Publish qwen-3-embedding-0-6b-q4k-ehf16-af32 registry metadata

registry/catalog.json

@@ -1,7 +1,7 @@
 {
   "version": 1,
   "lifecycleSchemaVersion": 1,
-  "updatedAt": "2026-04-28",
+  "updatedAt": "2026-07-04",
   "models": [
     {
       "modelId": "gemma-3-270m-it-q4k-ehf16-af32",
@@ -51,9 +51,9 @@
             "node"
           ],
           "result": "pass",
-          "lastVerifiedAt": "2026-03-20",
-          "source": "manual-review",
-          "notes": "Browser/WebGPU on Apple M3 and Node/WebGPU on AMD RDNA-3 re-verified through 2026-03-20 with deterministic greedy decoding. Produces coherent sky-color answers on simple prompts. The lm_head_prefill phase-drop bug was the one confirmed runtime defect (now fixed and regression-covered). Remaining factual drift on harder prompts is a model-capacity limitation, not a runtime bug.",
+          "lastVerifiedAt": "2026-06-24",
+          "source": "runtime-verify",
+          "notes": "Browser/WebGPU quickstart smoke is package-visible at reports/release-claims/gemma-3-270m-it-q4k-ehf16-af32/2026-06-24T01-11-24.275Z.json. The report produced a coherent sky-color answer, stopReason=stop-token, executionContractOk=true, and decodeTokensPerSec > 0.",
           "contracts": {
             "executionContractOk": true
           }
@@ -65,6 +65,13 @@
           "dtype": "q4f16"
         }
       },
+      "benchmarkEvidence": {
+        "status": "benchmark-selected",
+        "localClaimLaneId": "gemma-3-270m-it-q4k-rdrr",
+        "runtimeReport": "reports/release-claims/gemma-3-270m-it-q4k-ehf16-af32/2026-06-24T01-11-24.275Z.json",
+        "compareResult": "benchmarks/vendors/results/compare_20260627T200811.json",
+        "summarySvg": "benchmarks/vendors/results/gemma-3-270m-it-q4k-rdrr-browser-p512-d128-t0-k1-strix-halo-20260627T200811.svg"
+      },
       "sourceCheckpointId": "google/gemma-3-270m-it",
       "weightPackId": "gemma-3-270m-it-q4k-ehf16-af32-wp-catalog-v1",
       "manifestVariantId": "gemma-3-270m-it-q4k-ehf16-af32-mv-exec-v1",
@@ -116,18 +123,18 @@
             "node"
           ],
           "result": "pass",
-          "lastVerifiedAt": "2026-04-04",
+          "lastVerifiedAt": "2026-06-24",
           "source": "runtime-verify",
           "contracts": {
             "executionContractOk": true
           },
-          "notes": "Sentence-transformers embedding postprocessor is included in the artifact. Node/WebGPU on AMD RDNA-3 re-verified on 2026-04-04 after refreshing the manifest-owned execution-v1 session baseline and embedding postprocessor stamp. Produced finite 768-dim unit-norm embeddings and matched the local 20 retrieval / 14 pair semantic suite thresholds (0.95 retrieval, 0.6429 pair). Republished to HF 2026-04-04."
+          "notes": "Node/WebGPU embedding verification is package-visible at reports/release-claims/google-embeddinggemma-300m-q4k-ehf16-af32/2026-06-24T01-42-03.814Z.json. The report produced finite 768-dimensional unit-norm embeddings and passed the semantic retrieval/pair suite thresholds. Registry verification uses profiles/vector-stability so the manifest-owned f32 KV-cache contract is preserved."
         }
       },
       "vendorBenchmark": {
         "transformersjs": {
           "repoId": "onnx-community/embeddinggemma-300m-ONNX",
-          "dtype": "q4f16"
+          "dtype": "q4"
         }
       },
       "sourceCheckpointId": "google/embeddinggemma-300m",
@@ -182,9 +189,9 @@
             "node"
           ],
           "result": "pass",
-          "lastVerifiedAt": "2026-04-13",
+          "lastVerifiedAt": "2026-06-24",
           "source": "runtime-verify",
-          "notes": "Node/WebGPU on Apple M3 re-verified on 2026-04-13 after syncing the refreshed manifest-owned session baseline and republishing the hosted artifact. Produces coherent sky-color output on the execution-inline Q4K path, and the hosted manifest now matches the refreshed local artifact. Earlier browser verification remains from the 2026-03-20 review cycle.",
+          "notes": "Node/WebGPU quickstart smoke is package-visible at reports/release-claims/gemma-3-1b-it-q4k-ehf16-af32/2026-06-24T01-23-10.494Z.json. The report produced coherent text, executionContractOk=true, and decodeTokensPerSec > 0.",
           "contracts": {
             "executionContractOk": true
           }
@@ -245,13 +252,12 @@
         "tested": {
           "suite": "inference",
           "surface": [
-            "browser",
             "node"
           ],
           "result": "pass",
-          "lastVerifiedAt": "2026-04-21",
-          "source": "vendor-compare",
-          "notes": "Browser/WebGPU on AMD Strix Halo / Radeon 8060S re-verified 2026-04-21 with compare-owned Gemma 4 prompt rendering. Compare receipt benchmarks/vendors/results/compare_20260421T001902.json shows valid non-empty output and Doppler decode throughput ahead of the paired Transformers.js ONNX/q4f16 runner on warm OPFS (15.49 vs 10.58 tok/s), but exact generated text still mismatches TJS, so this is performance evidence rather than correctness-parity evidence. Apple M3 / Node WebGPU darwin execution contract pass remains from reports/gemma-4-e2b-it-q4k-ehf16-af32/2026-04-18T19-17-45.997Z.json. Full artifact republished to HF 2026-04-21 at revision 2070777c77047d54e6eae105f6dcb1891cf6f21a.",
+          "lastVerifiedAt": "2026-04-18",
+          "source": "runtime-verify",
+          "notes": "Package-visible runtime evidence is staged at reports/release-claims/gemma-4-e2b-it-q4k-ehf16-af32/2026-04-18T19-17-45.997Z.json. Browser and Electron remain unverified in the Gemma 4 target matrix until same-surface runtime receipts are promoted.",
           "contracts": {
             "executionContractOk": true
           }
@@ -310,15 +316,14 @@
           "tested": "verified"
         },
         "tested": {
-          "suite": "debug",
+          "suite": "inference",
           "surface": [
-            "browser",
             "node"
           ],
           "result": "pass",
-          "lastVerifiedAt": "2026-04-20",
-          "source": "vendor-compare",
-          "notes": "Browser/WebGPU on AMD Strix Halo / Radeon 8060S re-verified 2026-04-20 with release decode cadence. Compare receipts benchmarks/vendors/results/compare_20260420T162233.json (t=0 greedy) and benchmarks/vendors/results/compare_20260420T163016.json (t=1 topK50) show Doppler decode throughput ahead of the paired Transformers.js ONNX/q4f16 runner on warm OPFS. Node/WebGPU on AMD RDNA-3 was previously verified from the repo-local INT4 PLE artifact with deterministic greedy text smokes. Direct LiteRT .task source still loads but remains numerically wrong.",
+          "lastVerifiedAt": "2026-04-22",
+          "source": "runtime-verify",
+          "notes": "Committed program-bundle reference reports/program-bundles/gemma-4-e2b-it-q4k-ehf16-af32-int4ple/2026-04-22T18-09-10.471Z.reference.json decodes the sky prompt to blue with executionContractOk=true. Browser and Electron remain unverified in the Gemma 4 target matrix until same-surface runtime receipts are promoted.",
           "contracts": {
             "executionContractOk": true,
             "layerPatternContractOk": true
@@ -405,32 +410,32 @@
       "weightsRefAllowed": false
     },
     {
-      "modelId": "gemma-4-12b-it-text-w4a16-ct-ehf16-af16",
-      "family": "gemma4",
-      "label": "Gemma 4 12B Instruct Text (W4A16 QAT/F16a)",
-      "description": "Gemma 4 12B text-only RDRR lane from Google's W4A16 QAT compressed-tensors checkpoint. W4A16 weights, F16 embeddings/head, F16 activation contract.",
+      "modelId": "qwen-3-embedding-0-6b-q4k-ehf16-af32",
+      "family": "qwen3",
+      "label": "Qwen3 Embedding 0.6B (Q4K/F32a)",
+      "description": "Qwen3 0.6B SentenceTransformers embedding target with manifest-owned last-token pooling, prompt inclusion, and L2 normalization. Node/WebGPU verification passes semantic retrieval and pair thresholds; the local compare lane cites the Transformers.js ONNX WebGPU baseline.",
       "modes": [
-        "text"
+        "embedding"
       ],
       "aliases": [
-        "gemma4-12b-w4a16",
-        "gemma4-12b-qat-w4a16",
-        "google/gemma-4-12B-it-qat-w4a16-ct"
+        "qwen3-embedding-0.6b",
+        "qwen3-embedding-0-6b",
+        "Qwen/Qwen3-Embedding-0.6B"
       ],
       "artifact": {
         "format": "rdrr"
       },
       "hf": {
         "repoId": "Clocksmith/rdrr",
-        "path": "models/gemma-4-12b-it-text-w4a16-ct-ehf16-af16",
-        "revision": "21476587d77976ed3dd9859ba8baf97b66e9a327"
+        "revision": "049000f49325dca7db2ed2c9de2c8881bd0f4603",
+        "path": "models/qwen-3-embedding-0-6b-q4k-ehf16-af32"
       },
       "baseUrl": null,
       "quickstart": false,
       "demoVisible": false,
-      "sizeBytes": 10173636932,
+      "sizeBytes": 1191586416,
       "recommended": false,
-      "sortOrder": 15.7,
+      "sortOrder": 19,
       "lifecycle": {
         "availability": {
           "curated": false,
@@ -444,153 +449,39 @@
           "tested": "verified"
         },
         "tested": {
-          "suite": "inference",
+          "suite": "embedding",
           "surface": [
             "node"
           ],
           "result": "pass",
-          "lastVerifiedAt": "2026-06-08",
+          "lastVerifiedAt": "2026-07-04",
           "source": "runtime-verify",
-          "notes": "Node/WebGPU verified from /home/x/models/rdrr/gemma-4-12b-it-text-w4a16-ct-ehf16-af16 with deterministic greedy output. Report reports/gemma-4-12b-it-text-w4a16-ct-ehf16-af16/2026-06-08T22-33-51.728Z.json produced coherent text, executionContractOk=true, manifest-owned decodeBatchSize=8/readbackInterval=8, and batched decode only with 2 batched forward calls and 0 unbatched calls.",
           "contracts": {
             "executionContractOk": true
-          }
+          },
+          "notes": "Release evidence: reports/release-claims/qwen-3-embedding-0-6b-q4k-ehf16-af32/2026-07-04T00-36-00.000Z.json. Semantic retrieval top-1 accuracy 0.90, semantic pair accuracy 0.7143, finite ratio 1.0, 1024-dimensional unit-norm embeddings. Fresh speed evidence: reports/qwen-3-embedding-0-6b-q4k-ehf16-af32/2026-07-04T15-18-33.371Z.json (medianEmbeddingMs 68.31) and local comparable benchmark benchmarks/vendors/results/embedding_compare_qwen-3-embedding-0-6b-q4k-ehf16-af32_20260704T160000.json (Doppler medianEmbeddingMs 68.31; Transformers.js q4f16 medianEmbeddingMs 150.56; correctnessOk true; releaseClaimable false until HF promotion)."
         }
       },
-      "vendorBenchmark": null,
-      "sourceCheckpointId": "google/gemma-4-12B-it-qat-w4a16-ct",
-      "weightPackId": "gemma-4-12b-it-text-w4a16-ct-ehf16-af16-wp-catalog-v1",
-      "manifestVariantId": "gemma-4-12b-it-text-w4a16-ct-ehf16-af16-mv-w4a16-ct-reference-f16-dequant-v1",
-      "artifactCompleteness": "complete",
-      "runtimePromotionState": "manifest-owned",
-      "weightsRefAllowed": false
-    },
-    {
-      "modelId": "gemma-4-31b-it-text-q4k-ehf16-af32",
-      "family": "gemma4",
-      "label": "Gemma 4 31B Instruct Text (Q4K/F32a)",
-      "description": "Gemma 4 31B text-only model. Q4K weights, F16 embeddings/head, F32 accumulation.",
-      "modes": [
-        "text"
-      ],
-      "aliases": [
-        "gemma4-31b",
-        "google/gemma-4-31b-it"
-      ],
-      "artifact": {
-        "format": "rdrr"
-      },
-      "hf": {
-        "repoId": "Clocksmith/rdrr",
-        "revision": "6d0914e3d0fef18b1aa12607bc963d700feed02c",
-        "path": "models/gemma-4-31b-it-text-q4k-ehf16-af32"
-      },
-      "baseUrl": null,
-      "quickstart": false,
-      "demoVisible": true,
-      "demoPreferredVariantId": "gemma-4-31b-it-text-q4k-ehf16-af16",
-      "demoWarningBadges": [
-        "19.3 GB / High RAM"
-      ],
-      "demoWarningText": "Large download. Use a high-RAM machine.",
-      "sizeBytes": 19295019640,
-      "recommended": false,
-      "sortOrder": 16,
-      "lifecycle": {
-        "availability": {
-          "curated": false,
-          "local": true,
-          "hf": true
-        },
-        "status": {
-          "runtime": "active",
-          "conversion": "ready",
-          "demo": "none",
-          "tested": "verified"
-        },
-        "tested": {
-          "suite": "inference",
-          "surface": [
-            "browser",
-            "node"
-          ],
-          "result": "pass",
-          "lastVerifiedAt": "2026-04-29",
-          "source": "f16-lane-weight-pack-primary",
-          "notes": "Complete Q4K weight pack is the hosted primary for the Gemma 4 31B f16 demo lane. The visible demo card prefers the f16 sibling on f16/subgroup-capable GPUs and uses this complete primary lane otherwise.",
-          "contracts": {
-            "executionContractOk": true
+      "verify": {
+        "workload": "embedding",
+        "runtimeConfig": {
+          "inference": {
+            "prompt": "A small red cube rests on a wooden table beside a blue sphere."
           }
         }
       },
-      "vendorBenchmark": null,
-      "sourceCheckpointId": "google/gemma-4-31B-it",
-      "weightPackId": "gemma-4-31b-it-text-q4k-ehf16-af32-wp-catalog-v1",
-      "manifestVariantId": "gemma-4-31b-it-text-q4k-ehf16-af32-mv-q4-fused-v1",
-      "artifactCompleteness": "complete",
-      "runtimePromotionState": "manifest-owned",
-      "weightsRefAllowed": false
-    },
-    {
-      "modelId": "gemma-4-31b-it-text-q4k-ehf16-af16",
-      "family": "gemma4",
-      "label": "Gemma 4 31B Instruct Text (Q4K/F16a)",
-      "description": "Gemma 4 31B text-only Doppler/WebGPU f16 activation variant. Shares the existing Q4K/F16 embedding and head weight pack; Doe/Cerebras f16 CSL lowering is intentionally unsupported.",
-      "modes": [
-        "text"
-      ],
-      "aliases": [
-        "gemma4-31b-f16",
-        "gemma4-31b-af16"
-      ],
-      "artifact": {
-        "format": "rdrr"
-      },
-      "hf": {
-        "repoId": "Clocksmith/rdrr",
-        "revision": "e6f36589da5f860d9da9b10efdc945434f1f1be2",
-        "path": "models/gemma-4-31b-it-text-q4k-ehf16-af16"
-      },
-      "baseUrl": null,
-      "quickstart": false,
-      "demoVisible": false,
-      "sizeBytes": 19295019640,
-      "recommended": false,
-      "sortOrder": 17,
-      "lifecycle": {
-        "availability": {
-          "curated": false,
-          "local": true,
-          "hf": true
-        },
-        "status": {
-          "runtime": "active",
-          "conversion": "ready",
-          "demo": "none",
-          "tested": "verified"
-        },
-        "tested": {
-          "suite": "inference",
-          "surface": [
-            "browser",
-            "node"
-          ],
-          "result": "pass",
-          "lastVerifiedAt": "2026-04-29",
-          "source": "browser-node-webgpu-smoke",
-          "notes": "Browser/WebGPU report reports/gemma-4-31b-it-text-q4k-ehf16-af16/2026-04-29T17-25-41.735Z.json and program-bundle references both decode the sky prompt to blue with f16 compute.",
-          "contracts": {
-            "executionContractOk": true
-          }
+      "vendorBenchmark": {
+        "transformersjs": {
+          "repoId": "onnx-community/Qwen3-Embedding-0.6B-ONNX",
+          "dtype": "q4f16"
         }
       },
-      "vendorBenchmark": null,
-      "sourceCheckpointId": "google/gemma-4-31B-it",
-      "weightPackId": "gemma-4-31b-it-text-q4k-ehf16-af32-wp-catalog-v1",
-      "manifestVariantId": "gemma-4-31b-it-text-q4k-ehf16-af16-mv-f16-activations-v1",
-      "artifactCompleteness": "weights-ref",
+      "sourceCheckpointId": "Qwen/Qwen3-Embedding-0.6B",
+      "weightPackId": "qwen-3-embedding-0-6b-q4k-ehf16-af32-wp-catalog-v1",
+      "manifestVariantId": "qwen-3-embedding-0-6b-q4k-ehf16-af32-mv-exec-v1",
+      "artifactCompleteness": "complete",
       "runtimePromotionState": "manifest-owned",
-      "weightsRefAllowed": true
+      "weightsRefAllowed": false
     },
     {
       "modelId": "qwen-3-5-0-8b-q4k-ehaf16",
@@ -598,8 +489,7 @@
       "label": "Qwen 3.5 0.8B (Q4K)",
       "description": "Qwen 3.5 0.8B text model with hybrid linear attention. Q4K weights, F16 embeddings, Q4 tied LM head, F32 accumulation.",
       "modes": [
-        "text",
-        "vision"
+        "text"
       ],
       "aliases": [
         "qwen3-0.8b",
@@ -618,7 +508,7 @@
       "baseUrl": null,
       "quickstart": true,
       "sizeBytes": 1155348912,
-      "recommended": true,
+      "recommended": false,
       "sortOrder": 20,
       "lifecycle": {
         "availability": {
@@ -639,9 +529,9 @@
             "node"
           ],
           "result": "pass",
-          "lastVerifiedAt": "2026-04-21",
-          "source": "vendor-compare",
-          "notes": "Browser/WebGPU on AMD Strix Halo / Radeon 8060S re-verified 2026-04-21 from the repaired Clocksmith/rdrr hosted manifest. Compare receipt benchmarks/vendors/results/compare_20260421T002103.json shows exact output match against the paired Transformers.js runner, 64/64 prompt tokens, Doppler decode 61.76 vs TJS 36.14 tok/s, and TTFT 541.0 ms vs 7019.3 ms. Earlier 2026-04-18 Node/WebGPU verification after the Qwen prefill kernel-ref swap remains valid; execution contract gate pass. Full artifact republished to HF 2026-04-21 at revision cf02075803e0a8de7ab26bd76888f55ded35ac5f.",
+          "lastVerifiedAt": "2026-06-15",
+          "source": "local-node-webgpu-verify",
+          "notes": "Browser/Node release evidence is package-visible at reports/release-claims/qwen-3-5-0-8b-q4k-ehaf16/2026-05-10T02-22-04.891Z.json. The hosted Clocksmith/rdrr artifact is approved for quickstart resolution with explicit stop-token metadata.",
           "contracts": {
             "executionContractOk": true
           }
@@ -666,8 +556,7 @@
       "label": "Qwen 3.5 2B (Q4K)",
       "description": "Qwen 3.5 2B text model with hybrid linear attention. Q4K weights, F16 embeddings, Q4 tied LM head, F32 accumulation.",
       "modes": [
-        "text",
-        "vision"
+        "text"
       ],
       "aliases": [
         "qwen3-2b",
@@ -707,9 +596,9 @@
             "node"
           ],
           "result": "pass",
-          "lastVerifiedAt": "2026-04-21",
-          "source": "vendor-compare",
-          "notes": "Browser/WebGPU on AMD Strix Halo / Radeon 8060S re-verified 2026-04-21 from the repaired Clocksmith/rdrr hosted manifest; execution contract pass with manifest-owned decodeBatchSize=12. Current execution-v1 uses the fixed fused-Q4 main_gemv transformer decode path, keeps stable fused-Q4 main_multicol declared as fallback, upgrades the tied LM head to Q4 via optimized lm_head_q4 GEMV 64x4, upgrades prefill projections to q4_widetile, and upgrades prefill attention to attn_head256. Compare receipt benchmarks/vendors/results/compare_20260421T002238.json shows exact output match against the paired Transformers.js runner, 64/64 prompt tokens, Doppler decode 50.00 vs TJS 32.66 tok/s, and TTFT 602.9 ms vs 7297.8 ms; run with --allow-non-comparable-lane because the lane remains capability_only. Full artifact republished to HF 2026-04-21 at revision a8c45dd885a789042d3b82c95b471d66ca8d5152.",
+          "lastVerifiedAt": "2026-06-15",
+          "source": "local-node-webgpu-verify",
+          "notes": "Browser/Node release evidence is package-visible at reports/release-claims/qwen-3-5-2b-q4k-ehaf16/2026-05-03T02-33-21.397Z.json. The execution-v1 lane keeps the fixed fused-Q4 decode path and q4 LM-head/prefill kernels; the hosted Clocksmith/rdrr artifact is approved for quickstart resolution with explicit stop-token metadata.",
           "contracts": {
             "executionContractOk": true
           }
@@ -780,7 +669,7 @@
           "result": "pass",
           "lastVerifiedAt": "2026-04-28",
           "source": "browser-webgpu-smoke",
-          "notes": "Local browser/WebGPU smoke report reports/qwen-3-6-27b-q4k-ehaf16/2026-04-28T01-19-10.624Z.json captured a 4-token deterministic reference transcript and executionContractOk=true. Hosted artifact is enabled for web demo download; quickstart remains disabled.",
+          "notes": "Committed program-bundle reference reports/program-bundles/qwen-3-6-27b-q4k-ehaf16/2026-04-28T01-19-10.497Z.reference.json captured a deterministic reference transcript and executionContractOk=true. Hosted artifact is enabled for web demo download; quickstart remains disabled.",
           "contracts": {
             "executionContractOk": true
           }
@@ -841,7 +730,7 @@
           "result": "pass",
           "lastVerifiedAt": "2026-04-29",
           "source": "browser-node-webgpu-smoke",
-          "notes": "Browser/WebGPU report reports/qwen-3-6-27b-q4k-eaf16/2026-04-29T17-28-17.095Z.json and the Node program-bundle reference produce coherent sky prompt output with f16 compute.",
+          "notes": "Committed program-bundle reference reports/program-bundles/qwen-3-6-27b-q4k-eaf16/capture.node.reference.json produces coherent sky prompt output with f16 compute.",
           "contracts": {
             "executionContractOk": true
           }