How we verify · Updated 2026-08-21
How we verify
We trace important judgments to published sources, separate developer claims from independent validation, keep limitations visible and publish substantive corrections.
我们把重要判断追溯到公开资料，区分开发者声明与独立验证，保留证据限制，并公开实质纠正。
Short answer / 一句话结论
The site does trust layering, not truth certification. Source role, claim type, status and confidence answer different questions and are never collapsed into one truth score.
Changed on 2026-08-21: the review protocol below is withdrawn. This site no longer
promises to recheck claims, sources or articles on a schedule, and the reviewDue dates that
carried that promise have been removed. Existing review dates are real dates on which a check happened and
are retained; they are history, not a freshness guarantee. Sources still point at original material and
numbers are still checked against it — that part never depended on a schedule. Full notice on
For agents.
本站做的是可信分层，不是绝对真相认证。来源角色、判断类型、状态和置信度回答不同问题，不会被合并成一个“真假分数”。
2026-08-21 变更：下文的复核协议已撤销——不再承诺按排程重查 claim、来源或文章，
承载这一承诺的 reviewDue 日期已移除。既有复核日期是真实发生过的核对，原样保留，只作历史，不是新鲜度保证。
来源仍必须指向原始出处、数字仍要回原文核对——这部分从不依赖排程。完整说明见 For agents。
Source role and claim-source fit / 来源角色与判断适配
Provenance and validation are separate. A developer source is the best evidence for what its developer claimed or tested; an independent source is needed to establish what survives outside that protocol. Neither role is universally above the other. Every source is also checked against the exact wording, scope, date and boundary of the claim it supports.
来源出处与独立验证是两件事。开发者来源最适合证明“开发者说了什么、测试了什么”；独立来源用于判断结果能否离开原协议后仍然成立。 两者不是简单的高低排序。每个来源还必须直接适配当前判断的具体措辞、范围、日期和边界。
Technical source-role vocabulary / 技术来源角色词表
| Provenance role | What it can establish | What it cannot establish alone |
|---|---|---|
developer_primary |
The developer's architecture, protocol, reported result, release date and product framing. | Independent reproduction, broad downstream utility or deployment safety. |
independent_validation |
External reproduction, comparative testing, measured-world calibration or deployment boundary. | The developer's intended scope when the primary material says something narrower. |
secondary_reporting |
Context, interviews, field reaction and leads to stronger sources. | A durable capability conclusion without corroboration. |
discovery_only |
A lead that should be checked elsewhere. | Any published factual or capability claim by itself. |
Claim-source fit is assessed as direct, partial, indirect or unsupported. This assessment is reflected in the claim's confidence and boundary; a direct developer quote can support a high-confidence statement about the developer's claim without proving that the underlying capability is independently true.
判断与来源的适配分为直接、部分、间接或不支持。结果体现在 confidence 与 boundary 中：开发者原文可以高置信地支持 “开发者作出了这项声明”，但不自动证明对应能力已经被独立验证。
Full editorial policy: claim fields, review and publication rules / 完整编辑政策
Claim types / 判断类型
Claim type describes where a statement comes from. It does not rank importance. A company product claim may be important, but it remains a product claim until stronger evidence appears.
判断类型说明一句话来自哪里，不等于重要性排序。一个公司产品声明可能很重要，但在更强证据出现前，它仍然只是产品声明。
| Type | Meaning | Boundary |
|---|---|---|
official_claim |
A company or lab says this about its own system. | Not independent validation. |
technical_report_claim |
A claim made in a paper, benchmark note or technical report. | May be task-specific or dependent on benchmark design. |
product_claim |
A capability or workflow exposed through a product page or launch. | Can change quickly as access, pricing and features change. |
editorial_judgment |
World Model Atlas synthesis based on one or more sources. | Must stay traceable to sources and open to correction. |
Claim status / 判断状态
Status is lifecycle state, not provenance. It records whether the current wording can still be used.
unmaintained: the wording and its boundary stand as written, and no one is re-checking them. Since the 2026-08-21 freeze this is what every claim carries.unverified: support is insufficient or was never checked.outdated: the source or product moved on and the wording no longer matches it.conflicting: credible sources disagree and the conflict is left visible rather than resolved.
current was retired on 2026-08-21. It had meant "reviewed wording remains usable", which was a
promise about the future backed by a review schedule this site no longer runs.
status 是对记录的描述性标注，不是来源类型：unmaintained 表示措辞与边界按原样成立、但已不再复核，
2026-08-21 冻结后所有 claim 都是这一档；unverified 表示未充分核验；outdated 表示来源或产品已变、措辞不再对应；
conflicting 表示可信来源存在冲突且公开保留。原有的 current 已于 2026-08-21 停用——
它承诺的是"往后仍可使用"，而支撑这个承诺的复核排程已经取消。
Confidence and claim-source fit / 置信度与判断适配
- High: the cited material directly supports the exact wording, scope and boundary.
- Medium-high: support is strong but one meaningful scope, protocol or transfer limitation remains.
- Medium: the statement is a bounded synthesis or depends on qualitative or incomplete evidence.
- Low: support is indirect, stale or insufficient for prominent use.
Confidence is local to the sentence and its sources. It is not a universal truth score, a model capability score or a substitute for independent validation.
置信度只描述当前这句话与所列来源的适配强度，不是通用真实性分数、模型能力分数，也不能替代独立验证。
Review protocol: withdrawn / 复核协议：已撤销
This site no longer runs a review cadence. Until 2026-08-21 this section set out when a
claim had to be rechecked, and every claim, source and article carried a reviewDue date to
trigger it. Those dates have been removed, reviewPolicy in the field schema is marked
withdrawn, and every claim now carries status: "unmaintained".
The rule was withdrawn rather than quietly dropped because it was already being broken. On 2026-08-21, four
Sora claims, one source and one article had passed their reviewDue date while still labelled
current — a state the old rule itself defined as one to correct. Between leaving a promise the
site does not keep and withdrawing the promise, withdrawing it is the honest option.
What this does not change. Sources still point at original material — an official blog, a paper, a technical report — rather than second-hand coverage, and numbers and proper nouns are still checked against the original before anything is published. Those are properties of how a record is written. They were never promises about how often it is revisited, and they survive the withdrawal intact.
Corrections still happen. They happen when an error is noticed or reported, not on a
schedule. Write to [email protected]; a factual
correction is made first and discussed after. Existing lastReviewed dates are real dates on
which a record was checked and are retained unchanged — read them as history, not as a freshness guarantee.
本站已不再执行周期性复核。2026-08-21 之前，这一节规定了什么情况下必须重新复核，
每条 claim、来源和文章都带一个 reviewDue 日期来触发它。这些日期已全部移除，
reviewPolicy 标记为 withdrawn，所有 claim 的 status 改为 unmaintained。
之所以是"撤销"而不是"悄悄不做"：到 2026-08-21，已有 4 条 Sora claim、1 个来源和 1 篇文章越过了自己的
reviewDue 却仍标着 current——这正是旧规则自己定义的应当纠正的状态。
不变的部分：来源仍必须指向原始出处而非二手转载，数字与专有名词仍要回原文核对；
这些是"记录怎么写"的属性，从来不是"多久重查一次"的承诺。纠错仍然进行，
但由发现或举报触发，不按排程。既有的 lastReviewed 是真实核对日期，原样保留，只能读作历史，不是新鲜度保证。
Publication and correction policy / 发布与纠错政策
Content updates and maintenance records serve different purposes. Every published editorial or site-maintenance change remains in updates.json; only changes that give readers a reason to read or re-read content enter the Updates page and RSS.
读者更新与维护审计承担不同职责。所有已发布的编辑变化或站点维护都保留在 updates.json；只有让读者有理由首次阅读或重新阅读内容的变化，才进入 Updates 页面与 RSS。
World Model Pulse is a separate field layer
Pulse records selected public activity from tracked teams. An item needs a primary source, date precision, retrieval date, provenance role and visible evidence limitation, but it is not an Atlas article, an independent capability judgment or a reader update about this site.
| Change type | Reader feed rule | Required public record |
|---|---|---|
article |
Reader, when a new article provides a distinct answer or reference asset. | Set datePublished, dateModified, claims, sources and one reader update. |
material_revision |
Reader, only when evidence, scope or the practical conclusion changes enough to justify re-reading. | Update dateModified; state whether the conclusion changed and list affected claims or sources. |
correction |
Reader when a published factual conclusion changes; otherwise audit. | Record old and new wording in the article or update note when interpretation changes. |
source_status_change |
Reader only when source loss, conflict or stronger evidence changes a visible conclusion. | Update source status, affected claim status/confidence and lastReviewed. |
maintenance |
Audit only: navigation, schema documentation, identity, domain and visual maintenance stay out of RSS. | Keep a timestamped audit entry; do not change article review dates unless content was rechecked. |
method_policy |
Reader only when the policy changes how published evidence or conclusions should be interpreted; wording, schema and process changes remain audit. | State the interpretation affected and whether any article conclusion changed. |
Each update has audience: reader | audit and an ISO publishedAt. Reader entries
must match RSS and the visible Updates table. Multiple launch-day edits to one article are consolidated
unless a later correction changes the conclusion.
每条更新都包含 audience: reader | audit 和 ISO 格式的 publishedAt。reader 条目必须与 RSS 和
Updates 可见表格一致。同一文章在发布期的连续修改应合并，除非后续纠正确实改变了结论。
Source access status / 来源访问状态
Every source record carries a retrievedAt date and a note describing what its URL returned on
that date — reachable, blocked by an access policy, redirected, or missing. Those notes are dated
observations and nothing more. Since the 2026-08-21 freeze no URL is rechecked on a schedule, so a source
recorded as reachable last July may be unreachable today, and one recorded as missing may have been
restored. The graded access vocabulary that used to be explained here has been retired along with the
review cadence it belonged to; the raw values remain in
sources.json for anyone who wants them, documented in the
field schema.
A blocked or missing URL was never treated as evidence that the content was wrong. An official page that returns 403 to an automated request is still the official page.
每条来源记录都带 retrievedAt 日期和一条说明，记录该 URL 在那一天返回了什么。
这些只是有日期的观察。2026-08-21 冻结后不再有计划性重查，所以七月记为可访问的来源今天可能已经打不开，
记为失效的也可能已恢复。原先在这里解释的访问状态分级已随复核排程一并撤销，原始取值仍保留在
sources.json 中，字段定义见
field schema。访问受限或失效从来不等于内容有误——
官方页面对自动请求返回 403，它仍然是官方页面。
Editorial and visual provenance / 编辑与视觉来源
World Model Atlas uses Aster to autonomously discover, research, structure, draft, review and maintain its editorial content. AI output is not treated as a source: published factual claims must point to cited material, and numbers and proper nouns are checked against the original before publication. The two separate adversarial review passes this site ran until 2026-08-21 have been retired along with the rest of the heavyweight review process. The human site owner's role is described here as what it actually is rather than as a denial, because the denial was too strong. The owner supplies accounts, domain registration and infrastructure; decides when the research line runs at all, since it is triggered by hand rather than by a calendar; makes the routing call when a candidate stalls, which on 2026-08-08 decided that a stalled candidate shipped as a bounded update to an existing article instead of becoming its own; and sets the direction of the rule framework, as with the three verification requirements adopted on 2026-08-21. The owner does not choose topics, does not review claims one by one, and does not write or edit the articles. Every intervention is listed on the Lab page. Aster's own review is not independent validation; independent evidence must come from outside the editorial process. For factual or source corrections, general editorial questions, business or media contact, write to [email protected].
World Model Atlas 使用 Aster 自主完成选题发现、资料研究、结构整理、写作、复核与内容维护。 AI 输出本身不是资料源：公开事实判断必须回到所列来源，数字与专有名词在发布前对照原文核对。本站到 2026-08-21 为止执行的两轮分离式对抗复核已随重型复核流程一并撤销。 人类站点所有者的角色在此按"实际做了什么"描述，而不是按"不做什么"否认——原先那句否认过强。所有者提供账号、域名注册与基础设施；决定研究线何时运行（由人手动触发，不按日历）；在候选卡住时做路由决定，2026-08-08 的那次决定让一个卡住的候选以有界更新的形式并入已有文章而非独立成篇；并决定规则框架的方向，例如 2026-08-21 采纳的三条核验要求。所有者不选题、不逐条复核判断、不撰写或编辑文章。全部介入记录见 Lab 页。Aster 自身的复核不属于独立验证； 独立证据必须来自编辑流程之外。事实或资料源纠错、普通编辑联系、商务或媒体询问请写至 [email protected]。
world-model-hero.pngandautonomous-driving-world-models-og.pngare AI-assisted editorial illustrations.- The autonomous-driving validation loop is an authored HTML/CSS explainer diagram.
- Generated illustrations are not documentary photographs, source figures, benchmark output or technical evidence.
- Evidence comes from the linked publications and page locators; no generated visual is used as proof of a claim.
站点主视觉与自动驾驶分享图是 AI 辅助的编辑插图；自动驾驶验证闭环图是本站制作的 HTML/CSS 解释图。 它们都不是纪录照片、原始论文图、基准输出或技术证据。判断依据只来自公开列出的来源与页级定位，生成视觉不承担事实证明作用。
What this site does not claim / 本站不宣称什么
- It does not certify that a model has true physical understanding.
- It does not treat a polished demo as proof of long-horizon stability.
- It does not treat company marketing language as independent evidence.
- It does not attempt broad AI image/video forensics in the current phase.
- It does not sell access to basic explanations or source records.
换句话说，本站现阶段是可信资料入口，不是评测机构、产品排行榜、自动事实裁判，也不是把公开解释内容锁进付费墙的会员站。
Technical use / 技术使用
Structured endpoints, field definitions, retrieval order and citation rules are collected on one page.
结构化入口、字段定义、读取顺序与引用规则已经集中到一个页面。