01 — Why similarity cannot decide
Every pair below is measured live, right now, by the embedding model. Marks are coloured by what they actually are. Look at the colours before you read the labels.
measuring seven claim pairs against the embedding model — the first page load warms it, about eight seconds
measuring…
02 — Eight agents, one fact, at the same moment
Eight agents each learn that customer 4471 verified their identity. They phrase it differently and they extract the subject differently — Customer 4471, cust 4471, 4471. There is one fact in the world. A correct memory ends holding one belief.
Naive shared memory
read → think → insert
RETRACT
read → think → lock claim key → adjudicate → write
03 — The contradiction that is real
Now a belief chain, and a claim that is not a rephrasing. Same customer, same predicate, opposite fact — at a distance indistinguishable from a paraphrase.
04 — The money already moved
Concurrency control stops a bad fact being written. This is what reaches the bad fact that was written correctly, believed reasonably, and already paid out.
05 — The receipt, read back through a path that cannot write
Every other panel on this page reads the database over SQL. This one does not: it is served by CockroachDB’s Managed MCP Server, an endpoint with no write surface at all. So the agent that shows you what happened is structurally incapable of changing it. This heading used to say “what your fleet currently disagrees about” — it counted open contradictions, and once the story started resolving them it always read zero. A number that hits zero exactly when the product works is a badge, not a dependency.