← Safety

核验标记表示资料完整度，不是产品评分。下列未确认项可能影响选型；请按对应来源逐项确认。The confidence mark describes evidence completeness, not a product score. Unconfirmed items below may affect your choice; check them against the relevant sources.

English summaries and interface labels are available. Detailed source notes below are currently in Chinese.

模型与开放条件Model & openness

平台自己不产内容，评测量靠你配的模型——LLM-as-a-Judge 走「LLM Connection」， 凭据由用户自配（judge 文档当日直抓在册）；被观测的应用可接任意框架与模型， 自托管另有「Bring your own LLM API」口径（self-hosting 页当日在册）。 judge 可用模型全清单与版本映射本次未逐条清点，记为未知。

The platform produces no content itself; the evaluation capacity depends on models you configure — LLM-as-a-Judge runs through an "LLM Connection" with credentials you supply (judge doc direct-fetched today). Observed apps can use any framework or model, and self-hosting carries a "Bring your own LLM API" line (self-hosting page fetched today). The full judge-model list and version mapping were not itemized today — recorded as unknown.

运行位置Where it runs

开源可自部署（Docker Compose 起测试、Kubernetes/Helm 起生产，self-hosting 页当日在册）， 另有官方云（AWS 美/欧/日三区，security 页当日在册）；交互面是 Web 控制台加 SDK 与 API。 本机跑的是平台服务与界面，不在本机跑被评测的模型。

Open source and self-deployable (Docker Compose for testing, Kubernetes/Helm for production, per the self-hosting page fetched today), plus an official cloud (AWS US/EU/JP regions, security page fetched today); the interaction surface is a web console with SDKs and an API. What runs on your machine is the platform service and UI, not the models under evaluation.

本地文件Local files

自托管时 trace、score、prompt、数据集都落在自有 PostgreSQL、ClickHouse、Redis 与对象存储 （当日在册）；数据靠 SDK、OpenTelemetry、API 上报，平台不读本机代码库； 导出与删除接口面本次未逐条清点。

When self-hosted, traces, scores, prompts and datasets live in your own PostgreSQL, ClickHouse, Redis and object storage (fetched today); data arrives via SDK, OpenTelemetry and the API — the platform does not read your local codebase; the export and deletion interface surface was not itemized today.

关机后的任务Tasks after shutdown

在线评测在生产 trace 上异步打分（evaluation overview 与 judge 文档当日在册）， 实验（Experiments）批量跑数据集；关机续跑不是这类平台的功能面。 抽样比例、评分延迟与失败重试口径未逐条核验，记为未知。

Online evaluation scores production traces asynchronously (evaluation overview and judge docs fetched today); Experiments run datasets in batch. Resume-after-shutdown is not a feature surface for this kind of platform. Sampling ratios, scoring latency and retry semantics were not verified item by item — recorded as unknown.

工具与扩展Tools & extensions

Python 与 JS SDK、REST API、OpenTelemetry 摄入；官方文档 MCP server 当日直抓在册， 它只把文档暴露给 agent，能否读 trace 与 score 的鉴权 MCP 面未核验； CI/CD 用于「回归时挡住部署」（evaluation overview 当日在册）。

Python and JS SDKs, REST API, OpenTelemetry ingestion; the official documentation MCP server is on record from today's fetch and only exposes docs to agents — whether an authenticated MCP surface can read traces or scores is not verified. CI/CD is used to "block deploys on regressions" (evaluation overview fetched today).

上下文与记忆Context & memory

观测对象按项目组织，trace 与 score 带来源关联；分数类型 NUMERIC、CATEGORICAL、BOOLEAN、TEXT 四类当日在册；从 trace 沉淀数据集、prompt 版本管理在产品面在册。 跨项目记忆不是平台功能面，官方未声明。

Observed objects are organized per project, with traces and scores linked to their source; the four score types NUMERIC, CATEGORICAL, BOOLEAN and TEXT are on record from today; turning traces into datasets and prompt versioning are on the product surface. Cross-project memory is not a platform feature and is not declared by the vendor.

权限与限制Permissions & limits

自托管可 air-gapped，数据留在自有环境；云版 SOC 2 Type II、项目级 RBAC 与 SSO （security 页当日在册）；档位与配额事实（Hobby 与 Core、Pro、Enterprise 四档， 用量单位与数据访问时长随档位变化）当日在 pricing 页在册，价格数字按 D9 不采。 是否用客户数据训练模型未读到条款正文，记为未知。

Self-hosting can be air-gapped so data stays in your environment; the cloud is SOC 2 Type II with project-level RBAC and SSO (security page fetched today). Tier and quota facts (Hobby plus Core, Pro, Enterprise; usage units and data-access windows vary by tier) are on the pricing page from today, with price numbers excluded per D9. Whether customer data trains models was not found in clause text — recorded as unknown.

适合什么任务Best for

适合要把数据留在自有环境（自托管或气隙部署）、观测与评测共用同一套 trace、 把线上样本沉淀成数据集再在 CI 挡回归的团队。不适合只想用厂商侧托管红队或合规审计、 不愿自备 judge 模型密钥的场景；本站不测分也不排名。

Fits teams that must keep data in their own environment (self-hosted or air-gapped), want observation and evaluation on one trace store, and turn production samples into datasets that block regressions in CI. Not a fit for those who only want a vendor-hosted red-team or compliance-audit service, or who will not bring judge-model credentials. This site does not measure scores and does not rank.

本页由This page is generated from the ai-handbook 数据层生成（CC BY 4.0）。data layer (CC BY 4.0). 方法论与坐标系定义见仓库内For the methodology and coordinate definitions, see METHODOLOGY.md.