docs: require running the harness as it ships (no extra skills)
Address fairness concern from #24: the same harness can load completely different skills / plugins / MCP / custom system prompts, which shifts the output as much as swapping the harness and breaks Agent comparability. - README.md / README.en.md: add a bullet to the Model + Harness section stating the harness is evaluated in its default, out-of-the-box form. - AGENTS.md: add a "Run the harness as it ships" rule — no extra skills beyond what the harness bundles; disclose any deviation in `note`. Enforcement stays honor-system + the PR generation screenshot. Closes #24 Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
1 parent
eaf2737ecc
commit
5cd30bac2b
3 files changed
+14
No files matched your search
@@ -21,6 +21,7 @@ NAGI STUDIO 的 LLM 测评案例集:同一段提示词,不同「模型 × Ha
|
||||
- **模型(Model)**:权重与推理引擎本身,如 GPT-5.5、Gemini 3.1 Pro、Claude Fable 5。思考配额(effort)是模型自身的能力旋钮——同一份权重在高/低思考强度下的推理深度判若两个模型,因此它和权重同属「模型」这一维度,脱离 effort 谈模型表现并不公允;本仓库默认拉满(Max),并在每个组合里如实标注。
|
||||
- **Harness(运行环境)**:包裹模型的产品/脚手架,如 Codex CLI、Cursor、AntiGravity、Claude 网页版——它决定工具调用、系统提示词、上下文管理与续写策略,对最终产出的影响往往不亚于模型本身。
|
||||
- 因此本仓库的测评单位是一个 **Agent**:同一个模型(含其思考配额)换一个 Harness,就是另一个 Agent,记为不同条目(例如 GPT-5.5 Pro 与跑在 Codex CLI 里的 GPT-5.5(xhigh)是同一模型家族的两个 Agent)。
|
||||
- **测的是 Harness 开箱即用的默认形态**:请勿额外加载 Harness 自带以外的 skill / 插件 / MCP / 自定义系统提示——它们对产出的影响不亚于换 Harness,会破坏 Agent 之间的可比性。确实用了,就不是同一套 Agent,必须在 `note` 里如实说明。
|
||||
|
||||
## 已测组合 / Registry
|
||||
|
||||
|
||||
Reference in new issue
Block a user