每一条给三样东西：一句话定义、为什么它重要、出处。其他笔记引用这里的锚点，避免同一个词在不同文章里被解释成两回事。 Each entry gives a one-line definition, why it matters, and a source. Other notes link to these anchors so a term never means two different things in two articles.
定义：TypeSafe AI 提出的一类模型，目标是直接产出「软件能用的快速结构化决策」，而不是生成给人读的文本。 Definition: a class of models proposed by TypeSafe AI, built to produce fast structured decisions software can use directly, rather than text written for people.
为什么重要：把「模型输出」当成接口而不是文案，后面才能谈阈值、自动化与审计。名字取自 Kahneman 的 System 1（快、直觉）与 System 2（慢、推理）之分。[1] Why it matters: treating model output as an interface rather than copy is what makes thresholds, automation and audit possible. The name comes from Kahneman's System 1 / System 2 distinction.[1]
定义：Reinforcement Learning for Calibrated Decisions，官方给出的第三类训练范式，优化目标是「答案附带诚实、校准的概率」。 Definition: Reinforcement Learning for Calibrated Decisions — the third training paradigm in the vendor's framing, optimising for answers that carry honest, calibrated probabilities.
为什么重要：RLHF 优化人类偏好、RLVR 优化可自动验证的对错，两者都不直接约束「模型说自己有多确定」。RLCD 补的正是这一格。[1] Why it matters: RLHF optimises human preference and RLVR optimises verifiable correctness; neither constrains how confident the model claims to be. RLCD targets exactly that.[1]
RLHF：Reinforcement Learning from Human Feedback，用人类偏好当奖励信号，产出了今天所说的「对话模型」。RLVR：Reinforcement Learning with Verifiable Rewards，用可程序化验证的对错当奖励，是推理模型的路线。[1] RLHF uses human preference as the reward signal and produced today's chat models. RLVR uses programmatically verifiable correctness, the route behind reasoning models.[1]
定义：模型说「0.9」的那批判断里，实际对的比例接近 90%，就叫校准。反之叫过度自信。 Definition: among all judgements the model rates 0.9, roughly 90% should actually be correct. When that holds, the model is calibrated; when they are wrong far more often, it is overconfident.
为什么重要：概率只有在校准时才能拿来做阈值判断。一个 95% 能做对、但从不告诉你何时处于那 5% 的模型，无法被自动化的流程依赖。[1] Why it matters: a probability is only usable for thresholds when it is calibrated. A model that is right 95% of the time but never signals which 5% it is wrong on cannot be depended on inside an automated pipeline.[1]
定义：把预测概率分箱，比较每个箱内「平均置信度」与「实际正确率」的差距再加权平均，得到一个衡量校准好坏的数字。本文补充，非官方术语表内容。 Definition: bin the predicted probabilities, compare average confidence against observed accuracy in each bin, then take a weighted average. This entry is added by this note, not part of the vendor's own glossary.
定义：输出取值空间事先定义好（例如只能从五个标签里选一个），模型不可能产出范围外的值，也就不存在「解析失败」这一步。 Definition: the output space is defined in advance (say, one of five labels), so out-of-range values cannot occur and there is no parsing step to fail.
为什么重要：类型错误在 Agent 里只是麻烦，但在有延迟保证的系统里、或者在依赖链深处，是直接的故障。[1] Why it matters: a type error is an annoyance in an agent, but a hard failure inside a system with latency guarantees or deep in a dependency chain.[1]
定义：生成看起来合理但事实错误的内容。本文中的关键推论是——当输出空间被限定成有限选项、模型不生成自由文本时，这个失败模式的空间被大幅压缩。官方把它列为「零幻觉」，并承认该 0% 来自机制保证而非实测统计。[1][2] Definition: plausible-looking but false content. The key inference here: once the output space is a small finite set and no free text is generated, the room for this failure mode narrows sharply. The vendor claims "zero hallucinations" and notes that the 0% comes from a structural guarantee rather than a measured statistic.[1][2]
定义：在代码里规定「概率高于多少就自动执行，低于多少就交给人工」。这是把模型的不确定性翻译成业务流程的唯一支点。本文补充。 Definition: code that says how confident is confident enough to act, and below what level a human takes over. This is the single hinge that turns model uncertainty into an operating rule. Added by this note.
定义：一次决策可选结果的个数。官方提到单次支持到 255，更高时需要拆成「先打分、再显式选择」的两段式。这个上限会直接影响能不能把一个大问题一次问完。[1] Definition: how many possible outcomes a single decision may have. The vendor reports support up to 255 in one call, with a two-stage scoring-then-choosing scheme beyond that. The limit shapes whether a problem can be asked in one shot.[1]
定义：蒸汽机效率提升降低了单位煤炭消耗，却因为用得起的地方变多，总消耗反而上升。Jev 的名字取自 William Stanley Jevons；官方把它类比到「每单位智能的成本」：成本每降一个数量级，会解锁新的应用量级。[1] Definition: more efficient steam engines lowered coal per unit of work yet raised total coal use, because more uses became affordable. The name Jev comes from William Stanley Jevons; the vendor applies the analogy to the cost of intelligence — each order of magnitude cheaper unlocks an order of magnitude more uses.[1]
| 版本 | 日期 | 改动Change |
|---|---|---|
| v1.0 | 2026-09-25 | 初版：10 条术语，依据官方发布文与官网首页。First version: 10 entries, based on the launch post and the homepage. |