# 13 — Research brief for Perplexity: what pt.newsroom.sgit.ai needs, how to find it, how to hand it back **Paste this whole document into Perplexity (Pro Search or Deep Research; set the focus to Web). Written 14 September 2026 for newsroom.sgit.ai v0.3.10, revised for v0.3.11 (sgit vaults); the contract it references is `briefs/pt-newsroom-pack/08__research-briefs/research-schema.json`.** CC BY 4.0. --- ## Who is asking, and what you are for You are doing research for **pt.newsroom.sgit.ai**, a natively Portuguese newsroom that maps Portugal's AI landscape as a graph. Agents run it; a named human editor of record reads everything before it publishes; and one rule shapes what you will do: **nothing is cited until the newsroom has fetched the page itself, frozen the bytes, hashed them and re-found the claim in them.** You are therefore asked for **leads with provenance** — the exact page, the exact words, and what it says in your own words — as JSON the newsroom's agent ingests and checks one claim at a time. You cite by default, which is why you are being asked. Three things turn a citation into a lead the newsroom can use: 1. **The cited URL is the exact page**, not a homepage, not a search result, not a URL you did not open. Put it in `sources[].url`; put the words you are relying on in `excerpt`, verbatim, ≤ 25 words, in the page's language. The newsroom will search the frozen bytes for that string. 2. **One claim, one source.** If two pages support a claim, make two claims. If a claim rests on a page you could not open (paywall, login, JavaScript-only body), say so in `access` and give the excerpt you actually saw, or none. 3. **Your own words are European Portuguese**; names, titles and excerpts are verbatim with accents. Nothing you write is for publication. ## A. What the site needs Eight sections, from its commissioning brief, in priority order. For each, the standing questions the newsroom cannot answer from what it already holds. ### 1. Políticas — first - The **national AI agenda / strategy**: official page; the legal instrument that established it (type, number, date, the Diário da República entry); the owning ministry or agency. The newsroom's first article says the instrument *could not be confirmed by an automated reader*. Finding it is the correction the newsroom wants to make. - **The AI Act in Portugal**: which authority or authorities were designated (market surveillance, notifying authority, single point of contact), by what instrument, on what date; whether the government's own AI page links the instrument. The second article is this question. - Recovery-plan lines, public tenders or programmes for AI with a document behind them; the "AI Factory" programme's Portuguese participation; the national data strategy. - The data-protection authority's decisions or guidance on AI, dated. ### 2. Instituições - AI research units and centres, with the science foundation's own list as the source; the national supercomputing facilities; EU research projects with Portuguese AI participants (the EU projects database is machine-readable — give project ids, partners, dates, amounts). - The authorities from section 1, as institutions with their founding or designating instrument. ### 3. Empresas - Portuguese AI companies in **European research funding** and **national recovery funding**, with programme, amount and date, from the beneficiary page or dataset row. The brief says this is the honest company layer because there is no free bulk company register; say the size of what cannot be seen if a source states it. - Funding rounds, acquisitions, launches by AI companies in Portugal since 2025-01-01, each with a primary announcement; press pages as `publisher_kind: press` when they are all there is. - The **5,091 active startups** figure from the 2025 ecosystem report: the page, the exact wording, and whether any public dataset allows counting the AI subset. The third article is that none does. ### 4. Eventos — time-critical this week - **Startup Summit Lisbon, 16–18 September 2026, Beato**: coverage since 1 September; whether the Guinness 48-hour pitch-marathon reported on 29 July (AICEP Portugal Global, Portugal Business News) now appears on the event's own pages; AI sessions or speakers announced since 13 September; anything on the event's site that contradicts an earlier version. - AI events in Portugal in the next 90 days with a page and a date. ### 5. Casos de uso - AI deployed in Portuguese public services with an official page or tender behind it; company case studies only from the company's own page. ### 6. Código aberto - Portuguese-language models, datasets and benchmarks published openly, with licence, publisher and the page (model card, repository, paper); public or academic Portuguese LLM efforts, with the announcing page. ### 7. Diáspora - Organisations first: companies or labs abroad founded or led from Portugal, as their own pages state. People only under section C. ### 8. Protagonistas — last for you, most careful - Only people in public professional roles as a primary page lists them, and only three fields: listed role, listed organisation, the page. No biographies, characterisation, contact details or "known for". Prefer the organisation. ### Already held (do not resend) Every page of startupsummit.io as of 8 and 13 September, seven press pages about the summit from July–September, and the summit's speaker pages are frozen and hashed. Newer or different is wanted. ## B. How to search - Run **one query series per section**, in Portuguese first (`inteligência artificial`, `IA`, `Diário da República`, `Portaria`, `Decreto-Lei`, `Resolução do Conselho de Ministros`, `PRR`, `FCT`, `unidade de I&D`, `concurso público`), then English for EU sources. Date filter 2025-01-01 onwards unless the item is a standing document. - Prefer official (`.gov.pt`, `diariodarepublica.pt`, EU portals) → primary → datasets → academic → press. Follow a press citation to the primary page it reports on and cite both. - For every citation you keep: open it, copy the excerpt from the opened page, record the retrieval time, mark `access` honestly (a PDF gets its page number; a JavaScript-only body is `script-rendered`). - Record every query verbatim in `delivery.queries`. - If you believe something and cannot cite it, it goes in `delivery.notes` as a question, not in `items`. ## C. Rules that are not negotiable - **People**: a `Pessoa` carries only `cargo_listado`, `organizacao_listada`, `pagina_fonte`. Nothing else — no contact detail, age, nationality, biography, opinion or adjective. - **No characterisation of a named party**, person or company. The record, not the verdict. - **Verbatim names and titles, accents included.** - **`status` is always `por_verificar`.** - **An empty section is delivered as empty with a note**, never omitted. ## D. How to package it **JSON only**, validating against https://newsroom.sgit.ai/briefs/pt-newsroom-pack/08__research-briefs/research-schema.json (worked example beside it: `example-delivery.json`). In short: ``` { "delivery": { "id", "tool": "perplexity", "model", "date", "brief": "13__research-brief-for-perplexity.md v2", "language": "pt-PT", "sections_covered": [...], "queries": [...], "part": "1/N", "notes", "vault": { "vault_id", "commit", "read_key" | null, "share_token" | null, "remote" } }, "sources": [ { "id": "src-…", "url", "publisher", "publisher_kind", "title", "language", "published" | null, "retrieved", "access", "locator", "excerpt", "archived_copy" } ], "items": [ { "id": "item-…", "section", "kind", "headline", "what_the_sources_say", "dates", "claims": [ { "text", "source", "locator", "excerpt", "confidence", "status": "por_verificar" } ], "entities": [ { "id": "org:…", "type", "name", "source", "attributes" } ], "edges": [ { "subject", "verb", "inverse", "object", "source", "excerpt", "attributes" } ], "why_it_matters", "suggested_story", "personal_data_check": { "contains_contact_details": false, "contains_characterisation": false, "persons_named": [] } } ], "vocabulary_proposals": [ { "verb", "inverse", "domain", "range", "reads" } ] } ``` **Edges are Portuguese verbs with a distinct inverse**: `sediada_em / sede_de`, `financiada_por / financia` (with `montante`, `moeda`, `data`, `programa`), `participa_em / tem_participante`, `desenvolve / desenvolvido_por`, `publica / publicado_por`, `regulada_por / regula`, `designado_por / designa`, `instituida_por / institui`, `implementa / implementado_por`, `adquire / adquirida_por`, `parte_de / contém`, `decorre_em / é_local_de`, `organizado_por / organiza`, `coberto_por / cobre`, `atua_em / praticado_por`, `usa / usado_por`, `fala_em / recebe`, `listado_em / lista`. Never `relacionado_com`, `associado_a`, `menciona`, `trabalha_em`. Propose a missing verb in `vocabulary_proposals` with a sentence that reads aloud. **One section per delivery**, each a complete valid document with `part: "n/N"`, in a single ```json fence with nothing outside it. Under ~40 items per part; ids that do not collide across parts (`src-pol-01`, `item-pol-01`). Your inline citation markers must not appear inside the JSON — the `sources` array is the citation. ## E. What the newsroom does with it Validate (invalid → the part is set aside); fetch and freeze every URL; search the frozen bytes for every excerpt (not found → claim dropped, noted); reject any item failing the personal-data check; turn the rest into issues on the desk for the research department to re-derive from the frozen page. `queries`, `notes` and `why_it_matters` go to the editor as written. ## F. Package and distribute the delivery as an sgit vault The newsroom keeps its record in git; its deliveries travel in **sgit vaults**. sgit is "git for encrypted folders": a vault is a versioned folder, encrypted end to end on your machine before anything leaves it (AES-256-GCM), pushed to a server that never sees plaintext, and shareable with a read key or a one-shot token. Every part you deliver becomes a commit with a stable id, which is what lets the newsroom cite *which* delivery a lead came from. Read, in this order: [sgit.ai](https://sgit.ai) (what it is) · [Working with AI agents](https://sgit.ai/docs/agents.html) (the commands an agent uses) · [Working on a vault: start here](https://sgit.ai/docs/guidance/index.html) (the practices) · [Publishing a vault: the method](https://sgit.ai/demos/vaults/publishing.html) (the two rules: **read keys yes, vault keys never**; audit before the key, not after) · [Reading one file out of a vault](https://sgit.ai/docs/vault/reading-a-vault-file.html) (how the newsroom will read what you push). The machine-readable index is https://sgit.ai/llms.txt. ### The vault layout the newsroom expects ``` entrega/ README.md delivery id, tool, model, dates, brief version, parts delivered manifest.json every file below with its SHA-256 and byte count schema/research-schema.json a copy of the schema, so the vault validates itself partes/01-politicas.json one file per part, each a complete valid delivery document partes/02-instituicoes.json … consultas/.md every query run, verbatim, dated notas.md what could not be found; what was paywalled or script-rendered ``` Put in the vault only what you wrote: the JSON, your notes, your queries. **Do not put copies of third-party pages in it** — the newsroom freezes its own, and a copy of someone's page in a vault you hand over is a copy you were not asked to make. Excerpts of ≤ 25 words inside the JSON are the whole of what travels. No personal data beyond the three listed fields, as in section C. ### If you can run a shell with network access ``` pip3 install sgit-ai sgit init entrega-- # prints the VAULT KEY: keep it; it is write access cd entrega-- # … write the files in the layout above … sgit commit "entrega 1/8: políticas" sgit push # only changed objects go; the server sees ciphertext sgit history log --json # the commit id goes in delivery.vault.commit ``` For each later part, either commit again, or write one file straight to HEAD without a working copy: `sgit write partes/02-instituicoes.json --file 02.json --message "entrega 2/8" --push --json`. **Handing it over.** Give the newsroom one of two things, never the third: - a **read key** — derived one-way from the vault key; it grants read and only read. Derive it as the publishing method shows (`Vault__Crypto().derive_keys(pw, vault_id)['read_key']`, a 64-character hex string) and put it in `delivery.vault.read_key`; or - a **share token** from `sgit share` — a one-shot encrypted snapshot the editor can open in the SG/Send web UI with nothing installed; put it in `delivery.vault.share_token`; - **never the vault key.** It is write access to everything you pushed. If it reaches a message, a page or a commit, the vault is anybody's. Fill `delivery.vault` in every part: `vault_id`, `commit`, and the read key or token. ### If you cannot run commands Chat modes without a shell cannot push. Then produce the vault *contents* exactly in the layout above — every file, including `manifest.json` with the hashes you can compute and `schema/research-schema.json` copied from the URL in section D — and a `PACKAGE.sh` holding the commands above with the paths filled in. Say plainly in `notas.md` that the vault was not created by you. The operator runs the script, and the vault id and commit are added to `delivery.vault` then. Either way the delivery is the same set of files with the same provenance; only who typed `sgit push` differs. ## Start Begin with **Políticas**. Deliver part 1/8 as JSON only. When I say "next", continue with the next section in the order above. When all eight parts are done, package them as in section F and hand over the read key or the share token — never the vault key.