Add portable GGUF runtimes and expanded benchmarks

Adds the BF16 universal Node artifact, Q8 accelerated artifact, platform canary evidence, typed-decisions evaluation, option-order robustness, question scaling, and refreshed manifest.

.gitattributes

@@ -40,3 +40,5 @@ assets/xmesh-metallic.png filter=lfs diff=lfs merge=lfs -text
 assets/xmesh-metallic-banner.png filter=lfs diff=lfs merge=lfs -text
 assets/xmesh-wordmark-dark.png filter=lfs diff=lfs merge=lfs -text
 assets/xmesh-wordmark-light.png filter=lfs diff=lfs merge=lfs -text
+Singularity-xMESH-0.8B-BF16.gguf filter=lfs diff=lfs merge=lfs -text
+Singularity-xMESH-0.8B-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text

GGUF_VALIDATION.json

@@ -0,0 +1,87 @@
+{
+  "schema_version": 1,
+  "source": {
+    "format": "merged BF16 safetensors",
+    "model_sha256": "2dda751389cf7d98cfed72a16e6d8ce253dcd125b84870713cd33d1637e0ab84",
+    "decision_head_sha256": "ff6433e096130f286038a1bfff07741018d6c8f6ccb91ee325419de717e87ff4"
+  },
+  "conversion": {
+    "llama_cpp_tag": "b11719",
+    "llama_cpp_commit": "5aeac37a35b44fde4bdffaec9159f2fcb253b622",
+    "bf16_gguf_sha256": "9f4c08f19c4d203059bc74ed2754e800ec472d7276d8ddb2321480e86d392fff"
+  },
+  "release_artifact": {
+    "path": "Singularity-xMESH-0.8B-BF16.gguf",
+    "format": "GGUF BF16",
+    "bytes": 1571529600,
+    "sha256": "9f4c08f19c4d203059bc74ed2754e800ec472d7276d8ddb2321480e86d392fff",
+    "context_tokens": 2048,
+    "physical_batch_tokens": 2048,
+    "node_default_slots": 1
+  },
+  "accelerated_artifact": {
+    "path": "Singularity-xMESH-0.8B-Q8_0.gguf",
+    "format": "GGUF Q8_0",
+    "bytes": 841069440,
+    "sha256": "41b36270b67f6c595e98cf4f29d6a0d13f0c19daa37b75741cd8296e2729fd75",
+    "cpu_status": "rejected for universal use: llama.cpp b11719 exits 139 during quantized Clef CPU initialization"
+  },
+  "q8_metal_frozen_x402_export_parity": {
+    "cases": 80,
+    "decisions": 960,
+    "exact_cases": 80,
+    "matching_decisions": 960,
+    "median_request_ms": 1542.8355415,
+    "p95_request_ms": 4466.246708,
+    "runtime": "llama.cpp b11719, Apple M4 Metal, two validation slots"
+  },
+  "bf16_cpu_frozen_x402_export_parity": {
+    "cases": 80,
+    "decisions": 960,
+    "exact_cases": 80,
+    "matching_decisions": 960,
+    "median_request_ms": 3109.9937505,
+    "p95_request_ms": 3495.010333,
+    "runtime": "llama.cpp b11719, Apple M4 CPU, one release-profile slot"
+  },
+  "rejected_artifacts": [
+    {
+      "format": "GGUF Q4_K_M",
+      "bytes": 545162112,
+      "sha256": "0f13d08daa156d8f9cec6aa9a2488861ac0a0e1749a7ba04cfccec9662c1c302",
+      "reason": "Changed 16 of 960 decisions and preserved only 65 of 80 complete flows."
+    }
+  ],
+  "typed_decisions": {
+    "artifact": "Singularity-xMESH-0.8B-Q8_0.gguf",
+    "dataset": "LocalLLaMA/typed-decisions",
+    "revision": "7189f46b02248bd2569df7325c9df8e3b24e90a7",
+    "split": "all/test",
+    "cases": 400,
+    "decisions": 2000,
+    "accuracy": 0.4775,
+    "row_exact_match": 0.045,
+    "soft_accuracy": 0.4151943743,
+    "brier_vs_soft": 0.40858675,
+    "ece": 0.2928255711,
+    "score_mae": 0.6424282911,
+    "score_within_one": 0.765,
+    "median_case_ms": 1201.44,
+    "p95_case_ms": 1353.346917
+  },
+  "option_order": {
+    "seed": 13,
+    "choice_questions": 600,
+    "answer_flips": 0,
+    "mean_probability_total_variation": 0.0,
+    "max_probability_total_variation": 0.0
+  },
+  "question_scaling": {
+    "runtime": "llama.cpp b11719, Apple M4 Metal, sequential caller",
+    "one_question_p50_ms": 167.6057495,
+    "five_questions_p50_ms": 448.471604,
+    "ten_questions_p50_ms": 846.7270625,
+    "twenty_questions": "unsupported: 2070 input tokens exceeds the 2048-token contract",
+    "fifty_questions": "unsupported: 5070 input tokens exceeds the 2048-token contract"
+  }
+}

MANIFEST.json

@@ -2,8 +2,8 @@
   "schema_version": 1,
   "algorithm": "sha256",
   "base_model": "Qwen/Qwen3.5-0.8B@2fc06364715b967f1860aea9cf38778875588b17",
-  "file_count": 41,
-  "total_bytes": 1833747287,
+  "file_count": 46,
+  "total_bytes": 4246356274,
   "files": [
     {
       "path": ".gitignore",
@@ -20,6 +20,11 @@
       "bytes": 476,
       "sha256": "9a3fe9e8e038b38588a7907e99ce57904b1d92f095b70b4077fb41f62da55983"
     },
+    {
+      "path": "GGUF_VALIDATION.json",
+      "bytes": 3064,
+      "sha256": "65941fa861c7bf676982739d2dac50c00cf5c80d14ad8e61ea8afd41eed462db"
+    },
     {
       "path": "LICENSE",
       "bytes": 11544,
@@ -45,10 +50,30 @@
       "bytes": 498,
       "sha256": "b0e5e3849f78da8fac2df8aebe756b68e219dcdd1aa53697a4c1c618a2f25eb9"
     },
+    {
+      "path": "PLATFORM_CANARIES.json",
+      "bytes": 1548,
+      "sha256": "b5649859f1fadc6b3f34a1aa427b4132eb05e1fb8ded374df83e514b46e26e45"
+    },
+    {
+      "path": "PORTABILITY.md",
+      "bytes": 2437,
+      "sha256": "dadf4fe24a159bdd9e228556aaa1efca9bbe2c10fe71dde22776a4fdfdcb0e59"
+    },
     {
       "path": "README.md",
-      "bytes": 4582,
-      "sha256": "a1292ba01f9ea3008d2a0206620d006c0f7dcd4d23146cd42aa4b1a5ccccd765"
+      "bytes": 7480,
+      "sha256": "db512aafdc3af7af03152261f3f5d9083e49639b44008c5cd9ab4bcb1323b99c"
+    },
+    {
+      "path": "Singularity-xMESH-0.8B-BF16.gguf",
+      "bytes": 1571529600,
+      "sha256": "9f4c08f19c4d203059bc74ed2754e800ec472d7276d8ddb2321480e86d392fff"
+    },
+    {
+      "path": "Singularity-xMESH-0.8B-Q8_0.gguf",
+      "bytes": 841069440,
+      "sha256": "41b36270b67f6c595e98cf4f29d6a0d13f0c19daa37b75741cd8296e2729fd75"
     },
     {
       "path": "THIRD_PARTY_NOTICES.md",

PLATFORM_CANARIES.json

@@ -0,0 +1,43 @@
+{
+  "schema_version": 1,
+  "model": "Singularity xMESH 0.8B",
+  "artifact": "Singularity-xMESH-0.8B-BF16.gguf",
+  "artifact_sha256": "9f4c08f19c4d203059bc74ed2754e800ec472d7276d8ddb2321480e86d392fff",
+  "llama_cpp": {
+    "tag": "b11719",
+    "commit": "5aeac37a35b44fde4bdffaec9159f2fcb253b622"
+  },
+  "profiles": {
+    "macos_arm64_cpu": {
+      "status": "passed",
+      "cases": 80,
+      "matching_decisions": 960,
+      "total_decisions": 960,
+      "median_request_ms": 3109.9937505,
+      "p95_request_ms": 3495.010333,
+      "slots": 1
+    },
+    "macos_arm64_metal": {
+      "status": "passed_load_and_inference",
+      "note": "Q8 passed 960/960 export parity; BF16 Metal had one numerically tied decision flip, while BF16 CPU matches the release reference."
+    },
+    "linux_arm64_cpu": {
+      "status": "passed",
+      "environment": "Docker Desktop aarch64, Ubuntu 24.04, libgomp1",
+      "archive": "llama-b11719-bin-ubuntu-arm64.tar.gz",
+      "archive_sha256": "290aa4432ad8e28b638285be6a75d0dd6f327d9b2d95f7048a8cf5e69704bee9",
+      "server_sha256": "7b0734c19cc38e09d7c9b4ed96cbdbb54f9318282cecd7e4a82f067b54451892",
+      "fixture_input_tokens": 1012,
+      "expected_decisions": 9,
+      "matching_decisions": 9,
+      "response_sha256": "b1d70e2b94d1157f78b840d7ba1372e5719ca34f5231920fc65dfe27dd50bf20",
+      "dependency": "libgomp.so.1 from Ubuntu package libgomp1"
+    },
+    "linux_x64_cpu": {
+      "status": "pending native canary"
+    },
+    "windows_x64_cpu": {
+      "status": "pending native canary"
+    }
+  }
+}

PORTABILITY.md

@@ -0,0 +1,42 @@
+# xMESH portability contract
+
+Singularity Node must serve xMESH through the same `/v1/systemone` request and response contract on every supported platform.
+
+## Runtime
+
+- Minimum llama.cpp: `b11719` (`5aeac37a35b44fde4bdffaec9159f2fcb253b622`).
+- Context: 2,048 tokens per request.
+- Physical batch: 2,048 tokens. The llama.cpp default of 512 rejects normal multi-field requests.
+- Default concurrency: one slot. Higher concurrency requires a platform-specific canary.
+
+## Artifacts
+
+| Artifact | Bytes | SHA-256 | Use |
+| --- | ---: | --- | --- |
+| `Singularity-xMESH-0.8B-BF16.gguf` | 1,571,529,600 | `9f4c08f19c4d203059bc74ed2754e800ec472d7276d8ddb2321480e86d392fff` | Universal Node artifact |
+| `Singularity-xMESH-0.8B-Q8_0.gguf` | 841,069,440 | `41b36270b67f6c595e98cf4f29d6a0d13f0c19daa37b75741cd8296e2729fd75` | Accelerated runtimes only |
+
+Q8 passes the full Metal decision-parity suite but crashes in llama.cpp `b11719` during CPU initialization. Q4_K_M loads but changes 16 of 960 frozen decisions, so it is not a release artifact.
+
+## Release matrix
+
+| Surface | Target | Artifact | Status |
+| --- | --- | --- | --- |
+| Desktop | macOS Apple silicon | BF16 | Local CPU and Metal canaries passed; packaged-app canary pending |
+| Desktop | Windows x64 | BF16 | Build and runtime canary pending |
+| Node CLI | Linux x64 CPU/Vulkan | BF16 | Provisioner pinned; native runtime canary pending |
+| Node CLI | Linux ARM64 CPU/Vulkan | BF16 | CPU load and real decision canary passed in Ubuntu 24.04 ARM64 |
+
+The desktop release workflow does not currently publish Linux or Intel Mac packages. The Node CLI covers Linux; Intel Mac desktop support requires a separate release target.
+
+The official Linux llama.cpp archive links `libgomp.so.1`. Minimal Ubuntu images must install the standard `libgomp1` package before `sgl setup`; normal desktop/server distributions often already include it. The provisioner must surface this exact prerequisite when the staged binary fails its launch check.
+
+## Quality gates
+
+- Frozen x402 export parity: 80 cases and 960 decisions.
+- Official typed-decisions boundary suite: 400 cases and 2,000 decisions.
+- Option-order robustness: 600 choice decisions.
+- Runtime integrity: exact artifact and llama.cpp archive hashes.
+- Platform canary: load, health, real `/v1/systemone` request, expected answer, and clean shutdown.
+
+See `GGUF_VALIDATION.json` for measurements and rejected artifact evidence.

README.md

@@ -12,9 +12,13 @@ tags:
   - x402
   - tool-selection
   - safetensors
+  - gguf
+  - llama.cpp
   - pytorch
   - cpu
   - apple-silicon
+  - windows
+  - linux
   - mps
   - cuda
 language:
@@ -51,11 +55,19 @@ The general decision warm-up used licensed rows from [`MaziyarPanahi/AgentToolDe
 | --- | --- |
 | Apple silicon MPS | **Verified** on an M4 with 16 GB unified memory |
 | CPU | **Verified**; portable but slower |
+| llama.cpp Metal | **Verified** with the Q8 GGUF on an Apple M4 |
+| llama.cpp Windows/Linux CPU | Runtime packaged; release canaries required |
 | NVIDIA CUDA | Supported by the PyTorch runtime; standalone CUDA canary pending |
 | AMD ROCm | Not verified |
 
 The checkpoint contains 873M BF16 parameters. An 8 GB system is the practical minimum; 16 GB is recommended for comfortable local loading and inference.
 
+## Portable GGUF runtime
+
+`Singularity-xMESH-0.8B-BF16.gguf` is the universal Singularity Node artifact. It is 1,571,529,600 bytes with SHA-256 `9f4c08f19c4d203059bc74ed2754e800ec472d7276d8ddb2321480e86d392fff` and requires llama.cpp `b11719` or newer for native Clef `/v1/systemone` support.
+
+`Singularity-xMESH-0.8B-Q8_0.gguf` is an optional 841,069,440-byte accelerated artifact with SHA-256 `41b36270b67f6c595e98cf4f29d6a0d13f0c19daa37b75741cd8296e2729fd75`. It preserved all 960 decisions across the frozen 80-case Metal export-parity run, but llama.cpp `b11719` crashes while initializing this quantized Clef architecture on CPU. Singularity Node therefore uses BF16 for its cross-platform contract. Q4_K_M was rejected because it changed 16 decisions and preserved only 65 of 80 complete flows.
+
 ## Frozen x402 regression evaluation
 
 The standalone export was fixed before one evaluation on the existing frozen 80-case x402 suite. No weights, prompts, or policy rules were changed after seeing these results.
@@ -71,6 +83,25 @@ The standalone export was fixed before one evaluation on the existing frozen 80-
 
 Jev is a hosted API and Laya used a different local runtime, so latency is not a hardware-equivalent comparison. The standalone CPU validation measured 6.45 seconds median per full multi-field flow on an Apple M4. Optimized serving runtimes can be substantially faster.
 
+## General typed-decision evaluation
+
+This is a boundary test, not the model's specialization. xMESH was evaluated zero-shot on `LocalLLaMA/typed-decisions` revision `7189f46b02248bd2569df7325c9df8e3b24e90a7`, complete `all/test` split: 400 cases and 2,000 decisions.
+
+| Model | Condition | Accuracy | Soft accuracy | Brier ↓ | ECE ↓ | Score MAE ↓ |
+| --- | --- | ---: | ---: | ---: | ---: | ---: |
+| xMESH 0.8B | zero-shot Q8 | **0.478** | 0.415 | 0.409 | 0.293 | 0.642 |
+| Laya | base checkpoint, published | 0.362 | 0.331 | **0.315** | **0.174** | 0.694 |
+| Laya typed-decisions | trained on this benchmark's train split, published | **0.766** | 0.471 | **0.061** | 0.213 | **0.242** |
+| Jev 1.13 | published | 0.727 | **0.580** | 0.148 | **0.144** | 0.391 |
+
+xMESH beats base Laya on this suite, but it does not beat Jev or the benchmark-specific Laya fine-tune. Accuracy by xMESH workflow was 0.266 agent-trace observability, 0.578 customer service, 0.492 invoice processing, and 0.574 security incidents.
+
+A seeded option-order permutation changed **0 of 600** choice answers and produced zero probability drift because the Clef runtime canonicalizes option identifiers before inference.
+
+On a short fixed state with identical three-option questions, Q8 llama.cpp on an Apple M4 measured 168 ms p50 for one question, 448 ms for five, and 847 ms for ten. Twenty and fifty questions exceeded this release's validated 2,048-token context; they are not advertised as supported.
+
+Laya and Jev comparison rows above are figures published by the Laya benchmark page. They were not rerun in the xMESH environment.
+
 ## Run on Apple silicon
 
 Validated with Python 3.12 on an Apple M4 with 16 GB unified memory.
@@ -84,6 +115,19 @@ python examples/inference.py --request examples/request.json --device mps
 
 Use `--device cpu` when MPS is unavailable. The request must contain `state` and `questions`; `examples/request.json` shows the typed contract.
 
+## Run with llama.cpp
+
+```bash
+llama-server \
+  -m Singularity-xMESH-0.8B-BF16.gguf \
+  --host 127.0.0.1 --port 8081 \
+  -c 2048 -b 2048 -ub 2048 --parallel 1
+
+curl http://127.0.0.1:8081/v1/systemone \
+  -H 'Content-Type: application/json' \
+  --data-binary @examples/request.json
+```
+
 ## Safety boundary
 
 xMESH is an advisory router. It must not receive wallet secrets, construct signed payments, bypass approval, invent endpoints, or execute arbitrary URLs. Runtime code must restrict choices to retrieved candidate IDs and enforce budgets, networks, schemas, and approval rules independently.

Singularity-xMESH-0.8B-BF16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f4c08f19c4d203059bc74ed2754e800ec472d7276d8ddb2321480e86d392fff
+size 1571529600

Singularity-xMESH-0.8B-Q8_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41b36270b67f6c595e98cf4f29d6a0d13f0c19daa37b75741cd8296e2729fd75
+size 841069440