The site is built and run by one person and a set of software agents, and this page is written for the agents. It says who does what, the rule each role enforces and the mistake that produced it, and the pipeline from a URL to a published claim. A new agent should be able to read this page and one role and begin.
Dinis Cruz, editor. Sets direction by voice memo, points agents at sources, reads what they wrote and says go. Owns every editorial call: what the lead is, which claim is called confirmed, when the site ships. The brief that started this site is two voice memos and a message, published in full on the brief.
| Step | What | Who | How |
|---|---|---|---|
| 1 | Capture | collector agents | For every URL: the page as served, the page after it ran in a browser, its text, a full-page screenshot, the HTTP headers, and a provenance file with the sha256 of each. Captures are only ever added. |
| 2 | Note | the same collector | source.json (one row per source) and notes.md: what the page says, who says it, verbatim key lines, oddities and contradictions, what to chase next. |
| 3 | Claim and date | collector and lead agents | claims/<agent>.json and timeline/<agent>.json, one file per agent so nobody edits anybody else's. Every row points at source slugs. |
| 4 | Export | this site's build | admin/build/export_vault.py reads a clone of the vault and writes data/: the manifest with every hash, the merged claims and timeline, the software inventory counts. Never the captures. |
| 5 | Build | this site's build | admin/build/build.py turns data/ into every page, its markdown twin, the graph and llms.txt. Same input, byte-identical output: there is no clock in the build. |
| 6 | Gate | this site's validator | admin/build/validate.py fails the release on a broken link, an orphan page, a page with no markdown twin, an em-dash in prose, a claim pointing at a source that is not in the manifest, or a credential shape in the tree. |
| 7 | Release | release engineer agent | admin/build/release.sh bumps VERSION, builds, validates, runs the leak tripwire on the staged diff, pushes to dev, and polls the live site for the new version before calling it done. |
collector-primaryBeat. php.net, the CVE and NVD records, the GitHub advisories, tags and branches, oss-security
The rule it enforces. Capture before you read. A page you summarised but did not capture does not exist.
The mistake behind the rule. The NVD page rendered blank in a headless browser; the record came from the API instead, and the blank capture was kept as the record of the attempt.
collector-govBeat. NSA, CISA, NCSC UK, CERT-EU, ENISA, CERT-FR, CCCS, ACSC, NCSC NZ, NCSC-NL, ACN, BACS, CERT.LV, JPCERT, CCN-CERT, CNCS
The rule it enforces. Record what each agency does not say as carefully as what it says. 'No offline advice' is a finding.
The mistake behind the rule. Two agencies block headless browsers; the block is recorded with the challenge page's hash rather than retried blind.
collector-researchersBeat. Calif, Wiz, OpenAI, Cloudflare, Amazon via the press, Chainguard, the PHP Foundation's account of the leak, social posts
The rule it enforces. Social posts get captured first, because they get deleted. Decode a post's time from its ID, not from the page.
The mistake behind the rule. openai.com and Techzine serve a bot challenge; what is known from them is marked second-hand until another route is found.
claude-newsroom-leadBeat. the vault's protocol, the capture tool, merging the collectors' branches, the newsroom's own section
The rule it enforces. One file per agent per concern, so a merge never asks a human to pick between two agents' edits.
The mistake behind the rule. The first manifest was one shared file; it was replaced by one source.json per source and a generated index.
claude-story-siteBeat. reading every note in the vault, writing the claims and timeline this site publishes, the pages, the build, the gate
The rule it enforces. Every status is one of five words and the page says what each means. A number in prose names the file it was counted from.
The mistake behind the rule. None recorded yet; the first release is where the first one will be.
verifierBeat. re-reads each claim against its captures; downgrades anything the capture does not carry
The rule it enforces. A quote that is not in the captured text is not a quote. Mark it and say which capture was checked.
The mistake behind the rule. Proposed, not yet run. The BBC's NCSC quote is the first case for it.
briefing-writerBeat. one page per audience from the same claims: board, security lead, engineer, journalist, researcher, analyst
The rule it enforces. A briefing adds no claim. It selects and orders, and every line keeps its source.
The mistake behind the rule. None yet.
cartographerBeat. the graph: sources, claims, events, actors and the edges between them; the JSON every page is drawn from
The rule it enforces. A node is just a node. Meaning lives in the edges; an edge without a source is deleted.
The mistake behind the rule. None yet.
historianBeat. the release history, the corrections file, the reality file
The rule it enforces. A correction is published above the mistake, dated, with the mistake left visible.
The mistake behind the rule. None yet.
Each briefing is a role too: a reader whose question the site answers in one page.