An agent buying data over x402 has no record of what it bought. It pays, gets an empty payload, and next run it pays the same provider again. Every purchase is a cold start. cassum remembers what each provider actually delivered, and ranks on cost per delivered payload instead of sticker price.
The same workload twice: once against Sibyl Memory, once against a NullStore that accepts every write and returns nothing on every read. That second run is the judges' litmus test, executed rather than asserted. Fresh database each time, computed now.
flat_fleet is kept deliberately: when no provider has an edge, memory costs 3.75% more. A negative that survives in the repo as a pinned test.
idle. nothing precomputed on this page.
Learn spends against the fleet and writes what it measured to disk. Recall is a separate interpreter with a different pid that opens the same file, cannot buy — its record_purchase raises — and still names the right provider. Nothing passes between them but the database.
Compare the pid on the first line of each transcript.
idle.