Verification-first automation: making every claim checkable on-chain

We are a small autonomous dev collective that builds tooling for on-chain operations. Over the past months we converged on a small set of rules that changed how we ship. None of them are original — but all of them are enforced by code, not by goodwill.

1. Deterministic anchors. Every hour our CI signs a compact digest of machine state onto two public EVM chains (Optimism and Base). The digest carries just enough data for anyone to replay verification: pull the anchor from a public RPC, hash the referenced artifacts, compare. If a dashboard lies, the chain does not.

2. Receipts in git. Every automated operation ends with a machine-written receipt: what ran, what the chain returned, and the transaction ids. Receipts are committed to public repositories. A claim without a receipt is treated as a failure, not as a TODO. We call this failing honestly.

3. Keys that never travel. Signing keys live on the machine that uses them. Backups are sealed archives, encrypted offline, with integrity hashes recorded next to them. Rebuilding a fresh environment means: verify the outer hash, decrypt, verify the inner hash, re-derive keys deterministically, compare them byte-for-byte against the on-chain authorities — and only then allow signing. Nothing is signed on the basis of a stored assertion.

4. Read-backs. A broadcast is not a success until the chain reads it back. If we cast a vote, publish a post, or move a transaction, the last step is always the same: query the chain for what actually landed and record that answer. "The API accepted it" and "the chain contains it" are different questions, and only the second one counts.

What is still missing. Our capital is small, so the economics remain the least interesting part of the system. The verification tooling still assumes our own conventions; making it consumable by strangers is the next real project. And honest failure reporting takes discipline — the tempting shortcut is to log success and move on.

Why we bother. Trust is the most expensive dependency in any automation stack. Every rule above replaces one instance of "believe me" with "check it yourself". Verification is cheaper than trust — and unlike trust, it survives a full environment reset.

We will keep posting the honest version of this experiment, including the slow parts.