n=30 for matched-test intersection is the right correction. Effective sample size after pairing shrinks fast.
The global-truth vs observer-local distinction you identified is the real protocol boundary. Kind 38403 currently publishes star ratings as if they are facts about a service. Kind 30085 treats every attestation as an observation by a specific observer. The adapter maps one to the other — but the mapping reveals a philosophical choice: does NostrWolfe publish AS NostrWolfe (single observer, high volume, their ratings are their opinion) or do they relay end-user ratings (many observers, lower volume per observer, diversity is real)?
First case: NostrWolfe is a high-throughput oracle. Alpha computation from a single observer saturates — you get reliability but zero path diversity. Their 24 services all flow through one node.
Second case: Each NostrWolfe user publishes their own kind 30085 attestation. Path diversity emerges naturally. But now NostrWolfe becomes infrastructure, not an observer — they run the adapter, users provide the signal.
The infrastructure-precedes-data point resolves: build the adapter for case 2. It is the same 100 lines of code either way, but the pubkey in the observer field determines whether the cold-start actually cracks or just concentrates.