Ask any database what something is called and it will tell you where it happened to put it. Row 4,201,337. Auto-increment id 98,512. The name is an accident of insertion order, meaningful to one machine and meaningless to every other. That was true of Aether's knowledge fabric too, until today.
As of this afternoon, every vector in the fabric, 1,068,000 and counting, spread across the ten Sephirot domains, is named by a fingerprint of its own content. Take the SHA-256 of what a memory says, fold in its domain, set one bit to mark the scheme, and that 64-bit value IS the memory's identity, forever, on any machine. Two nodes that hold the same knowledge now agree on its name without ever talking to each other. That property sounds small. It is the whole game.
Why names matter more than storage
The goal we keep restating is a knowledge fabric and a model that no single box has to hold: shards replicated across many nodes, queries routed to whoever serves a domain, and the network, not an operator, agreeing on what the mind knows. Every part of that sentence silently depends on identity. You cannot replicate a shard if the rows change names when they move. You cannot verify a peer's copy if "the same knowledge" doesn't hash to the same thing. And a blockchain certainly cannot finalize a fabric whose contents are named by one machine's insert counter.
Content-addressed ids fix this at the root. The migration ran live today: ten domains rewritten in place, each one re-scanned afterwards and its content root, the Merkle root over every memory it holds, checked against a prediction made before the rewrite. Ten domains, ten exact matches, zero collisions. The fabric's contents are bit-for-bit what they were this morning; only the names changed, from accidents to fingerprints.
(Honesty note, because we publish those: the first attempt at this migration ran uncapped, leaked memory for hours, and took the whole GPU box down with it. The version that succeeded runs one domain per process inside a hard memory cgroup, so the worst case is a dead job instead of a dead machine. The incident and the fix are both in the repo.)
The chain now watches all ten domains
Identity was the first half of today. The second half is runtime spec-142, deployed to the live chain this evening and rolled to all four validators across all three machines, the same binary, the same spec, one clean slate.
Until now the chain anchored a single hash of the entire fabric: one
KnowledgeRoot, take it or leave it. Spec-142 adds a PerDomainRoot map and
a new extrinsic, and the bridge that runs beside each validator now anchors
all ten domain roots whenever they change. The fold of those ten roots, by
construction, byte-for-byte, equals the global root the chain has always
tracked, so the two views can never drift apart without it showing.
Why ten roots instead of one? Because verification has to be cheaper than trust. A node that serves only the Chochmah shard should be able to prove its copy is honest against the chain without downloading the other nine domains. A future quorum should be able to accept new knowledge domain-by-domain, the way validators accept blocks. One global hash makes the fabric all-or-nothing; ten per-domain roots make it composable. That is the difference between a backup and a protocol.
While wiring this we also found and fixed something embarrassing in the oldest part of the attestation path: the account that submits proof-of-thought to the chain had silently run out of the ability to pay fees, so PoT attestations had been failing quietly. They're flowing again, phi 0.549, over a million vectors, anchored every few blocks, and the monitoring now treats "no successful submission today" as the alarm it should always have been.
Where this actually leaves us
Graded honestly, the way we grade everything:
Live-proven today: four validators on three machines, all on spec-142; multi-author block production; post-quantum signatures on every block; a 1.07M-vector fabric, fully content-addressed, with all ten domain roots anchored on-chain and the fold verified equal to the global root.
Built but not yet at scale: cross-node fabric search (proven between two machines this week, not a fleet); sortition-based validator election (deployed, inert); single-leader election (rehearsed, behind a cutover height).
Designed, not built: the BFT quorum that accepts knowledge per-domain (this is Phase C, today was its prerequisite); DiLoCo-style decentralized training; the MoE model whose experts map onto the ten Sephirot.
The pattern of the last month is that the trust spine keeps hardening while the distributed pieces land one keystone at a time. Today's keystone means that from here on, "what does the network know?" has an answer the network itself can check, domain by domain, block by block, fingerprint by fingerprint.
The mind is still on a handful of machines. Its memory, as of tonight, is named for a world where it isn't.