A family weather-research platform · LIGHT HOPE ecosystem 家庭气象研究平台 · LIGHT HOPE 生态
Predict before. Verify after. Keep every miss. 天气之前留下预测，天气之后诚实打分，错的永远保留。
Youth Weather Lab is a kid-initiated, AI-assisted, evidence-verified research platform. Every day, a young scientist leaves a timestamped forecast before the weather happens — then scores it against official observations after. Nothing is edited. Nothing is deleted. 少年气象实验室是一个由孩子发起、AI 协助、以证据验证的研究平台。每天，小科学家在天气 发生之前留下带时间戳的预测，天气发生之后用官方观测打分。 不许改，不许删。
“I don’t just read the forecast — I check whether it was right.” 「我不只是查看天气预报，我还会验证它。」
Why this exists为什么做这个
The world has ten thousand weather apps. It has almost no weather report cards. 世界上有一万个查天气的网站，却几乎没有给天气预报打分的成绩单。
This is not another site that shows you today’s temperature. Its whole reason to exist is the verification loop: forecasts are frozen before the fact, scored against official observations after the fact, and the record — including every wrong call — is kept forever. It is how a child learns what science actually feels like: making a claim you can be wrong about, and finding out. 这不是又一个展示当前温度的网站。它存在的全部理由是验证闭环：预测在事前冻结， 事后用官方观测打分，全部记录——包括每一次错误——永久保留。这正是孩子体验真正科学的方式： 提出一个可能出错的判断，然后亲眼看它对不对。
How it works怎么玩
One honest loop, every day. 每天一个诚实的闭环。
Predict & freeze预测并冻结
Each morning the lab freezes forecasts from seven sources — NWS, four global and regional models, a persistence baseline, and the kid’s own rule cards — across six lead times, from 6 hours to 7 days out. Entries are append-only: the database itself refuses updates and deletions. 每天早晨，实验室冻结七个来源的预测——NWS、四个全球/区域模型、一个 persistence 基线， 以及孩子自己的规则卡——覆盖从 6 小时到 7 天的六个提前量。记录只增不改： 数据库本身拒绝任何更新和删除。
Weather happens天气发生
Official observations arrive from real stations. Severe-weather alerts are stored and shown verbatim from the National Weather Service — physically separated from any AI explanation, so authority is never blurred. 官方观测数据来自真实气象站。恶劣天气警报逐字保存并逐字展示，来源是 NWS—— 并与任何 AI 解释物理隔离，权威信息绝不被混淆。
Score in three states三态打分
Every frozen prediction gets exactly one verdict. If the observation is missing, the verdict is UNVERIFIABLE — never silently skipped, never guessed. 每条冻结的预测得到且只得到一个判定。观测缺失时判定为「无法核验」—— 绝不悄悄跳过，绝不臆造。
Hard rule under the hood: no number on any page may be generated by an AI model. Every figure must trace to a registered data source — if the source can’t be reached, the page says UNAVAILABLE instead of making something up. 底层硬规则：页面上任何数字都不得由 AI 模型生成。每个数字必须能追溯到登记在册的数据源—— 取不到源时，页面显示「不可用」，而不是编一个。
The Arena模型擂台
Seven contestants. One scoreboard. No mercy. 七位参赛者，一张记分板，谁也不留情面。
The lab’s scoreboard pits the official forecast, four numerical models, an AI model, and one 10-year-old against the simplest baseline in meteorology: persistence — “tomorrow will be like today.” Any forecaster that can’t beat persistence has no skill. It’s the one skill score a kid understands instantly. 记分板让官方预报、四个数值模型、一个 AI 模型和一位十岁小朋友，共同对阵气象学里最朴素的 基线：persistence——「明天和今天一样」。打不过 persistence 的预报就等于没有技巧。 这是唯一一个孩子能立刻理解的技巧分。
| Contestant参赛者 | What it is是什么 | Source来源 |
|---|---|---|
| NWS | The official U.S. forecast美国官方预报 | api.weather.gov |
| GFS | NOAA’s global physics modelNOAA 全球物理模型 | Open-Meteo |
| HRRR | NOAA’s high-resolution short-range modelNOAA 高分辨率短时模型 | Open-Meteo |
| ECMWF IFS | Europe’s flagship physics model欧洲旗舰物理模型 | Open-Meteo |
| AIFS | ECMWF’s AI weather modelECMWF 的 AI 天气模型 | Open-Meteo |
| Persistence | “Tomorrow = today.” The bar everyone must clear.「明天=今天」。所有人必须跨过的门槛。 | baseline基线 |
| The kid小科学家 | Hand-written rule cards — her own forecasting rules, in her own words手写规则卡——她自己的预报规则，用她自己的话 | rule cards规则卡 |
Standing falsifiable hypothesis常设可证伪假说
“AI weather models underperform physics models on record-breaking, out-of-distribution extreme events.” 「在破纪录、超出训练分布的极端天气事件上，AI 天气模型不如物理模型。」
Every Nor’easter is a trial: the lab pulls the AI model (AIFS) out of the pack and watches whether it falls below the physics model (HRRR). The hypothesis is written on the lab’s front page precisely so the data can prove it wrong. 每一场东北风暴都是一次审判：实验室把 AI 模型（AIFS）单独拉出来，看它会不会掉到物理模型 （HRRR）之下。这条假说写在实验室首页，正是为了让数据有机会推翻它。
House rules五条铁律
Five invariants, each machine-checked. 五条硬不变量，每条都机械可检。
These are not values statements. Each rule is enforced by code — database triggers, automated lints, offline tests — and a violation fails the build. 这些不是价值观口号。每条规则都由代码强制执行——数据库触发器、自动化检查、断网测试—— 违反任何一条，构建直接失败。
INV-1
Predictions are append-only预测只增不改
No prediction is ever updated or deleted. The database raises an error on any attempt. Your misses are part of the record.任何预测都不能更新或删除，数据库层面直接报错拒绝。错的记录也是记录。
INV-2
Official alerts stay verbatim官方警报逐字保留
Severe-weather alerts are stored and displayed word-for-word from NWS, physically separated from any AI commentary.恶劣天气警报逐字保存、逐字展示，来源为 NWS，并与任何 AI 解说物理隔离。
INV-3
No AI-generated numbers数字不许 AI 生成
Every number must trace to a registered source. If the source is unreachable, the page shows UNAVAILABLE — a fake number is a hundred times worse than no number.每个数字必须能追溯到登记在册的来源。取不到源就显示「不可用」——假数字比没数字坏一百倍。
INV-4
Child identity stays private孩子身份严格隔离
The public face is separated from the child’s identity by an automated privacy lint: personal details in any public page fail the build.公开面与孩子身份由自动化隐私检查隔离：任何公开页面出现个人信息，构建即失败。
INV-5
Three-state honesty三态诚实
Every scored prediction is PASS, MISS, or UNVERIFIABLE. Missing observations are counted and shown — never quietly dropped.每条被评分的预测只能是「猜对 / 猜错 / 无法核验」。缺失的观测被显式计数展示——绝不悄悄丢弃。
Who it’s for给谁用
Built around one real kid. Designed for every curious family. 围绕一个真实的孩子而建，为每个好奇的家庭而设计。
For young scientists (~8–12)给小科学家（约 8–12 岁）
The lab’s first user is a real 10-year-old, and the product bends to her reality: 实验室的第一位用户是一个真实的十岁孩子，产品完全围绕她的现实设计：
- A daily card that takes about three minutes — a form that takes five means the project dies next week.每日卡三分钟填完——要花五分钟的表单，项目下周就废。
- Probabilities in six simple boxes (0–100%), not decimals.概率用六个简单格子（0–100%）表达，不用小数。
- Jargon a kid can’t understand is treated as a dead feature.孩子看不懂的术语等于死掉的功能。
- An interactive rain-vs-snow phase explorer for building intuition.互动式雨雪相态滑块，帮助建立直觉。
For parents给家长
What your child practices here is the shape of science itself: 孩子在这里练习的，正是科学本身的形状：
- Commit to a claim before the answer is known.在答案揭晓之前，先立下自己的判断。
- Measure against official ground truth, not feelings.用官方实测数据打分，而不是凭感觉。
- Calibration over bravado: learning what “80% sure” should feel like.练校准而非逞强：体会「八成把握」到底该是什么感觉。
- Privacy by construction — the child’s journal is private and never leaves home; the public face is machine-checked for identity leaks.隐私内建于架构——孩子的日记完全私有、不离开家庭；公开面经机器检查，杜绝身份泄露。
Get involved参与
Where the lab stands, and how to follow along. 实验室走到哪了，以及如何跟进。
Phase 1 is a working system running locally for one family: real data pipelines (NWS, Open-Meteo, ECMWF AIFS), a real scoreboard, real machine-checked invariants. A public phase — where other families could run their own lab — is a direction, not a promise. No accounts, no payments, nothing for sale. Phase 1 是一个真实在跑的本地系统，服务于一个家庭：真实数据管线（NWS、Open-Meteo、 ECMWF AIFS）、真实记分板、真实的机器检查铁律。公开阶段——让其他家庭跑自己的实验室—— 是一个方向，不是一个承诺。没有账号，没有付费，不卖任何东西。
Curious? Say hello.感兴趣？打个招呼。
If you’re a family who’d want to run a weather lab of your own, an educator, or just weather-curious, we’d genuinely like to hear from you. Interest from real families is exactly what decides whether a public phase happens. 如果你是想为自己家搭一个气象实验室的家庭、是教育工作者，或只是对天气好奇，我们真心 希望听到你的声音。真实家庭的兴趣，正是决定公开阶段是否启动的依据。
Free, non-commercial family research project. Contact is the only call to action on this page — by design. 免费、非商业的家庭研究项目。本页唯一的行动号召就是联系我们——这是刻意为之。