Support DGX Spark metadata and benchmark contributor credit

.github/pull_request_template.md

@@ -0,0 +1,22 @@
+## Benchmark submission
+
+- Models, hardware, runtime, precision:
+- Harness and LibreYOLO versions/commits:
+- Dataset and number of images:
+- Validation and generated-data rebuild:
+- Known measurement or metadata limitations:
+
+## Benchmark authorship and display credit
+
+- I ran these benchmarks myself: yes / no (if no, identify the runner with their permission).
+- Benchmark author display: chosen name / GitHub handle / Anonymous contributor.
+- Optional public links chosen by the author: GitHub, LinkedIn, X, website.
+- I confirm the named author agreed to this credit and these links: yes / not yet.
+
+Maintainers record confirmed credit in `submission-credits.json` after the PR
+has a URL. Submission authorship is separate from who measured the benchmark.
+Anonymous display hides the name and links on the website; the public source PR
+and Git history remain accessible. Omit personal details you do not want public.
+
+Measurements are contributor-reported. Acceptance checks format and protocol
+consistency; it does not certify independent reproduction of the results.

.github/workflows/benchmark-submissions.yml

@@ -4,6 +4,7 @@ on:
   pull_request:
     paths:
       - "submissions/**"
+      - "submission-credits.json"
       - "schemas/**"
       - "support-matrix.json"
       - "scripts/**"
@@ -13,6 +14,7 @@ on:
       - main
     paths:
       - "submissions/**"
+      - "submission-credits.json"
       - "schemas/**"
       - "support-matrix.json"
       - "scripts/**"

.github/workflows/website.yml

@@ -7,6 +7,7 @@ on:
       - "article-pipeline/v2/**"
       - "generated/verified-results.v1.json"
       - "submissions/**"
+      - "submission-credits.json"
       - "scripts/**"
       - ".github/workflows/website.yml"
   push:
@@ -17,6 +18,7 @@ on:
       - "article-pipeline/v2/**"
       - "generated/verified-results.v1.json"
       - "submissions/**"
+      - "submission-credits.json"
       - "scripts/**"
       - ".github/workflows/website.yml"
 

docs/benchmark-credit.md

@@ -0,0 +1,65 @@
+# Benchmark authorship and provenance
+
+Each run retains its original harness JSON. Publication metadata lives in
+`submission-credits.json`, keyed by the exact submission filename. Rebuild
+`generated/verified-results.v1.json` with the existing script after changes.
+
+- `source_pr`: the Vision Analysis PR supplying the run.
+- `submitted_by`: the public GitHub account submitting that PR, checked by the reviewer.
+- `author`: optional, contributor-confirmed identity of the person who ran the
+  benchmark. Do not infer authorship from submission, a Git commit, or a social profile.
+- `author.display_name`: the name or pseudonym requested by the benchmark author.
+- `author.links`: optional HTTPS `github`, `linkedin`, `x`, and `website` URLs
+  supplied and approved by that author. No scraped photos or inferred profiles.
+- `author.anonymous: true`: display Anonymous contributor and suppress author
+  profiles and the submitter handle in the result table. Do not include a name
+  or author links with this option. The source PR is still public, so this is
+  anonymous presentation, not a guarantee of anonymity.
+
+Example, after the author confirms the desired credit in their PR:
+
+```json
+{
+  "exact-harness-filename.json": {
+    "source_pr": "https://github.com/LibreYOLO/vision-analysis/pull/14",
+    "submitted_by": "example",
+    "author": {
+      "display_name": "Example contributor",
+      "links": {"github": "https://github.com/example"}
+    }
+  }
+}
+```
+
+Older rows without confirmed credit remain unattributed. The source JSON link
+remains available. Do not label them anonymous or project-authored by default.
+The first Spark batch records its verified public submitter and source PR;
+personal display credit is pending the contributor's choice.
+
+The reviewer checks the source PR and the author's confirmation, then runs:
+
+```sh
+python3 scripts/validate_submission.py
+python3 scripts/build_verified_results.py
+npm test
+```
+
+Validation checks structure, supported configurations and credit links. It
+does not independently reproduce a run or transfer the site's responsibility
+for accepting and presenting data. Preserve the original run, version pins,
+dataset fingerprint and source PR so errors can be investigated and corrected.
+
+# DGX Spark metadata
+
+The hardware catalogue describes installed capacity: 128 GB shared by CPU and
+GPU, per NVIDIA's hardware guide. It is not 128 GB of dedicated VRAM. The harness
+reports OS-visible memory in its existing GiB convention (`ram_gb` and
+`unified_memory_gb`), which can be lower. `gpu_memory_gb: null` means there is no
+dedicated GPU capacity reading. Older GB10 runs with zero in that field retain
+their original bytes; use the hardware catalogue's unified-memory label.
+
+Do not overwrite measured peak allocation with installed capacity. Do not edit
+past measurements to make the metadata look complete or require a timing rerun
+solely to replace unavailable hardware metadata.
+
+Source: https://docs.nvidia.com/dgx/dgx-spark/hardware.html