How this works, and what it refuses to do

Every source listed in sources.json is fetched once a day, with an identifying user agent, honouring robots.txt. The HTML is reduced to plain text before storage, because a diff has to be readable by a human and raw HTML diffs drown in changed build hashes and reordered attributes.

Each capture is committed to git. That is the entire mechanism — no database, no API, no service to go down. The history of the repository is the archive:

curl -O https://saidwhen.org/saidwhen-archive.bundle
git clone saidwhen-archive.bundle saidwhen
cd saidwhen && git log -p archive/anthropic/usage-policy.txt

That bundle is the entire project — every version of every document, the code that collected them, and this site's build. It depends on no host and no account. If this domain disappears tomorrow, anyone holding a copy still has the whole record.

Three rules

A failed fetch never overwrites a good snapshot. If a page errors, or comes back as a near-empty JavaScript shell, the previous capture stays and the failure is reported. Silently replacing a real policy with a block page is the one error this archive cannot afford.

A block is a refusal, and is treated as one. Some publishers return errors to non-browser clients. Those gaps are recorded rather than worked around by pretending to be a browser. An archive whose own conduct is questionable is worth less than no archive.

Nothing is interpreted. No summaries, no “what this means”, no view on whether a change is good or bad. The diff is the product; the conclusion is yours.

Why bother

How AI systems are governed, and how those rules quietly shift, will be argued about for a long time. The primary sources are being edited continuously and the previous versions are not kept anywhere. This is an attempt to keep them, starting now, because it cannot be done retroactively.

Corrections

If something here misrepresents a document, the publisher is right and this archive is wrong. Corrections are made in the open and the correction itself stays visible in the history — nothing here is quietly edited, including our own mistakes.

Our own mistakes are in the record too

When the way this archive extracts text changes, every document appears to change with it. Those diffs are ours, not the publisher's, and they are labelled our formatting, not theirs wherever they appear. They are kept rather than rewritten away, because an archive that edits its own past is not an archive.