **expert heads and routing for Laya** — a decision
+head fine-tuned on top of the frozen encoder loads on its own, any number of heads share one encoder pass
+per question, a prompted router picks one per request, and a mixture stands behind the typed decision
+contract. Verified against the published layaMOE evaluation. Plus the whisper StableHLO export as a shipped
+CLI.
+
+Version note: every API change is additive. The Laya Q8_0 lane on the desktop JVM with
+`skainet-backend-native-cpu` on the classpath needs engine 0.62.1 (SKaiNET #1436: the native Q8_0 / Q4_0 /
+Q5_0 / Q5_1 bindings zeroed every row but the last of a multi-row product); the FP32 lane and the browser's
+Q8_0 lane are unaffected.
+
 ### Added
 
+- **iree-android: a Whisper session over the `WhisperExportCli` graphs** — `IreeWhisperSession` with the prebuilt `libskainet_whisper_session.so`, the README section and `IreeWhisperContractTest` (#512).
+- **Whisper export ships as `WhisperExportCli`** in the whisper module's `jvmMain` (was `jvmTest`), like `MoonshineV2ExportCli`: `WhisperExportHarness` is configurable (`WhisperConfig` + `WhisperExportOptions`: audio context, KV length, prompt length, manifest language, parameter scope), `configFromHf` reads geometry and vocabulary from a Hugging Face `config.json` (tiny, base, small and `*.en`), and the Gradle task `exportWhisper` is env-driven (`WHISPER_MODEL_DIR`, `WHISPER_OUT_DIR`, `WHISPER_AUDIO_CTX`, `WHISPER_MAX_POSITIONS`, `WHISPER_LANGUAGE`), so a blueprint can produce the encoder, masked prefill and KV-step graphs with one merged `params.irpa` from the released artifact (#510).
+
 - **Expert heads and routing for Laya** (`sk.ainet.models.laya`): a Laya decision head can be loaded on its own and run on the base model's encoder output. `LayaExpertLoader.load(metaJson, weights, base, ctx)` binds an `expert.safetensors` (the `type_emb.*` / `head.*` / `scorer.*` tensors of a head copied from the base checkpoint and fine-tuned for a domain) with its `expert.json` (name, domains, calibration temperatures) into a `LayaExpert`; `LayaWeightPreparation.prepareHead` and `LayaDecisionModel.buildHead` are the steps underneath, and a head prepared on its own reproduces the base model bit for bit. `LayaDecisionModel.logitsFrom` / `logitsBatchFrom` and `LayaAgent.encodeFor` / `logitsFrom` run any `LayaDecisionHead` on an encoder output computed once, so several heads answer the same sequences for one encoder pass each. `LayaRouterSpec` is the prompted router (the base head answers a choice over kinds of text with descriptions; kind → expert; threshold; `general` fallback), `LayaMixtureAgent.decide(state, questions, expert, compare)` routes and answers with the chosen head (forced head reported as overridden; `compare` adds the base head's answers on the same encoder output; encoder and head timings), and `LayaMixtureDecisionAdapter` puts a mixture behind the typed decision contract with `onRouted` reporting the head that answered. Verified against the published evaluation of layaMOE (gated `LayaMoeEvalTest`, fixtures under `laya-moe/`, Apache-2.0).
 - `LayaAgent.decide` / `decideAll` / `logitsFor` overloads that take a head and a decoder; `LayaAgent.decodeAll`; `LayaDecisionModel.encodeBatch(sequences, padId)` returning a `LayaEncodedBatch`; `LayaDecisionModel.defaultHead`. The existing signatures are unchanged.
 
 ### Changed
 
+- Engine pin **SKaiNET 0.62.0** (`skainet-bom:0.62.0` re-exported by the BOM); kotest 6.2.6. The llama `ForwardScopeSteadyStateTest` pins the scoped decode to within four ulp of the unscoped one instead of raw-bit identity: since 0.62.0 the JVM ops route slab-backed and heap activations through kernels whose accumulation order can differ by one ulp.
 - `LayaEncoderConfig.requireSupported()` accepts `position_embedding_type: "absolute"` beside `"sans_pos"`: ModernBERT positions are rotary whatever the field says, and transformers 5 exports (`convaiinnovations/laya-typed-decisions`) write `absolute`. The English checkpoint loads and its `sequences.json` section reproduces (gated on `LAYA_TYPED_DECISIONS_PATH`).
 
 ## [0.62.2] — 2026-10-18

diff --git a/README.md b/README.md
@@ -113,7 +113,9 @@ Honest status — see the project-status note at the top of this README.
 
 ## Current release
 
-The current release is **0.62.2**, against **SKaiNET 0.61.0**: a transformers-only fix, the Android runtime's streaming Moonshine libraries are rebuilt with the utterance-start fix (#500; 0.62.0 and 0.62.1 packed the old build, which garbles transcripts on device). It builds on **0.62.1**, a transformers-only fix on 0.62.0, whose JVM `LayaModelLoader.load(directory, …)` had become an extension function that existing callers did not import; it is a member again (`@JvmOverloads`, so Java callers keep the 0.61 signature).
+The current release is **0.63.0**, against **SKaiNET 0.62.0**: expert heads and routing for Laya — a decision head fine-tuned on top of the frozen encoder (`expert.safetensors` + `expert.json`) loads on its own with `LayaExpertLoader`, any number of heads share one encoder pass per question (`LayaDecisionModel.logitsFrom`), `LayaRouterSpec` + `LayaMixtureAgent` route a request by asking the base head what kind of text it is, and `LayaMixtureDecisionAdapter` puts the mixture behind the typed decision contract; verified against the published layaMOE evaluation. The ModernBERT-large `laya-typed-decisions` checkpoint loads (`position_embedding_type: absolute`). The whisper StableHLO export ships as `WhisperExportCli`. Engine 0.62.0 brings the JVM ↔ Kotlin/Wasm parity: every kernel on the wasm tier, resident activations, fused attention / norms / softmax, browser IO parity. The Laya Q8_0 lane on the desktop JVM with the native provider needs engine 0.62.1 (SKaiNET #1436).
+
+It builds on **0.62.2**, against **SKaiNET 0.61.0**: a transformers-only fix, the Android runtime's streaming Moonshine libraries are rebuilt with the utterance-start fix (#500; 0.62.0 and 0.62.1 packed the old build, which garbles transcripts on device). It builds on **0.62.1**, a transformers-only fix on 0.62.0, whose JVM `LayaModelLoader.load(directory, …)` had become an extension function that existing callers did not import; it is a member again (`@JvmOverloads`, so Java callers keep the 0.61 signature).
 
 It builds on **0.62.0**, against **SKaiNET 0.61.0**:
 
@@ -301,7 +303,7 @@ The recommended way to consume is via the BOM. It pins every published `skainet-
 
 ```kotlin
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.62.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.63.0"))
 
     // Versions resolved from the BOM:
     implementation("sk.ainet.transformers:skainet-transformers-core")

diff --git a/docs/modules/ROOT/pages/how-to/export-moonshine-v2-streaming.adoc b/docs/modules/ROOT/pages/how-to/export-moonshine-v2-streaming.adoc
@@ -32,7 +32,7 @@ From your own build, resolve the published module and run its main class:
 ----
 val exportTool by configurations.creating
 dependencies {
-    exportTool(platform("sk.ainet.transformers:skainet-transformers-bom:0.62.2"))
+    exportTool(platform("sk.ainet.transformers:skainet-transformers-bom:0.63.0"))
     exportTool("sk.ainet.transformers:skainet-transformers-inference-moonshine")
 }
 tasks.register<JavaExec>("exportMoonshine") {

diff --git a/docs/modules/ROOT/pages/reference/moonshine-encoder.adoc b/docs/modules/ROOT/pages/reference/moonshine-encoder.adoc
@@ -24,7 +24,7 @@ a DSL decoder is future work.
 [source,kotlin]
 ----
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.62.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.63.0"))
     implementation("sk.ainet.transformers:skainet-transformers-inference-moonshine")
 }
 ----

diff --git a/docs/modules/ROOT/pages/reference/typed-decisions.adoc b/docs/modules/ROOT/pages/reference/typed-decisions.adoc
@@ -12,7 +12,7 @@ below put that on the JVM path with the same contract for every model.
 [source,kotlin]
 ----
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.62.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.63.0"))
     implementation("sk.ainet.transformers:skainet-transformers-decisions-api")      // the contract, no engine dependency
     implementation("sk.ainet.transformers:skainet-transformers-decisions-runtime")  // validation + acceptance policy
     implementation("sk.ainet.transformers:skainet-transformers-inference-laya")     // Laya as a DecisionAdapter

diff --git a/docs/modules/ROOT/pages/tutorials/android-getting-started.adoc b/docs/modules/ROOT/pages/tutorials/android-getting-started.adoc
@@ -55,7 +55,7 @@ dependencies {
     // self-registers it on ART at process start — nothing to call.
     runtimeOnly("sk.ainet.core:skainet-backend-jni-cpu")
 
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.62.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.63.0"))
     implementation("sk.ainet.transformers:skainet-transformers-core")
     implementation("sk.ainet.transformers:skainet-transformers-runtime-kllama")
     implementation("sk.ainet.transformers:skainet-transformers-inference-llama")

diff --git a/docs/modules/ROOT/pages/tutorials/getting-started-java.adoc b/docs/modules/ROOT/pages/tutorials/getting-started-java.adoc
@@ -25,7 +25,7 @@ In your `build.gradle.kts`:
 [source,kotlin]
 ----
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.62.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.63.0"))
 
     implementation("sk.ainet.transformers:skainet-transformers-runtime-kllama")
     implementation("sk.ainet.transformers:skainet-transformers-agent")
@@ -41,7 +41,7 @@ Or in Maven (Maven needs the `-jvm` classifier suffix on platform artifacts):
     <dependency>
       <groupId>sk.ainet.transformers</groupId>
       <artifactId>skainet-transformers-bom</artifactId>
-      <version>0.62.2</version>
+      <version>0.63.0</version>
       <type>pom</type>
       <scope>import</scope>
     </dependency>

diff --git a/docs/modules/ROOT/pages/tutorials/getting-started-leaf.adoc b/docs/modules/ROOT/pages/tutorials/getting-started-leaf.adoc
@@ -34,7 +34,7 @@ encoder output to the advertised dimensionality. The runtime applies it automati
 [source,kotlin]
 ----
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.62.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.63.0"))
     implementation("sk.ainet.transformers:skainet-transformers-providers")
 }
 ----

diff --git a/docs/modules/ROOT/pages/tutorials/llama3-tool-calling.adoc b/docs/modules/ROOT/pages/tutorials/llama3-tool-calling.adoc
@@ -52,7 +52,7 @@ The pieces you need live in three modules:
 [source,kotlin]
 ----
 dependencies {
-    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.62.2"))
+    implementation(platform("sk.ainet.transformers:skainet-transformers-bom:0.63.0"))
 
     implementation("sk.ainet.transformers:skainet-transformers-runtime-kllama")
     implementation("sk.ainet.transformers:skainet-transformers-agent")

diff --git a/gradle.properties b/gradle.properties
@@ -1,5 +1,5 @@
 GROUP=sk.ainet.transformers
-VERSION_NAME=0.62.2
+VERSION_NAME=0.63.0
 
 POM_DESCRIPTION=SKaiNET-transformers
 

diff --git a/gradle/libs.versions.toml b/gradle/libs.versions.toml
@@ -1,5 +1,5 @@
 [versions]
-skainet = "0.61.0"
+skainet = "0.62.0"
 agp = "9.4.1"
 jacksonDatabind = "2.22.3"
 jsonSchemaValidator = "3.0.8"

diff --git a/llm-inference/llama/src/jvmTest/kotlin/sk/ainet/models/llama/ForwardScopeSteadyStateTest.kt b/llm-inference/llama/src/jvmTest/kotlin/sk/ainet/models/llama/ForwardScopeSteadyStateTest.kt
@@ -19,9 +19,11 @@ import sk.ainet.lang.types.FP32
  * #343 steady-state pins for the per-step [sk.ainet.lang.memory.ForwardScope]
  * in [OptimizedLLMRuntime]'s DIRECT decode path:
  *
- * 1. **Bit-identity** — a scoped decode produces exactly the logits of an
+ * 1. **Identity to within a few ulp** — a scoped decode produces the logits of an
  *    unscoped one (`forwardSlabFloats = 0`), step for step. The slab changes
- *    where activations live, never what they hold.
+ *    where activations live, never what they hold; since engine 0.62.0 the JVM
+ *    kernels chosen for slab-backed and heap activations may accumulate in a
+ *    different order, so the pin allows four ulp instead of raw-bit equality.
  * 2. **Flat peak** — `peakFloats` stops growing once decode is warm. The KV
  *    history lives outside the scope (the cache detaches it to ambient
  *    storage), so per-step slab use grows only with the KV *views* concat
@@ -105,11 +107,16 @@ class ForwardScopeSteadyStateTest {
             val b = unscoped.forward(tokenId).data.copyToFloatArray()
             assertEquals(b.size, a.size)
             for (i in a.indices) {
-                // Bit-identical, not approximately equal: same kernels, same
-                // values, different storage.
-                assertEquals(
-                    b[i].toRawBits(), a[i].toRawBits(),
-                    "step=$step logit[$i]: scoped=${a[i]} unscoped=${b[i]}",
+                // Equal to within a few ulp. Up to engine 0.61 this was raw-bit identity (same
+                // kernels, same values, different storage); since engine 0.62.0 the JVM ops route
+                // slab-backed and heap activations through kernels whose accumulation order can
+                // differ (the fused native attention and norm kernels take heap operands, a
+                // segment-backed activation is bridged first), which moves a logit by one ulp.
+                // The slab still changes where activations live, never what they mean.
+                val tolerance = 4 * Math.ulp(b[i])
+                assertTrue(
+                    a[i] == b[i] || kotlin.math.abs(a[i] - b[i]) <= tolerance,
+                    "step=$step logit[$i]: scoped=${a[i]} unscoped=${b[i]} (tolerance $tolerance)",
                 )
             }