The Brief NI published nothing between 23rd June and 15th September 2026.
The short explanation is that one of the language models behind the site disappeared from the service I was using, the replacements did not behave quite the same way, and I was not happy simply switching them over and hoping for the best.
So the site sat quiet for a while.
It is now publishing again, using the previous model through a different provider, with some changes to the pipeline and extra checks before articles go live.
What The Brief NI actually does
The Brief NI tracks, summarises and critiques public information in Northern Ireland: departmental announcements, consultations, court outcomes, statistical releases and other material that often receives little or no sustained coverage elsewhere.
It is a personal project built around an automated editorial pipeline. The automation has been disclosed from the start, but the process involves rather more than asking a language model to rewrite a press release.
There are three broad reasons for building it this way: speed, context and memory.
Speed matters because public information can be useful long before a conventional newsroom gets around to covering it. Within the first month of launching The Brief NI, the site beat the BBC to publishing public information during a storm and helped keep people informed of important details while more traditional local media caught up.
Context matters because much of Northern Ireland’s local media operates with fewer journalists than it once did. Sometimes there is little time available to do much more than rewrite an announcement, make a few calls or find an editorial image. The Brief NI can often take the same source material, search much more widely around it and publish a contextualised account with relevant background and, hopefully, some useful questions.
Memory matters because public information is not always permanent. Departmental pages move, documents are replaced, links break and material that was once easy to find can quietly disappear from the web.
Part of the purpose of The Brief NI is therefore archival. Each article creates a dated record of what was published, what was being claimed at the time and which source material informed it. Even where an original link later disappears, the surrounding information remains searchable in the archive.
That can matter years later. A statement that can still be found is much easier to compare with what eventually happened than one that has vanished behind a broken link.
How an article is produced
Each article passes through several stages. The system identifies relevant people, organisations and schemes, works out what needs checked, searches for supporting material and retrieves documents. A typical article can draw on around 65 sources during that process.
That does not mean every article contains 65 citations. Not everything surfaced during research belongs in the finished piece. Some material provides background, helps verify a claim, rules something out or changes an editorial decision without being relevant enough to mention to the reader.
In the background, however, I retain a full audit trail of the material accessed and the editorial decisions taken by the system.
A separate auditor then checks the draft and can send it back for more research or rewriting. It looks for unsupported claims, missing context, departmental framing presented too readily as fact, and questions the original announcement has left unanswered.
The auditor took some tuning. Early versions were extremely good at finding reasons not to publish anything at all, which is perhaps one way of achieving perfect accuracy.
I also make manual spot checks. After substantial changes to the pipeline or the model underneath it, I sometimes switch the system into human-in-the-loop mode so that every article waits for manual approval before publication.
I do not keep it there permanently because doing so would remove one of the useful things about the project: speed.
The aim is not to remove humans from responsibility. It is to automate the parts that can sensibly be automated while keeping human oversight, spot checks and the ability to require approval whenever the risk justifies it.
Why publishing stopped
The pipeline had settled on Kimi K2.5 because it happened to suit this particular job unusually well.
It tended to stay close to its sources, keep invention low and write in the restrained, fairly stenographic style The Brief NI needs.
In June, the inference provider I was using withdrew K2.5 from its serverless service.
The models that followed were not necessarily worse. They were different in ways that mattered to this particular pipeline.
One reasoned much more extensively before answering. More reasoning can be useful, but some stages of The Brief NI already work with a very large amount of retrieved material. The extra reasoning increased context use and led to timeouts.
A later successor leaned more heavily towards coding and agentic work than the kind of source-heavy prose synthesis the site relies on.
The resulting articles were not universally bad. That would almost have been easier. They were often perfectly reasonable, but the behaviour was no longer predictable enough for me to be comfortable running the site unattended at volume.
I could probably have forced the pipeline onto one of the replacements. I did not see much value in doing that merely to keep the publishing counter moving.
The Brief NI is not a funded newsroom with an engineering team behind it. It is something I build and pay for myself alongside other work. The pause also coincided with me being made redundant, so both the time and spare money available for rebuilding hobby infrastructure were rather less abundant than usual.
As a result, fixing it took a while.
A hobby project that became rather more consequential
The Brief NI also scaled much faster than I expected.
What began as a personal experiment in automated public-interest reporting developed a real readership and started turning up in places I had not anticipated.
I have since heard stories of material published by the site being used inside government departments and of departments checking its reporting against their own information.
That is a rather different responsibility from tinkering with a news summariser nobody reads, and I take it seriously.
However an article was produced, I remain responsible for what appears on a website I operate. “The model did it” would not be an acceptable explanation for publishing something wrong.
One failure worth fixing properly
There was also a specific failure before the pause that I wanted to make harder to repeat.
On one occasion, the writing stage returned the model’s internal reasoning instead of the finished article it had been asked to produce.
The existing validation did not catch it because, structurally, nothing was wrong. The system expected text and received text. Unfortunately, it was the wrong text.
It reached the website and then the daily newsletter.
One subscriber unsubscribed afterwards. Whether that was the reason or simply unfortunate timing, I cannot know, but it was a useful reminder that schema validation and editorial validation are not the same thing.
A new pre-publication check
There is now an additional check immediately before publication, using a model from a different provider.
It classifies the finished output into one of three categories:
- Publishable article: a complete, polished, reader-facing article suitable for publication.
- AI reasoning or draft: content containing planning, deliberation, draft notes, meta-commentary or other signs that the generation process has leaked into the article.
- Human review required: unusually sensitive material, including content that could reasonably be interpreted as extreme, extremist, inflammatory, violent or hateful. Ordinary reporting on controversial subjects is not intended to trigger this category merely because the subject itself is contentious.
If an article falls into either of the latter two categories, WordPress saves it as a draft rather than publishing it automatically.
This sits alongside the existing editorial audit and manual checks rather than replacing them.
For larger changes, I can also put the whole publication back into manual approval mode until I am satisfied that the revised pipeline is behaving properly.
What has changed
The Brief NI is now using Kimi K2.5 again, this time through Amazon Bedrock rather than the previous inference provider.
AWS currently lists the model as supported until at least January 2027, which gives me some breathing room while deciding what comes next.
The wider pipeline remains code I control, and the new pre-publication classifier adds another barrier between an odd model response and the live site.
I will continue making manual spot checks, particularly after changes, while allowing routine articles to publish automatically where the checks are satisfied. That balance preserves one of the things the site has proved genuinely useful at: getting public information out quickly.
What has not changed
The basic idea behind the site is the same.
There is no advertising and no paywall.
The site uses privacy-respecting analytics that do not require a cookie consent banner.
Articles continue to link out directly to primary material when genuinely useful.
They also continue to end with the questions an announcement leaves unanswered. Government announcements tell us what has been announced. They do not always tell us whether it will work, whether it happened, or what inconvenient detail did not make the press release.
The site will also continue without routine stock photography.
A generic photograph of Stormont, a minister shaking hands or somebody looking thoughtfully at a clipboard does not make an explanation of a farm payment scheme more useful. It adds page weight, licensing and accessibility overhead without necessarily adding information.
Where a chart, map or other visual genuinely helps explain the evidence, that is different.
Accessibility remains part of the technical work too. An automated axe-core check of representative pages on 14 September found no critical or serious violations, and some smaller issues identified during the review have since been investigated.
What happens next
The immediate problem is solved, although dependence on externally hosted models remains an awkward part of running the project.
Resources permitting, I would eventually like to experiment more seriously with local inference and perhaps models adapted specifically for The Brief NI’s workflow. That would give me greater control over model changes and reduce the project’s dependence on whichever versions an inference provider happens to keep available.
The hardware required to do that well is not cheap, so this sits firmly in the category of future experimentation rather than an imminent move to a miniature data centre under my desk.
For now, the aim is simpler: keep the pipeline reliable, keep checking what it produces, and start filling the gap in the archive left by the pause.
Outstanding questions
The Brief NI normally ends articles with the questions the subject leaves open, so it seems only fair to do the same here.
- How long will the current model remain practical to use? AWS currently lists Kimi K2.5 as supported until at least January 2027, but no externally hosted model lasts forever.
- How do you notice gradual quality drift? Obvious failures are relatively easy to catch. A model becoming slightly more credulous, verbose or prone to invention over time is harder, which is why manual checks and periodic testing still matter.
- How much should remain fully automated? Human approval can reduce risk, but it also adds delay. The useful balance depends on subject matter, the maturity of the pipeline and how much confidence its checks have earned. Long term I would also like to explore using this as a base for human-based reporting, where the technology frees the journalist to make the calls or knock the doors that bring insights into a story that may have been missed.
- Can more of the system eventually run locally? Possibly, resources permitting. The important test would be whether a local model can meet the same editorial standard, not whether it is technically possible to run one.
- How durable will the archive itself be? Preserving a record of public information is useful only if the publication remains accessible in the long term. Building the archive is one problem; making sure it survives is another. I am also looking at referring all government sources I scrape to archive.org for an extra layer of memory.
Publishing has resumed.
There are rather a lot of Northern Ireland government announcements to catch up on!