DaoyuanLi2816 · Aug 14, 2026 · Aug 14, 2026
diff --git a/.github/ISSUE_TEMPLATE/benchmark_submission.yml b/.github/ISSUE_TEMPLATE/benchmark_submission.yml
@@ -8,18 +8,18 @@ body:
         Thank you. Measured numbers from hardware we do not own are the most
         useful contribution this project can receive.
 
-        Produce a submission file first:
+        Produce and validate a private-by-default hardware record first:
 
         ```bash
-        miniverl train <recipe>
-        miniverl export-benchmark runs/<run-id> --out my-result.json
+        miniverl hardware record --run runs/<run-id> --out hardware-record.json
+        miniverl hardware validate hardware-record.json
         ```
 
-        `export-benchmark` strips absolute paths and keeps only the GPU model,
-        VRAM, OS family and library versions. Validate it against
-        `benchmarks/schema/benchmark-result.schema.json`, compute its SHA-256,
-        and read it before posting. Never include tokens, private paths or
-        proprietary data.
+        The record binds profile, plan, models, quantization, token bounds,
+        batches, updates, memory, timing and retained artifact hashes. The
+        command never uploads. Read the JSON before posting and never include
+        tokens, private paths or proprietary data. Schema-valid community
+        records remain unreviewed until a maintainer validates their evidence.
 
   - type: input
     id: hardware
@@ -40,7 +40,7 @@ body:
   - type: textarea
     id: result
     attributes:
-      label: Contents of the exported result JSON
+      label: Contents of the validated hardware-record JSON
       render: json
     validations:
       required: true
@@ -58,7 +58,7 @@ body:
   - type: input
     id: artifact_hash
     attributes:
-      label: Exported JSON SHA-256
+      label: Hardware-record JSON SHA-256
       description: Hash the exact JSON attached to this submission.
       placeholder: "64 lowercase hexadecimal characters"
     validations:
@@ -81,9 +81,11 @@ body:
           required: true
         - label: I ran the recipe to completion and did not select the best of several runs without saying so.
           required: true
-        - label: I validated the JSON against the committed schema and recorded the exact artifact SHA-256.
+        - label: I ran `miniverl hardware validate` and recorded the exact JSON SHA-256.
           required: true
         - label: I reviewed the export and attachments for tokens, private paths and proprietary data.
           required: true
+        - label: I consent to publishing this sanitized record and its reported hardware/software metadata.
+          required: true
         - label: I am willing to open a pull request adding this file under `benchmarks/results/`.
           required: false
diff --git a/.github/workflows/docs.yml b/.github/workflows/docs.yml
@@ -34,6 +34,7 @@ jobs:
           python scripts/publish_alignment_lab_artifacts.py --check
           python scripts/publish_verl_bridge_diagrams.py --check
           python scripts/publish_verl_opd_compatibility.py --check
+          python scripts/publish_hardware_record_schema.py --check
       - name: Build complete development site strictly
         run: mkdocs build --strict
       - name: Install deterministic Chromium

diff --git a/PROJECT_STATE.md b/PROJECT_STATE.md
@@ -35,6 +35,8 @@ Development `0.10.0.dev0` has a closed typed profile registry and torch-free
 compatibility introspection. Profile-scoped plans, caches, checkpoints and
 exports bind an independent identity. The direct-GKD and sampled-k1 vanilla
 policy-loss profiles both have pinned conformance and measured RTX 4080 paths.
+Portable hardware records preserve measured, estimated and unknown states;
+schema-valid community submissions remain unreviewed until maintainer validation.
 
 Arbitrary verl YAML, other policy-gradient modes, rewards, PPO/GRPO, Ray, FSDP,
 Megatron, multi-GPU and distributed execution remain unsupported. The legacy

diff --git a/PYPI.md b/PYPI.md
@@ -190,6 +190,16 @@ updates. See [hardware planning](https://github.com/DaoyuanLi2816/mini-verl/blob
 automatic downgrade to a different model,
 teacher, context, top-k or loss when memory is tight.
 
+Share a sanitized result without uploading it automatically:
+
+```bash
+miniverl hardware record --run runs/my-opd --out hardware-record.json
+miniverl hardware validate hardware-record.json
+```
+
+Community records remain unreviewed until their hashes and provenance pass the
+[documented maintainer gate](https://github.com/DaoyuanLi2816/mini-verl/blob/main/docs/community-benchmarks.md).
+
 ## Data and artifact interoperability
 
 The profile consumes structured verl-style Parquet without substituting a toy

diff --git a/README.md b/README.md
@@ -190,6 +190,16 @@ updates. See [hardware planning](docs/hardware-planning.md). There is no
 automatic downgrade to a different model,
 teacher, context, top-k or loss when memory is tight.
 
+Share a sanitized result without uploading it automatically:
+
+```bash
+miniverl hardware record --run runs/my-opd --out hardware-record.json
+miniverl hardware validate hardware-record.json
+```
+
+Community records remain unreviewed until their hashes and provenance pass the
+[documented maintainer gate](docs/community-benchmarks.md).
+
 ## Data and artifact interoperability
 
 The profile consumes structured verl-style Parquet without substituting a toy

diff --git a/README.zh-CN.md b/README.zh-CN.md
@@ -161,6 +161,16 @@ BF16/FP16 自动选择取决于设备能力，而不是 3070、4080、5090 或 T
 [硬件规划](docs/hardware-planning.md)。
 显存紧张时不会静默更换模型、teacher、context、top-k 或 loss。
 
+已完成的 run 可以生成经过清理、且不会自动上传的硬件记录：
+
+```bash
+miniverl hardware record --run artifacts/runs/<run-id> --out hardware-record.json
+miniverl hardware validate hardware-record.json
+```
+
+社区提交即使通过 schema 校验也仍标记为 `unreviewed`；只有维护者复核并确认发布许可后，
+才会进入公开的“maintainer-validated measured”矩阵。
+
 ## 数据与产物互操作
 
 profile 直接读取结构化 verl 风格 Parquet，不会在缺失数据时静默换成 calculator 环境。

diff --git a/docs/community-benchmarks.md b/docs/community-benchmarks.md
@@ -27,3 +27,31 @@ To propose a real result, add the JSON plus any new versioned recipe record in
 one pull request. Describe unsupported hardware or software explicitly; a
 negative or failed run is useful evidence and should not be removed merely
 because it does not improve a benchmark.
+
+## OPD hardware records
+
+Completed OPD runs use a separate, stricter version-1 record:
+
+```bash
+miniverl hardware record --run runs/my-opd --out hardware-record.json
+miniverl hardware validate hardware-record.json
+```
+
+The record binds the profile identity, plan and config digests, GPU/VRAM and
+software stack, pinned actor/teacher identities, quantization, prompt/response
+bounds, teacher-target mode, logical and physical batches, update count,
+measured memory and phase timing, resume state, and artifact hashes. Every
+numeric evidence item is labelled `measured`, `estimated`, or `unknown`.
+Generation and validation are torch-free and perform no upload. The committed
+[JSON Schema](generated/hardware-record-v1.schema.json) is generated from the
+same Pydantic contract used by the CLI.
+
+| Record origin | Public status | Required gate |
+| --- | --- | --- |
+| `maintainer_measured` | measured only after repository review | retained run, hashes, schema/privacy validation and maintainer approval |
+| `community_submitted` | unreviewed candidate | validated JSON plus explicit publication consent |
+| `not_measured` | documented gap | no numeric value may be presented as measured |
+
+`hardware validate` proves schema and privacy conformance only. It does not
+change `review_status`, publish a file or add a row to the public matrix. An
+issue comment or unreviewed JSON can never become a measured row automatically.