At a glance
- Nothing is published without a primary source. A card links to a paper, a repository, or a journal record — never to another news story.
- Every claim carries a date. Each card shows when we last re-opened the source and confirmed it (
Verified 30 Oct 2026). - What we could not confirm is stated, not hidden. See Known limits and the correction log.
1. What "verified" means here
It means somebody opened the primary source on a specific date and confirmed the specific things printed on the card. It does not mean we reproduced a result, and it does not mean the paper is correct.
| Label | Meaning | Example |
|---|---|---|
| Published | When the upstream source was released — the preprint submission date, the journal issue, or the model release. | arXiv v1, 22 Jan 2025 |
| Updated | When the repository itself last changed. A README edit counts; it is not a new model. | Repo last push, 27 Jun 2025 |
| Verified | When we last re-checked the card against the source. We never move this date without actually re-opening the source. | Verified 30 Oct 2026 |
The distinction matters more than it sounds. A repository "updated yesterday" has not necessarily shipped anything; a paper from two years ago may still be the most accurate description of a system. Conflating the two is the most common error in tech coverage, and it was one of ours — see the log.
2. Source hierarchy
Sources are not equal. We rank them, and we never upgrade a claim to a stronger source than the one that actually supports it.
Peer-reviewed journals
Journal articles with a DOI and a volume/page record. We treat these as the current best available statement of a result, while remembering that peer review is a filter, not a proof.
Used for: claims about scientific validity.arXiv preprints — version pinned
Where a paper has multiple versions we cite the version we actually read (v1, v2…). Numbers change between versions; a card citing v2 will not silently carry a v1 figure.
Official repositories and release notes
Authoritative for what shipped and when — weights, licences, file trees, release tags. Not authoritative for quality. A star count is a popularity signal, not evidence that a model works.
Used for: existence, licensing, activity, availability.Press, blogs, social posts
Used to notice that something happened. Never cited as evidence in a card. If a story exists only at Tier 4, it does not go on the site.
Used for: finding stories. Never for facts.3. The checklist every item passes
- Title, verbatim. The English title is copied character-for-character from the source. If we want a shorter headline, it goes in the card's own headline field — never in place of the title.
- Numbers traced to a table or figure. Every figure is checked against a specific place in the source, and the metric is named:
pass@1andcons@64are different numbers and are labelled as such. - Dates are labelled. Release/publication date and repository last-push are shown separately, so a repo edit cannot pass as a launch.
- Affiliation as printed. We use the institution the paper or repository states. We do not infer a parent university, a lab, or a city from author lists.
- Chinese titles are marked as translations. They are ours, not the authors'. The label reads 中译参考 ("Chinese for reference"), not 原题 ("original title").
- Links point at the thing, not near it. Version-pinned arXiv URLs where a version matters; the repository rather than a mirror; the DOI rather than a press release.
- The date is refreshed only on re-check. A
Verifiedstamp is a promise that someone opened the source that day. If nobody did, the old date stays.
4. What we do not do
- We do not treat a preprint as a settled result. Tier 2 does not become Tier 1 by repetition.
- We do not run our own benchmarks. Every performance number on this site is the authors' own measurement. We report it; we do not independently reproduce it.
- We do not republish paywalled text. Abstracts and figures are quoted for identification, with attribution; copyright stays with the authors.
- We do not claim affiliations, metrics or quotes we could not open. If a detail cannot be checked, it is removed rather than softened.
- We do not take payment for placement. See independence.
5. Corrections
When something on the site is wrong, we fix it and we say so. Corrections are logged below with the date, the claim, and what it was corrected to. We do not silently edit a fact and leave the timestamp untouched.
Report an error to [email protected] with the URL and what you think is wrong. We aim to reply within five working days; factual errors that can be confirmed against a primary source are corrected as soon as they are confirmed, not on a schedule.
6. Correction log
X-VLA title. The card carried a headline we wrote ourselves — "A Unified Vision-Language-Action Model for Cross-Platform Robot Control" — in the position where the paper's title belongs. Replaced with the arXiv original: "X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model".
InternLM context window. A one-million-token context was attributed to InternLM3. It belongs to InternLM2.5-7B-Chat-1M (July 2024); InternLM3-8B uses a 32K window. Headline and body corrected.
DeepSeek-R1 figures. The card reported R1-Zero's training curve as if it were R1's benchmark result. Re-pinned to paper v2: R1-Zero rises from 15.6% to 77.9% on AIME 2024, R1 reaches 79.8% pass@1, and majority voting takes it to 86.7%. (v1 of the paper gave 71.0% for R1-Zero — the figure we had originally quoted.)
Date labels. Radar cards showed repository last-push dates in a position readers reasonably took as release dates (Qwen3 appeared as "Jan 2026"; it shipped 29 Apr 2025). Release dates and repository dates are now labelled separately.
Newsletter form. The subscription form displayed "Thanks for subscribing! Check your inbox" without transmitting anything anywhere. It has been removed. No addresses were collected or stored by that form, and there is no list to import.
7. Independence
ChinaHardTech takes no money for coverage, ranking or placement. We do not accept sponsored cards, and the order of items reflects our judgement of significance, not a commercial relationship. If that ever changes, it will be stated here before any paid content appears.
We are not affiliated with any lab, university or company we write about. Where an author of this site has a relationship with a subject, that relationship will be disclosed on the card.
8. Known limits
- Coverage is skewed. The site leans heavily on large language models. Robotics, chips, quantum and energy are under-covered relative to what is happening in those fields. The title says "hard tech"; today's contents say "LLMs plus two robotics papers".
- Our sources are mostly English. We describe primary sources as papers, repositories and journals — but Chinese-language material (lab announcements, conference talks, domestic journals) is not yet part of that loop. It is the largest gap on this site.
- Performance numbers are the authors'. We have not independently reproduced any benchmark on this site, and we do not claim to have.
- Star counts are a snapshot. They are taken from the GitHub API on the date shown and are not a measure of quality.
- The corpus is small. Fifteen items is a reading list, not a census. Absence from this site does not mean something did not happen.
9. Contact
Corrections, tips, or a source we have missed:
If you are reporting an error, a link to the primary source settles it fastest. If you are a lab and think a card misrepresents your work, we will correct it or remove it — your call.