← 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

评测用的模型由你接入：官方页面在册 OpenAI 与 Anthropic 凭据在 Settings 的 Models 里配（当日直抓）。 scorer 面分代码、LLM judge 与 autoevals 三类（guides/evals 当日在册）。 judge 支持的厂商全清单、以及是否统一经网关转发未逐条核验，记为未知。

You bring the models used for evaluation: official pages are on record that OpenAI and Anthropic credentials are configured under Settings → Models (direct-fetched today). The scorer surface splits into code, LLM judges and autoevals (guides/evals fetched today). The full list of supported judge vendors, and whether calls are unified through the gateway, were not itemized — recorded as unknown.

运行位置Where it runs

主要交互面是浏览器里的控制台（官方文档当日在册的定位是「active observability platform」）； 评测也能在本机与 CI 用 bt CLI 跑（/cli 当日在册）。数据面可选部署进客户环境， 控制面留在厂商侧管认证、用户与元数据（/self-hosting 当日在册）。

The main interaction surface is the browser console (the docs' positioning today is "active observability platform"); evals can also run locally and in CI through the bt CLI (/cli fetched today). The data plane can be deployed into the customer environment while the control plane stays vendor-side, holding authentication, user management and metadata (/self-hosting fetched today).

本地文件Local files

默认数据在厂商侧；走客户侧数据面时 PostgreSQL、Redis、对象存储与 Brainstore 装在你自己的环境 （当日在册），控制面只保存认证与元数据。平台不读本机代码库； bt CLI 在本地读写评测配置与数据同步（当日在册）。

By default the data sits vendor-side; with the customer-side data plane, PostgreSQL, Redis, object storage and Brainstore run in your own environment (fetched today), while the control plane keeps only authentication and metadata. The platform does not read your local codebase; the bt CLI reads and writes local eval configuration and syncs data (on record today).

关机后的任务Tasks after shutdown

生产 trace 用在线打分异步评测（guides/evals 当日在册），实验（experiments）批量跑数据集做回归对照。 抽样比例、评分延迟与告警时序的口径在官方故障页有提及，但未逐条核验，记为未知。

Production traces are scored asynchronously via online scoring (guides/evals fetched today), and experiments run datasets in batch for regression comparison. Sampling ratios, scoring latency and alert timing are touched on in official troubleshooting pages but were not verified item by item — recorded as unknown.

工具与扩展Tools & extensions

bt CLI 经 curl、npm、pnpm 或 mise 安装，可本地与 CI 跑评测、管理函数、查日志、同步数据 （当日在册）；SDK 面覆盖 Python 与 TypeScript；文档索引在册含 OpenTelemetry 发送 trace 的集成页， 目标页正文本次未抓。MCP 接入方式未核验到官方声明，记为未知。

The bt CLI installs via curl, npm, pnpm or mise and runs evals locally and in CI, manages functions, queries logs and syncs data (fetched today); SDKs cover Python and TypeScript; the docs index is on record with an OpenTelemetry trace-sending integration page whose body was not fetched. MCP support was not verified in official statements — recorded as unknown.

上下文与记忆Context & memory

trace、实验与数据集在同一面板里对照；人工评审面支持一行填多个分数 （官方页面当日在册）。数据集版本与对照口径未读到正文，记为未知； 「跨会话记忆」不是本类平台的功能面。

Traces, experiments and datasets are compared in one panel; the human-review surface supports multiple scores on a single row (on record from official pages today). Dataset versioning and comparison semantics were not read — recorded as unknown; cross-session memory is not a feature surface for this class of platform.

权限与限制Permissions & limits

合规与加密当日在册：AES-256、SOC 2 Type II、HIPAA，另称推理输入输出不由模型宿主留存（/security）。 档位与配额事实为免费至付费档、企业可自定义保留期，配额随 trace 量、数据量与每月分数条数变化 （pricing 当日在册），价格数字按 D9 不采。数据是否用于训练未读到条款正文，记为未知。

Compliance and encryption are on record from today's fetch: AES-256, SOC 2 Type II and HIPAA, plus the statement that inference inputs and outputs are not stored by the model host (/security). Tier and quota facts run from free to paid tiers with custom enterprise retention, and quotas vary with trace volume, data volume and monthly score counts (pricing fetched today); price numbers are excluded per D9. Whether data trains models was not found in clause text — recorded as unknown.

适合什么任务Best for

适合把评测当发布闸门、要在 PR 与 CI 里跑实验对照，又不想自运维整套观测栈的团队。 不适合要整机自托管（控制面仍在厂商侧）、或想把平台再分发对外提供的团队； 本站只记支持的评测类型，不记分数也不排名。

Fits teams that treat evaluation as a release gate — running experiments on PRs and in CI — and would rather not operate the whole observability stack. When data must stay in your own environment there is a hybrid "customer data plane plus vendor control plane" deployment, but the control plane remains vendor-side, so full self-hosting is not its model. Not a fit for teams intending to redistribute the platform. This site records supported evaluation types, not 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.