Frozen public sample · Apache-2.0 source

A small search benchmark you can inspect.

Ten independently written intents were judged against a frozen 24-resource slice of public GoPlausible x402 discovery metadata. The dependency-free lexical baseline reports every ranking and metric.

0.958

nDCG@10

0.950

recall@10

1.000

MRR@10

0%

zero-result rate

What this does—and does not—prove

The sample demonstrates a reproducible evaluation workflow over public, multi-network HTTP metadata. It is not Stellar-specific, does not cover MCP, does not independently test endpoint liveness, and does not endorse any listed service. The observed mean ranking time (2.992 ms) is local and environment-dependent, not network latency. The proposed commercial 40-query holdout would be newly built after written scope and funding agreement.

Snapshot captured 2026-08-23T13:53:24.080Z from a source catalog reporting 1545 entries; 24 were selected for capability diversity after excluding security, trading and gambling.

Per-query result

IDIntentnDCG@10Recall@10Hits
q01play one legal move in a shared chess game1.0001.0001 / 1
q02search Pacific data endpoints by country and category1.0001.0002 / 2
q03Pacific tuna stock findings summary0.9451.0002 / 2
q04commercial license and production mesh for a lamp model0.6990.5002 / 4
q05agricultural weather risk forecast for farmers0.9651.0003 / 3
q06summarize a long document article or transcript0.9731.0003 / 3
q07verify a signed Algorand document on chain1.0001.0002 / 2
q08compute a cryptographic sha256 hash of text1.0001.0001 / 1
q09check whether a claim survives independent web research1.0001.0002 / 2
q10explain an Algorand transaction in plain language1.0001.0003 / 3
report sha256:82050FD5F3F8D07A945EC4EBF72C983359E9AB6CC2E6E5889E9BF90F8F740A26
snapshot file sha256:A40C842A98C6A9B19A9EA7FE6B58DEA208AEA80D8DB124EC62DB30472C2B9685
judgments sha256:817C3D97BB457FC595D720B59E9F43CE2802D493BC039FB020BFC000E794903E