Independent · English first, Chinese sources named
The difference is method, not volume. Every claim here carries an evidence tier, conflicting numbers stay side by side, and corrections are published with a date instead of quietly edited away.
Standing rule Official document first. Wire report second. Aggregation is a map to the original, never the citation. If I cannot reach the first, the piece says so.
01
Five tiers, one definition each. The tier is shown at the claim, not buried in a footer.
A document published by the organisation making the claim: newsroom post, filing, standards document, model card, pricing page. I read the page itself, not a press summary of it.
A named reporter at an outlet that did the reporting, with a byline, direct quotes or original observation. Still a second-hand account: treated as a report, not as proof.
A rewrite of someone else’s reporting, a roundup with no named author, or a summary that does not link the source. Used to find the original; never used as the citation.
Benchmarks, usage figures, capacity and reliability claims published by the company about its own product. Labelled at the claim itself. Vendor-reported is not a synonym for false; it is not independent.
Single-source leaks, screenshots, “people familiar with the matter”. Written as an open question with the claim attributed. Never stated in the headline as if it happened.
When credible sources disagree
If two credible sources give different numbers, I do not split the difference or pick the friendlier one. I print both, with each date and definition, then add one line on what would settle it: a filing, a later document, or a direct measurement. If nothing settles it yet, the label is open — not a midpoint dressed up as a fact.
A benchmark split is a data point about the benchmark, not just about the model. Who ran it, on which version, with which prompt and decoding settings.
02
Three recent pieces, published in the AI Weekly archive. Chinese-language sources are named under each card.
110 AgiBot Expedition A3 units completed a 5:43 routine in Jinan with no reported failure, and a Guinness adjudicator certified the record. The checkable claim is quieter: these were production units, not specially prepared demo machines. The reports do not define “zero failure”.
中文原文：科技日报、上海证券报、证券时报对 10 月 24 日济南 110 台远征 A3 群舞的报道。
Sources: three independent Chinese outlets, not a third-party audit.
The 5,000th personal robot came off the line about a month after launch, at 2.5 minutes per unit and more than 200 a day. The ramp is real. Sell-through, return rate and retention are not published, and the usage figures are vendor numbers with no disclosed method.
中文原文：上海证券报、证券时报、IT 之家对启元机器人 5,000 台下线及 200 台/日产能的报道。
Sources: production figures reported by three outlets; usage figures Tier D.
M3.1 Flash weights are downloadable: 428B total, about 23B active per token, 1M-token context, API near $0.15/$0.60 per million tokens. The usage numbers — 56.7 trillion tokens in ten days, 100 trillion within 72 hours — are media reports, not independently audited.
中文原文：MiniMax 官方新闻；新浪财经对开放权重与定价的报道。
Sources: official product page for specs; media for usage claims.
Also in the archive: the UK’s AI Biosecurity Lab · all four articles and the issue
Watching next
03
Public, dated, and kept. A correction that only lives in a diff is not a correction.
Entry 01 · 25 October 2026
zh-TW) as a supported locale for the new Siri. The prose named five languages; the table listed more. We had quoted the sentence and skipped the table beside it.zh-TW) appears in the locale table.” The original sentence stays visible above the note. The headline was not changed, because the headline did not make the claim.zh-CN) is in the first wave. A table entry for zh-TW is not an answer to that, and we are not treating it as one.What I changed in my process: check the document’s tables and locale lists before writing a sentence about what a release “does not mention”. Absence is a claim, and it needs the same reading as presence.
04
Eight checks, each one an action rather than a principle. The full page has the working version.
Read two official pages line by line — the announcement and the documentation — and say which one a claim comes from.
Every number gets a source and a date. No date, no number.
Label vendor-reported benchmarks at the claim. Not independent is not the same as wrong.
Separate preprint from peer review, and name the venue and status.
Print conflicting numbers side by side; never average them into a midpoint.
Read the Chinese original for Chinese hardware and model releases; do not cite an English rewrite.
Treat a promise as a promise: “weights by 4 November” is a date to check, not a shipment.
Publish corrections with a date: the original claim, what was wrong, what changed, what is still open.
Subscribe
Full text, newest first, every claim with its source line. No tracking pixels, no third-party scripts, no sponsorship.
https://hhfa008.github.io/ai-resources/feed.xml
Corrections and tips: superapp.devin@gmail.com. Corrections are logged on this page with a date.
Elsewhere