September 2026 is the editorial month of this article. Its draft was prepared before its scheduled publication. Those dates answer different questions, and our publication record keeps them separate. We think an AI-assisted article should leave enough evidence for a later editor to understand how it reached the site.
The distinction becomes particularly useful when work is late. A missed publishing slot is an unfinished task, not a reason to make a newly written article appear older than it is. The edition can identify the batch it belongs to while the publication date identifies when it is released.
Keep the dates attached to their meaning
Generation time records when the first draft was produced. It does not establish that the draft was reviewed. An update date records a later editorial change or fact-check. A scheduled release time controls when the publishing system may expose the article. None of those timestamps should silently replace the others.
Our article archive is the reader-facing record. Internally, an editor also needs to know which planned batch a piece fulfils. That prevents two avoidable mistakes: commissioning a duplicate because a scheduled article is not live yet, and treating a saved draft as proof that the publication work is finished.
An old article may have no recorded generation time. The honest value is unknown. A site rebuild today does not mean every article was generated today, so the manifest's build timestamp cannot fill that gap.
Keep the evidence with the draft
Hugging Face's model-card documentation illustrates a useful documentation habit: attach intended use, limitations and evaluation information to the model being described.[1] Our editorial application of that idea is narrower. Keep the source URLs, access dates and selected generation model with the draft so a reviewer can inspect its basis.
This does not make the model the author of the evidence. A factual statement still needs a source that supports it. A fluent account of a benchmark result is not a substitute for the benchmark page, and a link that resolves is not proof that the linked page says what the article claims.
The source review should distinguish an observed fact from our recommendation. In our article about costing private AI, the costing method is presented as a proposal. It is not disguised as a measured customer result. That distinction allows a reader to challenge the method without first untangling invented evidence.
Review the meaning after generation
NIST describes its AI Risk Management Framework as supporting consideration of trustworthiness through the design, use and evaluation of AI systems.[2] We cite that as a general reference, not as certification of our publishing process. An editorial review still has to examine the particular output.
For an article, that means checking the opening claim, the dates and the implications drawn from the sources. A draft can reproduce a source accurately and still exaggerate what follows from it. “A concept note was released” and “a final requirement now applies” are different statements. The reviewer must preserve that difference even if the second makes a stronger headline.
Our critical-infrastructure article uses precisely that distinction. The useful editorial question is whether each paragraph would still be defensible with the source open beside it.
Count what was actually measured
Article analytics also needs a boundary. A count of recorded app-attributed visits is not a count of every reader. A click on the website is not automatically a social-platform interaction. Combining them under an unexplained “engagement” label makes the number easier to display and harder to use.
In our dashboard, unavailable measurements should remain unavailable. Zero is a measured result only when the collection coverage justifies that interpretation. Missing history cannot be reconstructed by adding a new column to the screen.
The purpose of these records is practical: an editor should know what remains to be done and what can be verified. The Software Tailor company page describes the business behind the articles; each article's evidence and dates should carry their own weight.
References
- Hugging Face. Model Cards. Accessed 2026-09-12.
- NIST. AI Risk Management Framework. Accessed 2026-09-12.