Add model card and release metadata

ARTIFACT.json

@@ -0,0 +1,43 @@
+{
+  "schema_version": 1,
+  "artifact": {
+    "filename": "Qwen3.6-27B-Q8_0_ROCMFPX.gguf",
+    "bytes": 27755345216,
+    "sha256": "ff4dbc9093c1df6fd1242294d15eb94c7bfe42ed67f98e9d58b9789ac6912c1b",
+    "quantization": "Q8_0_ROCMFPX",
+    "bits_per_weight": 8.25,
+    "importance_matrix": false,
+    "quantized_by": "1337Hero",
+    "quantization_command": [
+      "./build-r9700/bin/llama-quantize",
+      "Qwen3.6-27B-BF16.gguf",
+      "Qwen3.6-27B-Q8_0_ROCMFPX.gguf",
+      "Q8_0_ROCMFPX",
+      "16"
+    ]
+  },
+  "base_model": {
+    "repo_id": "Qwen/Qwen3.6-27B",
+    "license": "apache-2.0",
+    "source_gguf_sha256": "0438be1f5bc861ffa84e1d2d4036920f6f3d9759f3cdedbc40e554a321d1c9c5"
+  },
+  "runtime": {
+    "repository": "https://github.com/1337hero/ROCmFPX",
+    "base_commit": "6bf20cd688ba0af882d1f68ba50b292edf646ab4",
+    "implementation_commits": [
+      "eb38c6f67701ff9c74e8597f573eedf9ccecf774",
+      "45bcff509c4b1cff137e2cc1ea84671c61ceddea"
+    ],
+    "llama_server_sha256": "1471753a94ba007b094842474cc3b3ffd48106f15a7c71934899dd832ae4cbb3",
+    "llama_bench_sha256": "38bc30e53badc5ae5efb5e3d449d989a4672c52fe5801eb1ce74e70233283749",
+    "llama_quantize_sha256": "0aa545af25e8235613349987fb29323745b70980aa2a4fb45c169cdf561181c6"
+  },
+  "validated_platform": {
+    "gpu": "AMD Radeon AI PRO R9700",
+    "architecture": "gfx1201",
+    "rocm": "7.2.4",
+    "host_os": "Arch Linux",
+    "power_profile": "250 W, -75 mV"
+  },
+  "claims": "claims.yaml"
+}

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Alibaba Cloud
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

NOTICE

@@ -0,0 +1,13 @@
+Qwen3.6-27B Q8_0_ROCMFPX GGUF
+
+This artifact is derived from Qwen/Qwen3.6-27B.
+Copyright 2026 Alibaba Cloud.
+
+Modification notice:
+The original model was converted to GGUF and quantized to the experimental
+Q8_0_ROCMFPX representation in July 2026 by the 1337Hero community account.
+The derivative requires a matching experimental ROCmFPX/llama.cpp runtime.
+
+The Apache License, Version 2.0, governs redistribution. Qwen and related
+marks belong to their owners. This community artifact is not affiliated with
+or endorsed by Qwen, Alibaba Cloud, AMD, ROCmFPX, or llama.cpp.

README.md

@@ -0,0 +1,245 @@
+---
+license: apache-2.0
+license_link: https://huggingface.co/Qwen/Qwen3.6-27B/blob/main/LICENSE
+base_model: Qwen/Qwen3.6-27B
+base_model_relation: quantized
+pipeline_tag: text-generation
+quantized_by: 1337Hero
+tags:
+  - gguf
+  - qwen3.6
+  - quantized
+  - rocm
+  - amd
+  - rdna4
+  - gfx1201
+  - experimental
+---
+
+# Qwen3.6-27B — Q8_0_ROCMFPX GGUF (experimental, AMD gfx1201)
+
+An experimental Q8 quantization of
+[Qwen/Qwen3.6-27B](https://huggingface.co/Qwen/Qwen3.6-27B), tuned and
+validated for the AMD Radeon AI PRO R9700 (`gfx1201`).
+
+> [!IMPORTANT]
+> This file does **not** run on upstream llama.cpp, Ollama, LM Studio, or
+> vLLM. It uses the custom `Q8_0_ROCMFPX` tensor type and requires the pinned
+> ROCmFPX fork build described below. Unsupported runtimes should reject the
+> file; if a tool appears to load it anyway, do not trust the output. Hugging
+> Face's GGUF metadata viewer may also mislabel the custom tensor type or fail
+> to parse the file.
+
+## Should you use this?
+
+Use this model if all of the following are true:
+
+- you run a Radeon AI PRO R9700 (`gfx1201`) under ROCm;
+- you are willing to build the pinned fork from source;
+- you want a Q8 that is 2.94% smaller than upstream `Q8_0` and performs equal
+  observed model work about 6% faster on this hardware.
+
+Otherwise, use an ordinary `Q8_0` GGUF of Qwen3.6-27B. It has broad support in
+the standard GGUF runtime ecosystem, and its quality is equivalent within the
+tolerances measured here.
+
+## TL;DR — measured against upstream `Q8_0`
+
+| Measure | Result |
+| --- | ---: |
+| Model size | 2.94% smaller |
+| HumanEval | 140/164 — tied |
+| MBPP base / MBPP+ | 348/378 and 291/378 vs 350/378 and 293/378 — within the predeclared tolerance |
+| Proxy — full-model decode | 5.77–5.87% faster |
+| Proxy — `pp4096+tg512` combined | 4.75–4.96% faster |
+| Equal-work — agent-derived evaluation (18 matched pairs) | 5.82% faster median, 17/18 pairs, exact one-sided `p=0.000965` |
+| End-to-end — raw live-agent wall time | 2.21% faster — **failed** the predeclared 3% threshold |
+
+The supported conclusion is deliberately narrow: this format performs equal
+observed model work faster than upstream `Q8_0` on the measured deployment.
+The campaign did **not** confirm that complete live-agent jobs finish at
+least 3% faster, so no such claim is made. All thresholds were fixed before
+the runs; failed gates are reported, not reinterpreted.
+
+The public [claim registry](claims.yaml) records the scope, result class, and
+sealed aggregate-evidence hash behind each release claim.
+
+## What `Q8_0_ROCMFPX` actually is
+
+Each block stores 32 signed eight-bit codes and one finite UE4M3 scale byte:
+33 bytes per 32 weights, or 8.25 bits per weight. The HIP path decodes the
+scale and performs integer MMVQ/MMQ dot products with float accumulation.
+This is **not** native FP8 or FP4 matrix arithmetic.
+
+```text
+GGUF block
+  32 signed int8 codes + UE4M3 scale
+        |
+        v
+activation quantization to Q8_1
+        |
+        +--> decode:  gfx1201 MMVQ, VDR8 + measured wave policy
+        |
+        `--> prefill: integer MMQ path
+        |
+        v
+float accumulation / model output
+```
+
+The experimental format and execution path come from the upstream
+[ROCmFPX](https://github.com/charlie12345/ROCmFPX) project. This repository's
+contribution is the `gfx1201` decode tuning (VDR8 vector-dot width and a
+measured wave policy) and the validation below; other GPU targets retain
+ROCmFPX defaults and are unmeasured.
+
+The quantization is uniform Q8 with no importance matrix and no per-tensor
+routing.
+
+## Required runtime
+
+Build the pinned fork branch:
+
+```bash
+git clone https://github.com/1337hero/ROCmFPX.git
+cd ROCmFPX
+git checkout 45bcff509c4b1cff137e2cc1ea84671c61ceddea
+env JOBS=16 scripts/build-r9700.sh
+```
+
+The wrapper is the simplest supported build. The sealed performance runners
+used this HIP-only configuration:
+
+```bash
+cmake -S . -B build-r9700 -G Ninja \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DGGML_HIP=ON \
+  -DGGML_HIP_FORCE_MMQ=ON \
+  -DGGML_VULKAN=OFF \
+  -DGGML_CUDA=OFF \
+  -DCMAKE_HIP_ARCHITECTURES=gfx1201 \
+  -DGPU_TARGETS=gfx1201 \
+  -DCMAKE_HIP_FLAGS= \
+  -DLLAMA_BUILD_SERVER=ON \
+  -DLLAMA_BUILD_WEBUI=OFF \
+  -DLLAMA_USE_PREBUILT_WEBUI=OFF \
+  -DLLAMA_BUILD_TESTS=ON \
+  -DGGML_BUILD_TESTS=OFF
+cmake --build build-r9700 -j 16 --target \
+  llama-server llama-bench llama-quantize
+```
+
+The validated implementation is ROCmFPX base commit
+`6bf20cd688ba0af882d1f68ba50b292edf646ab4` plus commits
+`eb38c6f67701ff9c74e8597f573eedf9ccecf774` and
+`45bcff509c4b1cff137e2cc1ea84671c61ceddea`.
+
+Binaries validated by the lab:
+
+| Binary | SHA-256 |
+| --- | --- |
+| `llama-server` | `1471753a94ba007b094842474cc3b3ffd48106f15a7c71934899dd832ae4cbb3` |
+| `llama-bench` | `38bc30e53badc5ae5efb5e3d449d989a4672c52fe5801eb1ce74e70233283749` |
+| `llama-quantize` | `0aa545af25e8235613349987fb29323745b70980aa2a4fb45c169cdf561181c6` |
+
+## Example deployment
+
+This shape matches the validated two-card, 262,144-token envelope. Device
+names depend on the host; check `--list-devices` first.
+
+```bash
+./build-r9700/bin/llama-server \
+  -m Qwen3.6-27B-Q8_0_ROCMFPX.gguf \
+  --no-mmap \
+  -c 262144 \
+  -b 2048 \
+  -ub 512 \
+  -t 16 \
+  -ngl 99 \
+  -sm layer \
+  -ts 1,1 \
+  -dev ROCm0,ROCm2 \
+  --fit off \
+  -ctk f16 \
+  -ctv f16 \
+  -fa on \
+  -np 1
+```
+
+Both this model and the upstream `Q8_0` control allocated one 262,144-token
+slot across two R9700s with F16 K/V cache and generated bounded valid output;
+this model used 804,319,232 fewer resident bytes in the captured snapshots.
+That test proved allocation and bounded generation, not a timed full-context
+prompt.
+
+## Measurement conditions
+
+All performance numbers were measured on:
+
+- Qwen3.6-27B, this GGUF vs an upstream `Q8_0` control of the same
+  conversion, same host, same workload;
+- two non-display Radeon AI PRO R9700s at a pre-existing 250 W, −75 mV tune —
+  **not** stock 300 W;
+- ROCm 7.2.4 on Arch Linux, outside the official Radeon Linux support matrix;
+- pinned source revisions, forced model residency, reversed candidate order,
+  and predeclared promotion thresholds.
+
+The equal-work result comes from 18 common-seed matched agent-workload pairs
+with ABBA/BAAB counterbalancing, warmup before every measured run, full
+request accounting (1,899 logical requests, zero retries), and exact paired
+Wilcoxon signed-rank tests.
+
+## Artifact
+
+| Field | Value |
+| --- | --- |
+| File | `Qwen3.6-27B-Q8_0_ROCMFPX.gguf` |
+| Size | 27,755,345,216 bytes |
+| SHA-256 | `ff4dbc9093c1df6fd1242294d15eb94c7bfe42ed67f98e9d58b9789ac6912c1b` |
+| Source GGUF | BF16, SHA-256 `0438be1f5bc861ffa84e1d2d4036920f6f3d9759f3cdedbc40e554a321d1c9c5` |
+| Quantization | uniform `Q8_0_ROCMFPX`, no importance matrix |
+| Content | main text model only; no MTP artifact or multimodal projector |
+
+Quantization command (the qualifying run wrote to a temporary filename and
+then promoted the verified hash without overwriting an existing artifact):
+
+```bash
+./build-r9700/bin/llama-quantize \
+  Qwen3.6-27B-BF16.gguf \
+  Qwen3.6-27B-Q8_0_ROCMFPX.gguf \
+  Q8_0_ROCMFPX \
+  16
+```
+
+Verify after download:
+
+```bash
+sha256sum -c SHA256SUMS
+```
+
+## Limitations
+
+- Requires an experimental fork; upstream llama.cpp compatibility awaits a
+  separate code contribution.
+- Validated on exactly one model conversion, one Arch Linux host, ROCm 7.2.4,
+  and `gfx1201` GPUs at a tuned power profile. Other GPUs, hosts, and stock
+  power are unmeasured.
+- Aggregate quality stays within the declared tolerances, but this model and
+  upstream `Q8_0` do not produce identical per-task outcomes.
+- The raw private lab archive (prompts, outputs, paths, environment) is not
+  published; it awaits a separate sanitization review.
+
+## License and attribution
+
+- **Base model:** Qwen3.6-27B, Copyright 2026 Alibaba Cloud, Apache-2.0. This
+  repository redistributes a converted and quantized derivative under the same
+  license. See `LICENSE` and `NOTICE`.
+- **Format and execution path:** the experimental `Q8_0_ROCMFPX`
+  representation and HIP kernels are the work of the
+  [ROCmFPX](https://github.com/charlie12345/ROCmFPX) project, which builds on
+  [llama.cpp](https://github.com/ggml-org/llama.cpp).
+- **This repository:** the `gfx1201` decode tuning, quantized artifact, and
+  validation evidence.
+
+Qwen and related marks belong to their owners. This community quantization is
+not affiliated with or endorsed by Qwen, Alibaba Cloud, AMD, ROCmFPX, or
+llama.cpp.

SHA256SUMS

@@ -0,0 +1,6 @@
+34f4a2d2879ae6fb8b90af386c3ae78fd52e9d9e3c3b745d7b0c9241ea1001cf  ARTIFACT.json
+39b34cdcb38146a1eff65fb3b604291c0ea2a136f35334efbf61c8d7fe697772  claims.yaml
+689f0c220c9b4e857f35ca7d0dca51397e5a545dc4ad3987438c023e24ff0ab6  LICENSE
+0e93cce70f17bc22164a0a27596b6bc41073b9f42f0b95edcd03ed08cf4fc545  NOTICE
+985037f872776846bfd039479a7648c3cc6b5e7a99eb2eb531b545cd29e53ef3  README.md
+ff4dbc9093c1df6fd1242294d15eb94c7bfe42ed67f98e9d58b9789ac6912c1b  Qwen3.6-27B-Q8_0_ROCMFPX.gguf

claims.yaml

@@ -0,0 +1,145 @@
+schema_version: 1
+release: Qwen3.6-27B-Q8_0_ROCMFPX-GGUF
+artifact_sha256: ff4dbc9093c1df6fd1242294d15eb94c7bfe42ed67f98e9d58b9789ac6912c1b
+
+evidence_archive:
+  publication_status: private_pending_sanitization
+  gate_manifest_sha256: 4d9953b3698684fcf1f795d2303cdb80ee4e90bd9d8afba9cf8a7b3bab240198
+  note: >-
+    Raw prompts, outputs, telemetry, environment captures, and host paths are
+    intentionally excluded from this model release. The paths below identify
+    records in the sealed research archive; their hashes bind the aggregate
+    statements published here.
+
+claims:
+  - id: C01
+    class: compatibility
+    status: supported
+    claim: >-
+      The artifact requires the pinned ROCmFPX fork and is not compatible with
+      stock llama.cpp, Ollama, LM Studio, or vLLM.
+    scope: custom Q8_0_ROCMFPX GGUF tensor type
+    evidence:
+      runtime_commit: 45bcff509c4b1cff137e2cc1ea84671c61ceddea
+      llama_server_sha256: 1471753a94ba007b094842474cc3b3ffd48106f15a7c71934899dd832ae4cbb3
+
+  - id: C02
+    class: format
+    status: supported
+    claim: >-
+      Each weight block stores 32 signed int8 codes and one finite UE4M3 scale
+      byte (8.25 bits per weight); HIP uses integer MMVQ/MMQ dot products with
+      float accumulation. This is not native FP8 or FP4 matrix arithmetic.
+    scope: Q8_0_ROCMFPX representation and measured HIP execution path
+    evidence:
+      summary: reports/phase2/model-gate/20260713T214155Z/q8-rocmfpx/SUMMARY.md
+      summary_sha256: 0a8f780c98fd40560e254121542382d37bc1c2e69a43000bbe4a87303f7e7eb9
+
+  - id: C03
+    class: artifact_size
+    status: supported
+    claim: >-
+      The custom GGUF is 27,755,345,216 bytes, 840,418,208 bytes or 2.94%
+      smaller than the matched upstream Q8_0 control.
+    scope: the two named Qwen3.6-27B conversions only
+    evidence:
+      summary: reports/phase2/model-gate/20260713T214155Z/q8-rocmfpx/SUMMARY.md
+      summary_sha256: 0a8f780c98fd40560e254121542382d37bc1c2e69a43000bbe4a87303f7e7eb9
+
+  - id: C04
+    class: quality
+    status: supported_aggregate_only
+    claim: >-
+      Custom and upstream Q8_0 each score 140/164 on deterministic HumanEval.
+    scope: aggregate pass@1; per-task outcomes differ
+    evidence:
+      custom_score_sha256: bae8408c545a7738917d143a36b1b725beecb9ad59d8a0e6caf14fb18a7bb5a7
+      summary: reports/phase2/model-gate/20260713T214155Z/q8-rocmfpx/SUMMARY.md
+      summary_sha256: 0a8f780c98fd40560e254121542382d37bc1c2e69a43000bbe4a87303f7e7eb9
+
+  - id: C05
+    class: quality
+    status: pass_within_predeclared_tolerance
+    claim: >-
+      Custom scores 348/378 on MBPP and 291/378 on MBPP+, versus 350/378 and
+      293/378 for upstream Q8_0; both two-task deficits are within the fixed
+      seven-task tolerance.
+    scope: aggregate pass@1; not behavioral identity
+    evidence:
+      summary: reports/phase2/model-gate/20260713T214155Z/q8-rocmfpx/quality/mbppplus/SUMMARY.md
+      summary_sha256: 036add79e0b9f1b0c8e28c7d48ecb3edefcc14b1dc216486ad05ecf7d9c1e965
+
+  - id: C06
+    class: performance_proxy
+    status: pass
+    claim: >-
+      Full-model decode is 5.77% to 5.87% faster than the matched upstream
+      Q8_0 control.
+    scope: >-
+      Two non-display R9700s, ROCm 7.2.4 on Arch Linux, 250 W/-75 mV profile,
+      fixed runner and model hashes
+    evidence:
+      summary: reports/phase2/model-gate/20260713T214155Z/q8-rocmfpx/SUMMARY.md
+      summary_sha256: 0a8f780c98fd40560e254121542382d37bc1c2e69a43000bbe4a87303f7e7eb9
+
+  - id: C07
+    class: performance_proxy
+    status: pass
+    claim: >-
+      The pp4096+tg512 combined proxy is 4.75% to 4.96% faster than the
+      matched upstream Q8_0 control.
+    scope: same measured deployment as C06; not a live-agent wall-time claim
+    evidence:
+      summary: reports/phase2/model-gate/20260713T214155Z/q8-rocmfpx/SUMMARY.md
+      summary_sha256: 0a8f780c98fd40560e254121542382d37bc1c2e69a43000bbe4a87303f7e7eb9
+
+  - id: C08
+    class: equal_work_performance
+    status: pass
+    claim: >-
+      In 18 predeclared matched pairs, standardized evaluation of equal
+      within-pair prompt and completion work is 5.82% faster at the median,
+      favors custom in 17/18 pairs, and has exact one-sided p=0.000965.
+    scope: fixed six-scenario agent-derived pack and measured deployment
+    evidence:
+      protocol_sha256: 6b1ea93a91c288700086f593ac69ad39bb65d71eb8eebdd6e225936ef54d023f
+      analysis_sha256: 775ffe638a0df718e7c8d72eacb8471b4fe03b7495cc670c211cd46ace8e9ce3
+      summary_sha256: 6c86e5a2b19688436eb68ae80e221926bf0bde94aefea611ef6433193feb12c4
+      nested_manifest_sha256: 18adf876740d265182b97cbc9f95a4a4e14626d68e59843813c11343e6adc156
+
+  - id: C09
+    class: end_to_end_performance
+    status: rejected_below_predeclared_magnitude
+    claim: >-
+      Raw live-agent wall time improves 2.21% at the median and favors custom
+      in 13/18 pairs, but fails the fixed 3% minimum; this release does not
+      claim that complete live-agent jobs finish faster.
+    scope: same V4 campaign as C08
+    evidence:
+      analysis_sha256: 775ffe638a0df718e7c8d72eacb8471b4fe03b7495cc670c211cd46ace8e9ce3
+      summary_sha256: 6c86e5a2b19688436eb68ae80e221926bf0bde94aefea611ef6433193feb12c4
+
+  - id: C10
+    class: deployment_capacity
+    status: pass
+    claim: >-
+      Both models allocate one 262,144-token slot and generate bounded output
+      on two R9700s; custom uses 804,319,232 fewer resident bytes in the
+      captured snapshots.
+    scope: >-
+      Capacity allocation and bounded generation only; no timed 262,144-token
+      prompt and no peak-memory claim
+    evidence:
+      summary: reports/phase2/model-gate/20260713T214155Z/q8-rocmfpx/deployment/262144-two-card/SUMMARY.md
+      summary_sha256: 36c1b025dbb67baf858f1b56c76799ed5e25e65f97b4ea82fa659fdaebfd1b66
+
+  - id: C11
+    class: hardware_scope
+    status: measured_single_host
+    claim: >-
+      Results are validated only on Radeon AI PRO R9700 gfx1201 GPUs under
+      ROCm 7.2.4 on Arch Linux at a 250 W/-75 mV tune.
+    scope: other GPUs, hosts, ROCm versions, and stock power are unmeasured
+    evidence:
+      summary: reports/phase2/model-gate/20260713T214155Z/q8-rocmfpx/SUMMARY.md
+      summary_sha256: 0a8f780c98fd40560e254121542382d37bc1c2e69a43000bbe4a87303f7e7eb9