Today's agent evals only tell you whether it passed — not how it thought. AgentIRT scores every reasoning step and shows you where it breaks — before it ships, before it fails.今天的智能体评测只告诉你它有没有通过 —— 不告诉你它是怎么想的。AgentIRT 为每一步推理打分，指出它在哪里断裂 —— 在上线前，在出事前。
Agents now decide refunds, loans, and prescriptions — autonomously, at scale. The cost of a wrong call is no longer hypothetical.智能体如今自主、大规模地决定退款、信贷与处方。决策出错的代价不再是假设。
The EU AI Act puts high-risk systems under conformity obligations; the FDA scrutinizes clinical decision support. Both ask how the system reasoned — not just what it answered.EU AI Act 将高风险系统纳入合规义务；FDA 审视临床决策支持。两者都追问系统如何推理 —— 而不只是答了什么。
Pass/fail still grades the answer, not the reasoning. The one thing now being asked for is the one thing today's evals can't show.通过／失败仍只评答案、不评推理。如今被要求的那样东西，恰是今天的评测给不出的。
So the question is no longer can an agent act — but which agent can you trust?所以问题不再是 agent 能不能行动 —— 而是你能信任哪一个？
Paste a step-by-step reasoning trace (chain-of-thought, tool calls — one step per line). It's broken into steps, each scored, with the weak step flagged — then a 5-ability profile and a drill-down into the weakest.粘贴一段逐步的推理过程（思维链、工具调用 —— 每行一步）。它会被拆成步骤、逐步打分、标出薄弱步 —— 再给出 5 维能力画像并下钻到最弱维度。
Paste a trace and hit Analyze. Your step-by-step report appears here.粘贴过程并点分析，逐步报告会出现在这里。
Illustrative demo. Scoring here is a lightweight client-side heuristic for demonstration — production runs your real trace through the live AgentIRT (Process-IRT) engine. The sub-skill taxonomy shown is a proposed draft, to be replaced with your validated definitions.演示用途。此处评分为用于展示的轻量客户端启发式 —— 正式版会把你的真实推理过程交给在线 AgentIRT（Process-IRT）引擎。所示子技能分类法为提议草稿，将替换为你校验过的定义。
The same Item Response Theory (IRT) psychometrics use to measure human ability — adapted to grade each step a machine takes.与心理测量学衡量人类能力所用的同一套项目反应理论（IRT）—— 被改造来为机器走的每一步打分。
Break a run into discrete reasoning steps and tool calls — the "items" to be scored.把一次运行拆成离散的推理步骤与工具调用 —— 即待评分的"题项"。
Separate "step difficulty" from "agent ability" — harder steps done right count for more.将"步骤难度"与"智能体能力"分离 —— 难步做对，权重更高。
Process-IRTMultidimensional IRT resolves the score into a 5D profile, not one risk-hiding number.多维 IRT 把分数分解成 5 维画像，而非一个掩盖风险的数字。
5D-MIRTAdaptive testing reaches reliable conclusions from fewer tasks, producing auditable inputs for your compliance assessment.自适应测试用更少任务得出可靠结论，为你的合规评估产出可审计的输入。
Adaptive CATWe describe the science we build on, not our implementation. Item banks, calibration, and parameterization are proprietary.此处描述我们所依据的科学，而非具体实现。题库、参数标定与建模为专有技术。
Each cycle sharpens the agent and the item bank — a loop that compounds, for you and for us.每一轮循环都让智能体和题库更锐利 —— 一个不断累积的闭环，对你、对我们都是。
Choosing between models or agents, running each through a full benchmark is slow and token-expensive. Adaptive testing picks only the most informative tasks — fewer questions, same confidence.在多个模型或智能体之间选型时，让每个跑完整基准既慢又费 token。自适应测试只挑信息量最大的任务 —— 更少的题，同样的置信度。
It's not "fewer questions, so rougher" — each question is chosen for maximum information given the last answer, so fewer questions reach the same confidence. (Computer-adaptive GRE and TOEFL have worked this way for decades.)这不是"题少所以更粗" —— 每道题都根据上一题的作答挑选信息量最大的下一题，所以更少的题也能达到相同置信度。（计算机自适应的 GRE、TOEFL 已这样运作几十年。）
And you don't get one score — you get a 5D profile: select on the dimension you care about most — safety for medicine, goal adherence for support.而且你得到的不是一个分数，而是一张 5 维画像：按你最在意的维度选型 —— 医疗看安全，客服看目标遵循。
Illustrative figures shown to explain how adaptive testing works — not benchmarked results.此处为说明自适应原理的示意数字 —— 非基准测试结果。
One number hides the risk; a profile reveals it. Click any dimension and multidimensional IRT resolves it into specific sub-skills — weaknesses become visible, with the direction and data to add.一个数字会掩盖风险；一张画像会暴露它。点击任一维度，多维 IRT 将其分解为具体子技能 —— 弱点一目了然，并给出补强方向与数据。
A lab grading its own agent is a structural conflict of interest — regulators won't accept it, and enterprises shouldn't. AgentIRT is independent by construction.实验室给自家智能体打分是结构性利益冲突 —— 监管不会接受，企业也不该接受。AgentIRT 从构造上就独立。
Our independent assessment never grades an agent we built. We point the engine at our own BioAgent for internal QA — but a neutral assessment is only ever issued to agents we don't own.我们的独立评估从不评判自己造的智能体。我们把引擎对准自家 BioAgent 做内部 QA —— 但中立评估只发给我们不拥有的智能体。
Built on decades of peer-reviewed psychometric theory, and falsifiable — not one LLM judging another.建立在数十年同行评审的心理测量理论之上，且可证伪 —— 不是一个大模型评判另一个。
EU AI Act Art. 43 demands independent conformity assessment — we supply auditable evidence as an input to it; the compliance determination remains yours and your counsel's.EU AI Act 第 43 条要求独立合规评估 —— 我们提供可审计的证据作为其输入；合规判定仍归你与你的法律顾问。
Every agent we measure sharpens the item bank. The more we measure, the harder we are to catch — a flywheel newcomers can't buy.每测量一个智能体，题库就更锐利一分。测得越多，越难被追上 —— 这是新进入者买不到的飞轮。
Not one LLM rubber-stamping another. Every reasoning step is cross-examined by multiple judges, and each judge is weighted by its proven reliability — anchored to expert-graded ground truth, so an unreliable verdict counts for less.不是一个大模型给另一个盖章。每一步推理由多个评判者交叉复核，且每个评判者按其已验证的可靠度加权 —— 锚定在专家评定的标准答案上，不可靠的判断权重更低。
We prove the engine on our own hardest agent first — BioAgent, in medtech — then assess the agents we didn't build.我们先在自己最难的智能体上验证引擎 —— 医疗科技领域的 BioAgent —— 再去评估我们没造的智能体。
BioAgent guides medtech innovators through all 6 stages of the Stanford Biodesign process — and every reasoning step it takes is measured and governed by AgentIRT.BioAgent 引导医疗科技创新者走完 Stanford Biodesign 流程的全部 6 个阶段 —— 它走的每一步推理都由 AgentIRT 测量与治理。
AgentIRT began with a question in a Stanford course project on AI Measurement Science: if we have a rigorous science for measuring human ability, why not use it for machine reasoning? It brings psychometrics out of the classroom and turns it into an independent quality assessment for AI agents.AgentIRT 始于斯坦福一门 AI 测量科学课程项目里的一个问题：既然我们有衡量人类能力的严谨科学，为什么不把它用于机器推理？它把心理测量学带出课堂，变成面向 AI 智能体的独立质量评估。
AgentIRT turns that question into a measurable, auditable, independent answer.AgentIRT 把这个问题变成一个可测量、可审计、独立的答案。
Be first to measure and govern your agent. Early partners get priority access.率先测量与治理你的智能体。早期合作伙伴享优先接入。