What was published, and when, on the record
Forecasts, claims and writings are published, and then edited. The miss becomes a hit, the date moves, the wording softens, and a reader looking at the page today has no way to know it differs from the page as first published. Screenshots and archives help only if someone took them, and they say nothing about who published what.
The SolutionEtched In Stone is Watchfire AI's public, append-only record. Each application publishes entries into its own record with its own sequence; Fireball's forecasts were the first. Every entry is stored exactly as published, fingerprinted, timestamped by two independent timestamp authorities and anchored to the Bitcoin blockchain, then rendered with its proofs and a downloadable evidence archive. Change one byte and the proofs fail; remove an entry and the sequence shows the gap. Verification needs nothing from Watchfire: the record, the proofs and standard public tools are enough.
Each step produces a recorded artifact before the next begins, and nothing is published until every proof that must exist at publish time exists. What each step produces:
The bytes are fixed before any proof is made, and every proof is about exactly those bytes. Two independent timestamp authorities and the Bitcoin blockchain attest that the record existed by the times shown. Change one byte and verification against the original proofs fails. Entries are appended, never substituted; a correction or an update sits beside what it follows. The entry's id is inside the stamped bytes, so a removed entry leaves a visible gap in its sequence. If a timestamp receipt cannot be obtained, nothing is published. Verification needs nothing from Watchfire: the record, the proofs and standard public tools are enough.
Entries arrive through a private path on Watchfire's staff portal, gated by signed cookies, and reach a small API on AWS Lambda that validates the entry and issues a ticket. An Amazon SQS FIFO queue delivers each job to one serial worker, so publishes never interleave. The worker canonicalizes and fingerprints the entry, obtains the two RFC 3161 tokens and the OpenTimestamps receipt, writes the bytes, proofs and evidence archive to Amazon S3 and the registry row to Amazon DynamoDB, renders the static site and invalidates Amazon CloudFront. An Amazon EventBridge schedule runs the worker hourly to upgrade pending receipts to Bitcoin blocks, retry any missing token and re-render every page.
The proofs establish that each exact record existed no later than the attested times, and that it has not changed since. They do not establish who authored it beyond the domain it is served from, they do not bind the order of entries beyond their timestamps, and they do not prove that nothing has been removed after the last entry. Those three, a signing key per publisher, a per-series hash chain and periodic signed completeness roots, are the open items before the record is offered to third parties, and they are stated on the record's own verification page.
Deliberately absent from this page: the canonical form specification, the contract fields, calendar and authority endpoints, and infrastructure identifiers. The method is described to the level at which its guarantees can be judged, not reproduced.
Discuss putting your own published claims on the record: forecasts, analyses, statements, or the outputs of a system you run, each entry verifiable by anyone.