docs: record Open-Jev PR-by-PR optimization A/B results

[hsliuustc0106]

Purpose

Add a concise Open-Jev optimization post led by headline numbers and comparison figures, linked from the README and MkDocs navigation. The post is about 850 words, with four main charts, a compact mechanism table, and links to detailed figures, controls and raw records.

Cover native/graph ancestry (#19/#52), PR #55's backend gain and isolated RMSNorm/SiLU/graph-cache A/Bs, GDN preparation (#68), and processing/runtime regression checks (#78/#80). Distinguish bundled, isolated, kernel and HTTP results; keep missing ablations and missed gates visible.

Include the reusable write-optimization-blog skill, A/B evidence template and presentation references. The skill favors short explanations and linked detailed records for concise posts. Installed and repository copies match. Preserve six SVG/PNG pairs, CPU-only regeneration and 78 measured passes / 5,040 timing samples with provenance/parity hashes.

Test Plan

System1-Omni Version / Commit: docs head 5d5398147be15ac2e73409f0de694b5481d8e395; GitHub-reported PR base f594d7dfc4c2bef812e23f7ed73573be9625b287; fetched upstream main 99e368982ea3a9d4ede2f96133c1f1be2995008a; branch merge base d6aa3a9cedcc5c87cf0222d08fabce5accf521db. Measurement revisions are separate in the ledger.

Check the diff and claims against frozen source/evidence; audit hashes, raw timings, parity/gates, rendered links and figures; validate both skills and strict MkDocs build. New inference A/B is unnecessary for documentation changes.

Test Result

• mkdocs build --strict --site-dir profile/blog-writing-20261005/site: passed.
• Task-local evidence/link audit: passed. Ten source-hash entries, 12 raw GDN kernel records, 296 GDN HTTP responses and 78 exported passes / 5,040 samples match the ledger. Refactor p95/throughput/gates and headline ratios recompute; Markdown/rendered links resolve.
• Both skill copies pass quick_validate.py and match byte-for-byte.
• All six figure pairs and measurement inputs are unchanged in this revision; four figures are embedded and two detailed figures linked. Regeneration and visual inspection passed earlier on this branch.
• git diff --check origin/main...HEAD: passed. Commit is DCO signed.
• GitHub CI on 5d53981: Docs build, Rust and benchmarks passed; Pages deployment skipped for pull requests.
• No new GPU measurements or local Rust/GPU tests: documentation-only changes. Full archives/task-local helpers remain outside the PR.

Demo / evidence

Read the post, PR #55 figure, ledger, timing export, and writing skill.

Historical H200/BF16 JevBench results: 362.21 → 48.50 ms (7.47×) versus raw HF, 74 single-candidate requests/pass, concurrency 1, two warm measured passes after excluded feasibility. Separate RMSNorm/SiLU/Graph-64 A/Bs save 2.58% / 2.48% / 2.03%; they cannot be added into a cumulative ladder. GDN's ~13% longer-call gain yields 0.42% HTTP, missing its 2% gate. Refactor checks establish no speedup.

HF/native decisions match on the subset while probabilities differ. LoRA/native/gate-only attribution and GDN + Graph latency remain unmeasured. Figures plot preserved project evidence; timing exports contain IDs/timings/hashes without request bodies or credentials. Full response/protocol/trace archives remain local. Video is N/A for documentation. These are historical results, not measurements of this docs head or current main.

Self-review

Agent-assisted diff and evidence review found no blocking issues. Contributor checklist state is preserved; the PR remains draft.

• I have reviewed the full diff and addressed the issues I found.
• I have checked that the change follows the project's architecture and stays focused on the stated purpose.
• I have run the checks appropriate to this change and reported commands, results, and anything I could not verify above.
• I have checked that the PR description, documentation, and any accuracy or performance claims match the implementation and available evidence.