Reproduce the proof yourself

This page is not a claim. It is an instruction: run the code below and you will obtain the same numbers we publish, on real public data, in under two minutes.

What you are checking

That a vector search engine can prove, for each document it did not return, that the document could not belong in the top-K — and that the proof holds on a corpus you download yourself, not one we ship.

1. Get a real corpus

2,167 court decisions from the Brazilian Superior Court of Justice (STJ), published under open data. No account, no key.

curl -sL -o stj.json \
  "https://dadosabertos.web.stj.jus.br/dataset/bab65848-f287-4146-8273-e58faee39ad2/resource/de1881ee-8527-4b35-b1c6-99ad2b2fef57/download/20260930.json"

2. Build the index with the proof enabled

import json, numpy as np
import winnex_madhava as mad
from winnex_tracer.core import InstrumentedEngine
from sentence_transformers import SentenceTransformer

d = json.load(open("stj.json"))
docs = [f"{r.get('descricaoClasse','')}. {r['ementa']}"
        for r in d if len((r.get('ementa') or '').strip()) > 120]

enc = SentenceTransformer("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")
C = enc.encode(docs, normalize_embeddings=True).astype(np.float32)

engine = mad.build_engine(
    C, dim=384, metric="cosine", k=10,
    basis="pca_corpus",    # <-- the parameter that produces the proof
    stage1_dim=256,
    k1_fraction=1.0,
    postfilter=True,
    audit_exhaustive=True,
)

3. Measure

q = enc.encode(["penhora de bem de família por dívida de condomínio"],
               normalize_embeddings=True)[0]
c = InstrumentedEngine(engine, mode="commitment").search_with_commitment(q, k=10)

print(c["total_excluded_count"])   # documents PROVEN outside top-10
print(c["bound_pairs"])            # documents examined
print(c["bound_violations"])       # must be 0

4. What you should see

Metric	Result	Meaning
total_excluded_count	2,157 of 2,167	99.5% of the corpus proven irrelevant to this query
bound_pairs	2,167	every document was examined — none skipped
bound_violations	0	no exclusion was wrong
recall_guarantee	exact_global	the returned top-10 is the true top-10
latency	~92 ms	the cost of producing proof versus 0.2 ms without it

5. Compare with the default setting

Change one line — remove basis="pca_corpus" — and run again:

basis	Excluded with proof	Latency
`random` (default)	0 of 2,167	0.2 ms
`pca_corpus`	2,157 (99.5%)	92 ms

Why we show you this

The default configuration does not produce per-document proof. It is faster and the top-K is still exact, but nothing is proven excluded — every document is resolved by exact comparison instead.

We could have hidden this. A vendor page usually shows the best configuration and omits the default. We show both, because a CTO who discovers it later is a lost customer, and a CTO who discovers it here is a reference.

What this does not prove

Not that the retrieval is semantically good. That depends on the encoder. A poor encoder produces a correct proof over a useless ranking.
Not that the corpus is complete. A document never indexed appears in no counter and no proof.
Not that the answer is legally correct. This is a retrieval guarantee, not a legal opinion.

Run it on Kaggle — public output

Same measurement, executed on Kaggle's free CPU, with the output visible and reproducible:

Open the Kaggle notebook →

Configuration	Excluded with proof	Examined	Latency
`basis="random"`, s1=128	0	2,167	2.2 ms
`basis="random"`, s1=256	70 (3.2%)	2,167	2.4 ms
`basis="pca_corpus"`, s1=128	2,153 (99.4%)	2,167	6.6 ms
`basis="pca_corpus"`, s1=256	2,157 (99.5%)	2,167	21.7 ms

Latency on Kaggle's shared CPU is lower than on our machine because the index was already built — the numbers that matter are the excluded counts and the zero violations.

Verified against brute force

brute-force top-10 : [48, 857, 872, 878, 983, 1384, 1471, 1828, 2103, 2120]
engine top-10      : [48, 857, 872, 878, 983, 1384, 1471, 1828, 2103, 2120]
identical          : True

documents proven outside : 2157 of 2167
bound violations         : 0
recall guarantee         : exact_global
signature                : VALID — Ed25519

If you want this on your own corpus

We will run it on data you already have, and you compare the result against what your current system produces. Fifteen days, no cost for the first one.

Request an audit See the code examples