How it works
Custody first. Intelligence after.
The system has two halves: one that cannot wait — saving what is published before it changes — and one that can be built calmly on top of what was saved.
The chain
Listen
Collectors sweep the 28 state outlets — RSS where it exists, sitemap or front page where it doesn't — detecting every new URL minutes after publication.
Capture
For each article the full HTML, extracted text, metadata and a cryptographic fingerprint of the body are stored. The copy is immutable: nothing is ever overwritten.
Re-check
Every URL is revisited on a decaying schedule — hourly at first, wider later. If the body fingerprint changes, a new version is stored beside the old one. If the page disappears, the absence is recorded too.
Diff
The engine compares only the article body — not the ads or sidebars, which change on their own and would produce false positives — and classifies the change: minor correction, altered figure, deleted paragraph.
Extract
On ingestion each article is processed once: dated commitments, figures, officials and organizations are identified and stored as structured records pointing at their frozen copy.
Answer
A user's question queries those records, not a pile of loose text. The language model only writes what the database returns: the archive decides what is true; the AI only decides how to say it.
Rules the system never breaks
Citation or silence
No claim without a link to a frozen copy. If there is no record, the answer says so.
Verbatim, never paraphrased
Quotes appear as published, with their date and outlet.
Declared scope
Every answer states how many articles and what period it was built from, and where the gaps are.
Open source
Software and datasets are published under an open license: anyone can verify or replicate the archive.
What we don't do. CubanPress does not interpret or judge. It shows the contradiction, the edit and the silence with their date and source; the conclusion belongs to the reader. That restraint is exactly what makes its material quotable anywhere.