nextX · Verify

Proof,
not promises.

Run our benchmarks against your own data. Live-verifiable via Qdrant read-only API and cross-vendor Bit-Identity fixtures. 30-minute round-trip from request to result.

terminal · verify-compress.sh● LIVE

From your terminal:

No SDK install. No account. Just a read-only API token.

terminal · verify-compress

$ aqea-bench encode --corpus msmarco-1M \

--qdrant-readonly $AQEA_VERIFY_TOKEN

# results in ~30 minutes

✓ Recall@10 = 0.8836

✓ HN@10 = 0.000   (vs 0.561 baseline)

✓ HitC@10 = 0.832 (vs 0.800 baseline)

# compression ratio: 117× (1024D → 11D)

From request to result in 30 minutes

No demos. No sales calls. Just a token, a command, and your own data.

Step 1 · 2 min

Get the token

Fill the engineering-eval form. We email a read-only API token within minutes.

Step 2 · 25 min

Run the bench

Execute the script against msmarco-1M. Watches the encode + retrieval pipeline live.

Step 3 · 3 min

Compare

Your numbers match the published results. Or they don't — and we want to know.

Live-verifiablePublic test fixturesSHA-256 manifestsCross-vendor reproducible

Manifest

If we can't show you, we don't claim it.

Most AI vendors ask you to trust their numbers. We don't. Every headline metric on this site is reproducible from public test fixtures, on hardware you control. This page is the shortest path between our claims and your terminal.

Two paths

Pick the verification that maps to your problem.

Path 01 · Live API

Verify Compress

117× compression with 0% false positives against Qdrant Cloud Read-Only. Public endpoint, returns metrics in under 5 minutes.

  • Read-only Qdrant Cloud Partner API
  • LexGLUE 50k pool, deterministic seed
  • Compare against published WP numbers
Run the Compress bench

Path 02 · Public fixtures

Verify Gen-5 Bit-Identity

Encode on one platform (CUDA / Vulkan / Metal / WebGPU / AVX-512 / NEON), verify byte-identical results on another. SHA-256 manifests included.

  • Apache-2.0 test fixtures
  • 9 / 9 fixtures byte-identical at last release
  • Reference SHA-256 in manifests/
Run the Bit-Identity test

Public test fixtures

What we publish.

  • Pre-registered benchmark thresholds — declared before measurement.
  • SHA-256 manifests for every encoded corpus.
  • Hardware specs for every published number (Hetzner AX102 for CPU, Lambda Cloud H100 PCIe for GPU).
  • Energy measurements via RAPL (CPU) and nvidia-smi (GPU) at 100 ms resolution.
  • Our own falsified hypotheses — see Phase J in the Shader Whitepaper §3. We publish what we did not prove.

Trust anchors

Independent verification, coming soon.

Partner-published benchmark results will appear here as they land. Until then — three anchors you can verify today.

Founder track

Inventor of record on all 25 USPTO filings. Engineering at Bosch, BMW, Mercedes, Trumpf, OWIPEX, Forschungszentrum Jülich.

Cross-vendor reproducibility

Bit-identical encodings across CUDA, Vulkan, Metal, WebGPU, AVX-512, NEON. Public SHA-256 manifests.

25 USPTO filings

Patent posture covers all three pillars end-to-end across six tracks — substrate, encoder, cortex, embedding, governance and response-authorization.

Ran our benchmark? Share your result.