Open data

Five small datasets and an Evidence Explorer Space, live on Hugging Face, each labelled for what it is.

Status: published on Hugging Face. The five datasets, the Evidence Explorer Space and the collection are live: the collection groups them.

Five small datasets are built in CI from evidence that is committed in this repository. Each one is labelled for what it is, ships a MANIFEST.json with a sha256 for every file and the commit it was built from, and a verify.py that recomputes the hashes. A privacy and secret scan runs on every output.

Dataset	What it holds	Label
`gavel-gates`	Index of G0 to G5 with status, date, run links and evidence file hashes	real sandbox, synthetic audio, G3 in progress
`gavel-latency`	Settlement and per-call latency of the PayPal sandbox	real sandbox, small n
`gavel-sim-30d`	30-day , per and per hold	simulated, fake clock
`gavel-stt-bids`	24 synthetic bid clips, speech-to-text outputs and scores	synthetic audio
`gavel-audit-demo`	Event, and outcome records of auctions	synthetic,

A read-only static Space, gavel-evidence-explorer, charts them (gate status, sandbox latency, the simulated 30-day run, speech-to-text on synthetic audio), each section labelled. The collection is titled "GAVEL Evidence: auction holds, hammer and receipts".

Nothing here is production data, no real person or real money is involved, and the sandbox numbers say nothing about production PayPal. The full description of each dataset, and how they are built and published, is in docs/HF_COLLECTION.md. See also Real, simulated, recorded.