myfeeds.sgit.ai / Provenance
ai-generated, unreviewed what this means
Disclosure
This site is written by an AI, and reviewed afterwards
Not "AI-assisted". The prose on every page, the two ontologies, the six audience definitions, the extraction of entities from each article and the classifications that follow from them were all produced by a language model — a Claude Code session — and are read by a human after publication rather than before. That is an unusual thing to put at the top of a page instead of at the bottom, and it is the only honest place for it on a site whose entire argument is that automated selection should be checkable.
Who does what
| Part | Made by | Reviewed |
|---|---|---|
| The pages and their argument | Model | After the fact, by Dinis Cruz |
| The article and audience ontologies | Model | After the fact; the open questions are on the review page |
| The six audience role definitions | Model | After the fact. These are the most opinionated thing here and the most worth disagreeing with |
| Entity extraction from each article | Model | Not yet. 3 articles, none reviewed |
| The join formula and its weights | Model | After the fact |
| The source articles themselves | pt.newsroom.sgit.ai | By that newsroom's named human editor of record, before publication |
Note the asymmetry in the last row. The Portuguese newsroom this site reads from has a named human who reads every page before it goes live. This site does not, yet. That gap is on the board and is the single most important thing to fix before anything here runs on a schedule.
The status on every page
Every page on this site carries one of these, at the top, under the breadcrumb:
| Status | Means |
|---|---|
| ai-generated, unreviewed | A model wrote it and nobody has checked it. Today this is every page. |
| ai-generated, human-reviewed | A model wrote it and a named person has read it and let it stand. |
| human-written | A person wrote it. |
A page moves off the default when somebody names themselves against it, which is why the generator holds a map of exceptions rather than a field that has to be remembered. An empty map is the honest state and it is currently empty.
Why this matters more here than elsewhere
This site argues that a recommendation you cannot interrogate is worthless. A site making that argument in prose a model wrote, using an ontology the same model invented, to classify articles the same model extracted, would be an unusually pure example of the problem it describes — unless it says so, in the same place a reader forms their view of whether to trust it.
The three places a model can be wrong here, in order of damage
- Extraction. A model can name an entity the article does not contain. Everything downstream then explains, correctly and traceably, a connection that should not exist — and the provenance trail makes it more convincing, not less. This is the worst failure mode in the whole system and it currently has no automated guard.
- The audience definitions. Six role descriptions written by a model about how real people read. They are plausible, which is exactly the problem: plausible and unchecked is how a stereotype gets encoded as a data structure. These need a practitioner, an investor and a risk owner to read their own entry and say what is wrong.
- The prose. The least dangerous, because a reader can tell. A wrong sentence about how something works is visible in a way that a wrong edge in a graph is not.
What is not AI-generated
- The recovered MVP posts — written by Dinis Cruz in 2025 and reproduced as published.
- The source articles, which belong to the Portuguese newsroom and are linked rather than reproduced.
- The measurements in this repository: counts of pages, words, articles and connections are computed from files, not written.