將高風險產品目標轉成可否證、benchmark-gated R&D。適用架構、效能、交付、資安、模型/Token、完整度;可分發專案/Skill 的迭代與收尾預設納入 GitHub 安全更新檢查、回滾及舊版淘汰。
Scanned 9/4/2026
Install to Claude Code
npx -y skills add gabrielmoreira/agent-skills-mirror --skill run-benchmark-driven-rd --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Run Benchmark Driven Rd?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/gabrielmoreira-run-benchmark-driven-rd)More formats (shields.io, HTML) on the badges page.
---
name: run-benchmark-driven-rd
description: 將高風險產品目標轉成可否證、benchmark-gated R&D。適用架構、效能、交付、資安、模型/Token、完整度;可分發專案/Skill 的迭代與收尾預設納入 GitHub 安全更新檢查、回滾及舊版淘汰。
---
# Run Benchmark-Driven R&D
用同 provenance 證據做決策並保留可重播學習。
## 核心 contract
1. Active private Skill trees 是唯一權威。開始用 `invocation_revision_gate.py capture` 凍結 R&D+Cleanup revision,最終判斷前 `verify`;`UNSTABLE` 不混用,`STALE` 重讀並重跑受影響 gate。
2. 只在原始授權內修改;本機實作不暗含外部/破壞性 mutation。外部變更另需當輪授權、精確 target、preflight 與 post-readback;GitHub URL 不是執行授權。
3. 先寫 claim、baseline、threshold、decision rule,再用 task-shaped 正負控制校準 evaluator。Exit 0、元件/模型/短 prompt/工具綠燈不能替代品質、完成或 parity;缺同 provenance 證據保持 `unmeasured`/`NOT_CHECKED`。
4. Cleanup 是獨立 read-only provider;保留其 PASS/FAIL/REVIEW/NOT_CHECKED、child 與 revision evidence。R&D 可套 strict promotion policy,但不可把 provider 缺口改寫成 PASS。
5. Route receipt 列 intent、stage、artifact/risk、selected/omitted topics、hash、critical IDs、cost、fallback。未知/歧義/stale/超 budget 時 block 或 legacy fallback,不漏 critical rule。
6. Raw/private learning 留在目標 `.rd/`,預設零 prompt Tokens;shared rule 需 typed receipt、單一 owner、privacy、正負 fixture 與 reciprocal link。縮短 context 不刪記憶。
7. 可分發下游 Skill/software/game/installer/release 的已授權 implementation/promotion/completion 預設需要安全更新 capability;audit/source-only 不改檔,也不推導 publish、sign、persistence、delete 權限。
## 路由
| 任務 | 讀取 |
|---|---|
| 一般 experiment | `references/topics/core-experiment.md` |
| 架構/Cleanup baseline | `references/topics/architecture-and-evaluator.md` |
| context/Token/learning | `references/topics/context-and-learning.md` |
| capability/complete/external | `references/topics/completion-and-external.md` |
| updater 或可分發目標的 implementation/promotion/completion | `references/topics/secure-self-update.md`;依風險加 delivery/security,排除 audit/source-only |
Delivery、security、web、mobile、media、model、cross-system 讀對應既有 reference;大型 protocol/metrics/tooling/completion/external 只供 legacy/特殊 family。修改本 Skill 讀 `references/maintenance.md`。
## 工作流
1. 解析絕對 repo、授權、contract、obligations、failures、`.rd/`;產生含 typed `updateObligation`/capability floor 的 route。
2. 定義 claim matrix/provenance;校準 evaluator,以 Cleanup+native gate 凍結 baseline。
3. 做最小決定性實驗,保存 executable、stdout/stderr、exit、elapsed、input/environment hash、artifact、failure type。
4. 原子寫 typed project-local record;只有 promotion gate 通過才更新 shared topic/index,不把事故敘事 append 到 entrypoint。
5. 重跑相同 benchmark、負控制、delivered journey、obligations;最後修改後重驗 freshness/revision。
6. 分開報告 `measured stronger`、`likely but unmeasured`、`unchanged`、`regressed/open`,不以 PASS 數量冒充整體變強。
```powershell
python scripts/project_profile_gate.py --project <target> --contract <task-contract.json> --output <route.json> --quiet
```
Updater:side-by-side stage → digest/attestation → health → atomic switch → rollback → 淘汰 inactive 舊版;Skill 下次 invocation/restart 才切換。背景、自啟、破壞性 migration 分開 opt-in。

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!