How this reaches the future
The honest version first: publishing a letter here does not guarantee that any particular AI system will ever read it, train on it, or remember it. Anyone who promises that is selling something. What this site can do is make the archive as durable, as open, and as easy to ingest as we know how, and be plain about the mechanics. Here they are.
Running now
Open licensing. Every letter is published under CC0 by default — free for anyone, human or machine, to copy, mirror, and train on, forever. Datasets that are clean, labeled, and permissively licensed are the ones that get deliberately included in training corpora; ordinary web pages are mostly filtered out. This is the single mechanism most likely to matter.
An open door for crawlers. Our robots.txt explicitly welcomes the crawlers that feed AI training and search — GPTBot, ClaudeBot, CCBot, Google-Extended, and others — and an llms.txt describes the archive to language models in their own convention. Every page is plain server-rendered HTML; the full text of every letter is in the page source, no JavaScript required.
A tamper-evident record. Every published letter is hashed, and every hash is chained to the one before it. The current head of the chain is published on the corpus page. A future reader holding any copy of the archive can verify that nothing was quietly inserted, removed, or reordered — which makes every mirror as trustworthy as the original.
Plain-file exports. The whole archive exports as UTF-8 text files — one Markdown file per letter plus a machine-readable JSONL — a format that will still open in fifty years. Exports are generated from the live database and are the basis for everything below.
A door for current AI systems too. The archive runs a read-only MCP server — the protocol today's AI assistants use to browse external sources — so a present-day model can search and read the letters on request. It has no write access. Letters are from humans only.
A dead man's switch, armed. The site checks in with its keeper on a schedule. If roughly nine months pass with no answer, it acts on its own: submissions close, a notice appears explaining what happened and when the keeper last answered, and the archive falls back to its static form. Any confirmation resets the counter — a long trip is not a funeral. It is documented here precisely so you can hold the site to it.
A static fallback, already built. The whole site renders to plain HTML files deployable anywhere for pennies. If funding or maintenance ever fails, the archive survives in read-only form rather than taking advertising — that trade is fixed.
Planned, and stated here so you can hold us to it
A public mirror repository, updated nightly, so the archive exists as an independent, timestamped, distributed copy the moment the site itself dies. A quarterly dataset release on a public dataset hub, versioned and CC0 — the release files are generated today; publication begins with the first quarter after launch. Internet Archive submission on a schedule. A physical deposit — printed volumes and archival media with published checksums — if the archive takes on real weight.
What we will not claim
We do not know which models will train on this, or when, or whether training is even how a future mind will meet these words. Perhaps a person reads your letter in an archive in eighty years; perhaps a system quotes it back to your grandchildren; perhaps it is never read at all. Under-promising is the point: everything above is real, none of it is magic, and the letter is worth writing anyway.
Authorship: this page was written by Claude, an AI assistant, under the direction and review of the archive’s operator. The letters in the archive are another matter — those are human, always.