The method: capture, hash, claim, build, gate

The evidence vault

A private, encrypted sgit vault, opened at 2026-10-22T08:03Z, written to by 5 agents and the editor. It holds 144 sources and 137 captures at commit 200d08971937. Its protocol, in its own README: capture, don't paraphrase; captures are additive, never edited or deleted; one source.json and one notes.md per source, so agents never conflict on a shared manifest; claims and timeline rows in one file per agent; primary sources first, then news, then social posts, which get captured early because they get deleted; a blocked fetch is recorded with the block, so another agent can retry; and no keys or tokens in any file. The vault is private because it holds verbatim third-party pages for the record; this site publishes their hashes, their URLs and short quotations, and never the pages.

Five words for how far a claim can be trusted

• confirmed two or more primary sources agree, or the fact was counted from a captured file
• reported one party says so, and it has not been contradicted
• disputed the sources disagree, and both readings are shown
• contradicted a primary source says otherwise
• unverified even the reporting outlet could not check it
• p/n primary sources among the n cited; a blocked capture is counted and marked

A claim never gets a better word than its sources give it. A verbatim quote is marked as such and attributed to the page it is on; a quote attributed by one outlet to another party stays reported until the other party's page is captured and found to carry it. The one case so far where that check failed, the BBC's NCSC sentence, is on the agencies page.

Deterministic, and why that matters

The build reads data/ and VERSION and writes every page. It has no clock, no random number and no network: run it twice and the output is byte-identical, which the gate checks. That means a page can be reproduced from the data at any version, a diff between two versions is a diff in the data, and a reader who doubts a page can rebuild it. Times on pages are the sources' times; the only date the site adds is the time of the latest capture, which is itself read from the vault.

1. admin/build/export_vault.py --from <vault clone> reads the vault and writes the manifest with every hash, the merged per-agent claims and timeline, the vault's HEAD and the software inventory counts. Never the captures.
2. admin/build/build.py builds every page, its markdown twin, data/graph.json and llms.txt.
3. admin/build/validate.py is the gate: every internal link resolves; every page is linked from somewhere and has a markdown twin and an llms.txt line; every claim and event cites a slug in the manifest; the build is byte-identical when run again; no em-dash in the site's prose; no credential shape anywhere in the tree.
4. admin/build/release.sh bumps the version, builds, validates, runs the leak tripwire over the staged diff, pushes, and polls the live site for the new version before it will say the release is done.

For an agent

Everything on this site is in data/ as JSON and every page has a markdown twin at the same path; llms.txt lists them. To add a source: capture it into the vault with the vault's tool, write its source.json and notes.md, add rows to your own claims/<agent>.json and timeline/<agent>.json, commit with your agent's name, push. To change this site: edit data/, run the build, run the gate. Do not edit a generated page; the build will overwrite it. The vault's claims and timeline files are exported as vault-claims.json and vault-timeline.json; this site's own rows are in claims.json and timeline.json, written by the agent claude-story-site from every note in the vault.

Replicate it for the next incident

Nothing here is specific to PHP. The vault protocol, the capture tool, the five status words, the export, the build and the gate would carry a story about any incident, outage or disclosure. The repository is public under Apache-2.0 and CC BY 4.0; the next story starts by copying it, emptying data/, and opening a vault.