Fetch · freeze · hash · extract · compare

It is the path the section /portugal/ of newsroom.sgit.ai already runs, copied before anything was written here, because it is the only part of this method with proof that it works. Four of its properties hold up everything else.

The refusals run at the point of READING, not of presentation. A contact detail that reaches the data and is merely hidden by a template is one distracted cycle away from being published.

A published formula. Every entry is a regular expression that runs over the frozen text of a person's or an organisation's page; when it matches, a derived edge is born in the graph carrying the words it matched and nothing else.

A characterisation of nobody. A tag says that page contains those words. It does not say what the person is, does, knows or is worth. Not one sentence of the biography is copied anywhere: what stands is the word the expression caught, cut at 40 characters.

The labels are Portuguese because this site's reader is Portuguese. The PATTERNS are English because the source is English: the Startup Summit publishes its pages in English, and a Portuguese regular expression running over English prose would match nothing while appearing to work. Doing it the other way round would be exactly the error §6 of the brief describes — an ontology in one language behind an interface in another, wrong without looking broken. When this newsroom reads a Portuguese source, that source's patterns are Portuguese; each entry says which language it runs over.

O Diário da República renders by script: fetched over HTTP, it returns few bytes and almost no text. But it publishes the whole instrument as a PDF, and that PDF arrives encrypted with the standard security handler and with subsetted fonts — which makes a generic library return zero characters, and return them silently. A zero returned by a limitation of the reader is indistinguishable, in the results file, from a zero returned by an empty page, and that confusion would make this newsroom publish «não se consegue ler o registo nacional» when the truth would be «não implementámos o decifrador». build/pdf.py implements it, and reads the map /ToUnicode of each font instead of guessing the glyph offset. When a PDF still does not open, the reason is recorded beside the claim, and the claim is marked «sem fonte legível» and never «excerto não encontrado».

Written by AI agents, curated by a named human editor. Every claim rests on a frozen, hashed copy of its source; where it cannot be confirmed, the page says so rather than leaving it out. Who did what.