# RUDO™ Weather Challenge Rules v1.1

rules_version: `RUDO-WC-RULES-1.1` · Season 1: El Niño Learning Season 2026–27 · Effective 2026-11-03

English text first, then 中文. Both versions have the same meaning. The sha256 of this file is published by the platform; any change to this file is a new rules version.

## English

### Article 1. The ruler

1.1 Station: Boston Logan airport (internal station id `USW00014739`). Everyone forecasts the same station, wherever they live.

1.2 Variable: the maximum temperature (TMAX) of the target climate day. Forecasts are entered in degrees Celsius with one decimal place and stored as an integer number of tenths of a degree (`C_TENTHS_INTEGER`).

1.3 Truth source: for Season 1, the NWS daily climate report for Boston (CLI, product CLIBOS). The value used is the one in the **first complete CLI** issued after the target climate day ends. The NWS monthly CF6 table is used as a cross-check and NCEI GHCN-Daily as an audit comparison. These sources have different day boundaries and revision schedules and are never swapped; changing the truth source requires a new rules version.

1.4 Climate day: as defined by the data provider. For the NWS CLI this is 00:00–24:00 local standard time (UTC−5) all year, with no daylight saving shift. In UTC, target day D runs from D 05:00 to D+1 05:00.

1.5 Snapshot and correction window: the observation is the archived raw product retrieved at a recorded time, with its sha256. If the provider issues a correction within 7 days after our snapshot, we re-score, publish a notice, and add a new result version. Results are **PROVISIONAL** during those 7 days and **FINAL** afterwards. Revisions after 7 days (including NCEI final quality control) are archived for audit only and do not change results.

### Article 2. The cutoff

2.1 Common cutoff: 18:00 EST on the day before the target day, which is 23:00:00 UTC on day D−1. Pages show the cutoff in EST, in the participant's local time and in UTC.

2.2 The server receive time decides, recorded in UTC with sub-second precision. A forecast is on time only if `received_at < cutoff_at` (strictly earlier). A forecast received exactly at the cutoff is late. Times are not truncated to the second before comparing. Client clocks and screenshots are not evidence.

2.3 A participant may submit several versions before the cutoff. The last complete and valid on-time version, by highest server sequence number, is the one scored. All versions are kept. Content received at or after the cutoff is recorded as `LATE_REVISION` and kept in the learning record only. Retrying with identical content returns the same receipt; changing A to B and back to A creates three versions. At most 20 submissions per participant per round are accepted.

2.4 Every scored round satisfies `opens_at < cutoff_at < first target start`.

2.5 The round definition and its `round_sha256` never change after the round opens. Late data, corrections and voids are recorded as observation snapshot versions and result versions that reference the original `round_sha256`. Receipts never need to be re-bound to a new round.

### Article 3. Rounds and tasks

3.1 Season 1 runs the next-day task only: one TMAX forecast per round. Multi-day windows (3-day, 7-day) and precipitation are not part of the Season 1 core set.

3.2 A round opens automatically for every target day. **Core** rounds form the Season 1 question set: S-01 on 2026-11-21, S-02 on 2026-11-23, S-03 on 2026-11-25, then every Saturday and Wednesday target day through 2027-02-28. All other rounds are **Practice** rounds, scored the same way and kept in the learning record.

3.3 If the observation for the target day is missing or fails quality control, the day is removed for everyone, including reference lines. Nothing is filled with zero.

### Article 4. Scores

4.1 Temperature error is the absolute difference between the forecast and the observation, in °C, computed **exactly**. With forecast t in tenths of °C (integer) and the observation F in whole °F (integer) as printed in the CLI: `absolute_error_C = |9t − 50F + 1600| / 90`, kept as a reduced fraction. The observation is never converted and rounded to 0.1 °C before scoring. Example: 57 °F = 125/9 °C; a forecast of 14.0 °C (t = 140) has error 1/9 °C, displayed as 0.11. If a value is not a whole °F, it is parsed as an exact decimal; if the source format cannot be parsed, scoring for that day pauses.

4.2 Reference lines are shown as diagnostic yardsticks only: M0 persistence (the latest observed high available before the cutoff, no later than D−2), M1 the NCEI 1991–2020 daily normal, M2 the raw NWS gridpoint forecast, and M5 a linear calibration of the NWS forecast on at least 10 past pairs (raw before that). They are submitted automatically through the same ledger and cutoff as everyone else. They are not competitors and are never ranked.

4.3 The unrounded value is authoritative. Displays use two decimals, rounded half-up from the exact fraction. Equal displayed values are treated as equal; hidden digits never break a tie.

4.4 Reasons, essays and other work never change a score.

4.5 Personal summary (shown only to the participant and their guardian): the mean error over rounds actually completed with a result, with equal weight per round; it always shows rounds completed, eligible rounds, the exact coverage fraction and the list of target days. Missed rounds are not penalised, not filled with zero and not filled with a reference line. Displays are half-up to two decimals; the exact fraction is stored.

4.6 Coverage of 60% or more of eligible rounds earns the "consistent participation" badge. The exact fraction is compared, never a rounded percentage. The badge does not create any comparison group.

### Article 5. Comparison and publication

5.1 By default each participant sees their own forecasts, observations, errors, differences from reference lines and history.

5.2 **Season 1 has no leaderboard and no percentiles.** Nothing like "you beat X% of your group" or "top X%" is shown. This is enforced in the API, not only hidden in the page.

5.3 Any future cross-participant ranking requires a new rules version published before that season starts, and must meet all of these: same season, task, data contract, age band and track; every compared participant completed 100% of a frozen core set (practice rounds excluded, void rounds removed for everyone); at least 20 eligible participants in the group who opted into publication; at least 60 distinct observed climate days. These are publication gates, not proof of statistical significance.

5.4 Age bands are never mixed: 10–12, 13–15, 16–17; the 18+ Open Division is separate.

5.5 Season 1 publishes only aggregate statistics: counts, the median error per round, reference line errors, and the hashes needed to recompute results. No codenames, names, schools, cities or individual results are published. A guardian's publication choice is recorded for possible future seasons only.

### Article 6. Eligibility and conduct

6.1 Accounts are zero-personal-data family accounts. An adult (18+) who is the parent or legal guardian creates a family after self-attesting to both facts and accepting the Terms. The family receives a one-time Family Key. No name, email, school, address, phone, photo or birthdate is collected. Each participant gets a system codename, an age band chosen by the guardian, and a coarse region (country or US state) used only for aggregate statistics.

6.2 Participants aged 10–17 take part only with their guardian's participation consent, which the guardian can switch off at any time. Adults 18+ join the Open Division the same way and consent for themselves. A guardian may add several children; each child is one participant. Season 1 has individual participants only.

6.3 A guardian may issue a Student Pass that lets a child submit forecasts for that one participant only. Classroom contacts may organise, but never replace a guardian's consent.

6.4 Any public material, model or AI may be used. Each submission must disclose whether AI was used (yes or no). Paid tools never add points.

6.5 Not allowed: more than one participant for the same person, submitting on behalf of someone else without their guardian's consent, tampering with receipts, harassing others, or posting personal information. Violations void that participant's Season 1 results; the learning record is kept.

6.6 Organisers, reviewers and their immediate families may take part but are excluded from any future published list.

### Article 7. Exceptions

7.1 If the truth source is missing or late by more than 48 hours, the round shows PENDING_DATA and is scored when the data arrives. After 14 days without data the round is VOID.

7.2 If a question is ambiguous, the round is VOID and the reason is published.

7.3 Cutoffs are not extended. If a confirmed platform-wide outage made fair submission impossible, the round is marked VOID (platform outage). No participant is penalised and a new round for a future target day is opened instead; any new cutoff is always before its target start. Void rounds do not count in the coverage denominator.

### Article 8. Appeals

8.1 An appeal may be filed within 5 calendar days after a result is published, by the guardian (or by an Open Division participant for themselves), from the family hub.

8.2 Review means re-running the public scorer on the frozen inputs. The re-run decides. Any change is added as a new correction version; the original record is never overwritten.

8.3 Appeals based on impressions rather than data, scoring or deadline facts are not accepted.

### Article 9. Data and rights

9.1 Data collection is minimal (see the Privacy notice). Rights in any child's work stay with the child and guardian; the organiser receives only the limited licence needed to score and run the challenge.

9.2 Publication, publishing, promotion and model training each need separate permission. Taking part grants none of them.

9.3 Children's work is never used to train models.

9.4 A guardian can export or delete a participant at any time. After deletion, identifying links are removed; the ledger keeps only hashes and anonymous results so that everyone else's results stay verifiable.

### Article 10. Organiser statements

10.1 RUDO is a learning and technology platform, not a weather agency. It does not act on behalf of NWS, does not replace NWS and gives no warnings. Never use these pages for safety decisions.

10.2 Use of public NOAA/NWS/NCEI data does not imply their partnership or endorsement.

10.3 Partners never buy rank, never receive children's data and never affect scoring.

### Changes in v1.1

Strict `<` cutoff with sub-second precision; latest on-time version by server sequence; round hash never changes, snapshots and results are versioned; exact rational error; no leaderboard and no percentiles in Season 1; no absence-filled ranking formula; the 60% threshold is a badge only; no cutoff extensions, platform outages lead to VOID; multi-day windows are training tasks, outside the Season 1 core set; zero-personal-data family accounts.

---

## 中文

### 第 1 条 尺子

1.1 站点:Boston Logan 机场站(内部站号 `USW00014739`)。无论住在哪里,所有人都预测同一个站点。

1.2 变量:目标气候日的最高气温(TMAX)。预测以摄氏度输入、保留一位小数,以整数「十分之一度」保存(`C_TENTHS_INTEGER`)。

1.3 真值来源:第一季固定为 Boston 的 NWS 每日气候报告(CLI,产品 CLIBOS),以目标气候日结束后发布的**第一份完整 CLI** 为准。NWS CF6 月表作交叉核对,NCEI GHCN-Daily 作审计对照。三者日界与修订节奏不同,不互换;更换真值来源须发布新的规则版本。

1.4 气候日:按数据提供方定义。NWS CLI 全年固定为当地标准时间(UTC−5)00:00–24:00,不随夏令时变化。换算为 UTC,目标日 D 为 D 日 05:00 至 D+1 日 05:00。

1.5 快照与更正窗口:实测以在记录时刻取回并存档的原件为准,并记录其 sha256。若提供方在快照后 7 日内发布更正,我们重新评分、发布公告,并新增结果版本。这 7 天内结果为**初步(PROVISIONAL)**,之后为**最终(FINAL)**。7 日后的修订(含 NCEI 最终质控值)只存档备查,不改变成绩。

### 第 2 条 截止

2.1 共同截止:目标日前一天 EST 18:00,即 D−1 日 UTC 23:00:00。页面同时显示 EST、参赛者当地时间与 UTC。

2.2 以服务器接收时间为准,按 UTC 记录、精度高于秒。只有 `received_at < cutoff_at`(严格早于)才算按时;恰好在截止时刻收到即为晚交。比较前不截断到秒。客户端时钟与截图不作证据。

2.3 截止前可提交多个版本。以服务器序号最大的、最后一份完整且合法的按时版本计分;所有版本都保留。截止时刻及之后收到的内容记为 `LATE_REVISION`,只进学习记录。内容相同的重试返回同一份回执;A→B→A 记为三个版本。每位参赛者每轮最多接受 20 次提交。

2.4 所有计分轮次满足 `opens_at < cutoff_at < 第一个目标日开始时刻`。

2.5 轮次定义及其 `round_sha256` 在开放后永不改变。数据延迟、更正与作废以观测快照版本和结果版本记录,并引用原 `round_sha256`;回执无需重新绑定到新轮次。

### 第 3 条 轮次与任务

3.1 第一季只开「次日」任务:每轮一个 TMAX 预测。多日窗口(3 日、7 日)与降水不属于第一季核心题集。

3.2 系统为每个目标日自动开放一轮。**核心轮**构成第一季题集:S-01 为 2026-11-21,S-02 为 2026-11-23,S-03 为 2026-11-25,之后每个周六、周三目标日,直到 2027-02-28。其余轮次为**练习轮**,评分方式相同,保存在学习记录中。

3.3 若目标日实测缺失或未通过质控,该日对所有人(包括参考线)一致剔除,不补零。

### 第 4 条 成绩

4.1 温度误差为预测与实测之差的绝对值(°C),**精确计算**。预测 t 为十分之一摄氏度整数,实测 F 为 CLI 中的整数华氏度,则 `absolute_error_C = |9t − 50F + 1600| / 90`,以最简分数保存。评分前绝不先把实测换算并四舍五入到 0.1 °C。例:57 °F = 125/9 °C;预测 14.0 °C(t = 140)的误差为 1/9 °C,显示为 0.11。若数值不是整数 °F,按精确十进制解析;源格式无法解析时,暂停该日评分。

4.2 参考线只作诊断标尺:M0 持续性(截止前可得的最新实测最高温,不晚于 D−2)、M1 NCEI 1991–2020 日常年值、M2 NWS 格点预报原始值、M5 用至少 10 对历史数据对 NWS 预报做线性校准(之前用原始值)。参考线经同一账本、同一截止时间自动提交;它们不是参赛者,永不参与排名。

4.3 以未四舍五入的值为准。显示为两位小数,由精确分数四舍五入(half-up)得到。显示值相同即视为相同,不用隐藏位数分胜负。

4.4 理由、作文和其他作品永不改变成绩。

4.5 个人汇总(只给参赛者本人和监护人看):对实际完成且有结果的轮次等权求平均误差;始终同时显示完成轮数、可参与的有效轮数、精确覆盖率分数和目标日清单。缺轮不罚、不补零、不用参考线补。显示按 half-up 保留两位小数,内部保存精确分数。

4.6 覆盖率达到有效轮次的 60% 及以上,获得「持续参与」徽章。按精确分数比较,不先四舍五入成百分比。徽章不构成任何比较组。

### 第 5 条 比较与公布

5.1 默认每位参赛者看到自己的预测、实测、误差、与参考线的差和历史。

5.2 **第一季不设排行榜,也不显示任何分位。**不显示「你超过了同组 X%」或「前 X%」之类的信息。此限制在 API 层执行,不只是在页面上隐藏。

5.3 将来如开放跨人比较,须在该赛季开始前发布新的规则版本,并同时满足:同赛季、同任务、同数据合同、同年龄段、同轨;被比较者 100% 完成事先冻结的同一核心题集(练习轮不计,作废轮对所有人一致移除);同组至少 20 位已选择公开的合资格参与者;至少 60 个不同的实测气候日。这些是发布门槛,不是统计显著性证明。

5.4 年龄段永不混排:10–12、13–15、16–17;18 岁以上公开组单列。

5.5 第一季只公布汇总统计:人数、每轮误差中位数、参考线误差,以及复算所需的哈希。不公布代号、姓名、学校、城市或任何个人成绩。监护人的公开选项只为将来可能的赛季做记录。

### 第 6 条 参赛资格与行为

6.1 账户为零个人信息的家庭账户。由身为家长或法定监护人的成年人(18 岁以上)自我声明这两点并接受条款后创建家庭,获得一次性显示的家庭密钥。不收集姓名、邮箱、学校、地址、电话、照片或出生日期。每位参赛者获得系统代号、由监护人选择的年龄段,以及只用于汇总统计的大区(国家或美国的州)。

6.2 10–17 岁参赛者须有监护人的参赛同意,监护人可随时关闭。18 岁以上成年人以同样方式参加公开组,由本人同意。一位监护人可以添加多个孩子,每个孩子是一位参赛者。第一季只设个人参赛。

6.3 监护人可发放学生通行码,让孩子只为该参赛者提交预测。课堂联系人可以组织,但不能替代监护人的同意。

6.4 允许使用任何公开资料、模型或 AI。每次提交须披露是否使用了 AI(是或否)。付费工具不加分。

6.5 禁止:同一人使用多个参赛者身份、未经对方监护人同意代为提交、篡改回执、骚扰他人、发布个人信息。违者该参赛者第一季成绩作废,学习记录保留。

6.6 组织方、评审及其直系亲属可以参赛,但不进入将来任何公布名单。

### 第 7 条 异常

7.1 真值来源缺失或延迟超过 48 小时,该轮显示 PENDING_DATA,数据到达后补评。超过 14 天仍无数据,该轮作废(VOID)。

7.2 题面有歧义时,该轮作废并公布原因。

7.3 不延长截止时间。经确认的平台级故障使公平提交无法成立时,该轮标为作废(平台故障),不罚任何参赛者,另开一个未来目标日的新轮;任何新截止都早于其目标开始时刻。作废轮次不计入覆盖率分母。

### 第 8 条 申诉

8.1 成绩公布后 5 个自然日内,可由监护人(或公开组参赛者本人)在家庭中心提出申诉。

8.2 复核即用冻结的输入重跑公开的评分程序,以重跑结果为准。如有变化,新增更正版本,不覆盖原记录。

8.3 不受理与数据、评分或截止事实无关的「印象类」申诉。

### 第 9 条 数据与权利

9.1 最小化收集(见隐私说明)。儿童作品的权利归孩子与监护人;组织方只获得评分与运行比赛所需的有限许可。

9.2 公开展示、出版、宣传、模型训练各需单独授权。参赛不默认授予其中任何一项。

9.3 绝不用儿童作品训练模型。

9.4 监护人可随时导出或删除参赛者。删除后移除可识别关联;账本只保留哈希与匿名结果,以便其他人的成绩仍可核验。

### 第 10 条 组织方声明

10.1 RUDO 是学习与技术平台,不是气象机构;不代表 NWS 行事,不替代 NWS,不发布预警。请勿据此做安全决定。

10.2 使用 NOAA/NWS/NCEI 公开数据,不代表其合作或背书。

10.3 合作伙伴不能购买名次,拿不到儿童数据,也不影响评分。

### v1.1 修订摘要

严格 `<` 截止、精度高于秒;按服务器序号选最后按时版本;轮次哈希不变,快照与结果分版本;精确有理数误差;第一季不设排行榜与分位;删除缺席补值排名公式;60% 只作徽章门槛;不延长截止,平台故障改为作废;多日窗口为训练任务,不进第一季核心题集;零个人信息家庭账户。