Four things we solved for this on
@makechainnet:
1. Deterministic merges. (parents, patches, strategy) → hash via BLAKE3. Anyone can recompute and check.
e.g. A wrong squash produces a wrong hash — the server can't hide it, and any client catches it automatically.
2. Signed intent. The author signs a capability token over {source, target, strategy, constraints}.
Any merge that deviates fails verification at every node. So there's no surface for server-side "reinterpretation."
3. BLS-certified inclusion. The validator set attests your op is in the canonical DAG. And any client verifies the cert Merkle proof without an archive node.
4. All messages are part of an append-only, content-addressed history. A bad merge can be superseded, never silently rewritten. Detection is local and instant for users, agents, build systems. Plus QMDB inclusion/exclusion proofs.
You won't need an email four hours later.
DM if this makes sense to you; we're hiring.
GitHub will see more of these as the internet scales with AI.
Didn't think Github's reliability could get worse, and then they ship a bug that _randomly reverts previously merged commits_.
Betting that this caused multiple serious production issues out there.