Reproduce the proof
The claim is arithmetic, so you should be able to run it. Each step below was executed against the published packages; the numbers are what came back.
What this shows
A corpus of 3,000 real contract clauses is indexed, a query is run, and every clause not returned carries a Cauchy-Schwarz bound below the K-th score. The record of that run is signed with Ed25519 and written to a hash-chained log.
Nothing here needs an API key. The packages install from PyPI and the run takes about two minutes on a laptop.
What came back
Measured with all-MiniLM-L6-v2 on 3,000 clauses from the CUAD corpus.
| Metric | Value | What it means |
|---|---|---|
| Clauses indexed | 3,000 | annotated by lawyers, 41 categories |
| Provably excluded | 2,982 | each carries a bound below the K-th score |
| Bound violations | 0 | no relevant clause fell outside by chance |
| Signature | Ed25519 | tampering with the count invalidates the record |
| WORM chain | SHA3-256 | removing a record breaks the link and names the block |
Run it yourself
Three packages, one call. The proof is on by default once the basis is aligned to the corpus — with a random projection the bound excludes nothing.
pip install winnex-madhava winnex-tracer winnex-ai-normalize (ver 1.9.25 / 1.1.4 / 1.4.6 or later). The quality gate routes the basis from a measurement, so you do not have to choose it.