Vector Retrieval Metrics — Ground Truth First
Resolves Recall@k ambiguity between baseline-overlap and ground-truth-based evaluation. Defines canonical metric names.
2026-01-03 · CC BY 4.0 · 281 KB
Research
nextX is rooted
in research.
25 USPTO patent filings across six tracks. End-to-end patent-pending stack covering substrate, encoder, cortex, governance and response-authorization layers. Whitepapers under CC BY 4.0 open access. We publish what we falsified — too.
25 USPTO6 patent tracksPre-registeredPhase-J FalsificationCC-BY-4.0
Patent portfolio
25 USPTO filings — end-to-end stack coverage.
Six patent tracks cover substrate, encoder, cortex, embedding, governance and response-authorization end-to-end. The entire stack is patent-pending today.
| Track | Theme | Codes | Coverage |
|---|---|---|---|
| A | Substrate + Hardware | M33, P1, P2, P3 | Storage layer + GPU pipeline + CPU SIMD trit substrate |
| B | Cortex / Memory / Audit | C1, C2, C3, C4 | Audit-proof memory + LLM compute reduction |
| C | Search-Engine Layer | A1, A2, A3, A4 | Adapter + search algorithms |
| D | Encoder/Decoder Reversibility | A5, R1 | Auditable Trit-Quantizer + Reversibility Master Method |
| E | Embedding (TSE / FM-Golay-QAT) | L1, L2 | Embedding-class patents |
| F | Governance · Response · Verifier-Trust | N6, N7, N8, N9a, N9b | Pre-inference pipeline · Verifier-only trust-state · Sub-ms sensor gate · Multi-modal response-authorization gate · Defense continuation |
Application numbers, draft PDFs, and patent prosecution details available under NDA-Full with strategic buyers and patent counsel.
Phase-J · Falsification story
We publish what we couldn't prove.
Why this matters
Most AI labs quietly drop hypotheses that don't deliver. We don't. Pre-registration plus public falsification is the only way to build infrastructure-grade trust.
In Phase J of our cross-modal validation, we pre-registered the hypothesis that AQEA's substrate would EXCEED the Float baseline on mass-spectrometry archives — based on prior noise-resistance results on raw industrial sensor streams.
The empirical result was PASS (96.7% retrieval-equivalent), not EXCEEDS. Archive pre-processing strips per-bin noise, eliminating the mechanism that drives Supra-Trit effects.
We documented the negative result. We characterised the exact condition hierarchy (C1+C2+C3) that determines when the property holds. We did not quietly drop the hypothesis. This is what scientific honesty looks like in AI infrastructure.
Reference: Shader Whitepaper v5.0 §3 Fig 11 · Edge Whitepaper v2.2 §3.4
Library
Whitepaper index.
Public · Product
Product papers — available now.
| Whitepaper | Topic | Date | Size | Access |
|---|---|---|---|---|
| AQEA 2-Page Summary v1.5 | Executive one-pager · Two-Flagship · Numbers · Engineering proof | 2026 | 388 KB | Download PDF |
| AQEA Shader v5.0 | Generation-5 Information-Geometry Substrate · Vector-Search Application | 2026 | 4.0 MB | Download PDF |
| AQEA Edge Sensor v2.2 | Cross-Modal Generation-5 Edge Processing · 13 Domains validated | 2026 | 2.1 MB | Download PDF |
| AQEA Gen-5 Data Sheet v2 | Substrate-property reference · Production benchmarks · Spec card | 2026 | 720 KB | Download PDF |
| AQEA Sales Deck v2 | Two-Flagship product overview · Customer-facing pitch (public-safe) | 2026 | 11.4 MB | Download PDF |
| CRONOS Investor Brief (EN) | Temporal Analytics product brief · 13 domains overview | 2026 | 54 KB | Download PDF |
| CRONOS Investor Brief (DE) | Temporal Analytics Produktbrief · 13 Branchen-Übersicht | 2026 | 613 KB | Download PDF |
Public · Methodology, Theory & Standards
Beyond product papers — open methodology for the industry.
We publish open methodology, theoretical frameworks, and propose industry standards for vector-retrieval evaluation. These four papers are foundational — cite and reuse freely under CC BY 4.0.
Ground-Truth-Aware Metric Terminology for Vector Retrieval
Extended standards spec (12 sections) with minimal reproducible evaluation protocol. Worked examples: legal search, e-commerce.
2026-01-05 · CC BY 4.0 · 321 KB
Multi-Layer Network Theory Resolves the Semantic Compression Problem (MLNT v0.0.3)
Formalises why single-embedding spaces cannot preserve domain expertise and cross-linguistic connectivity at the same time. Three-layer solution with O(d) complexity bounds. 15% higher Pearson correlation (r=0.831 vs 0.748) at 1.1× compute; MTEB +12% across 14 tasks; validated on 783K concepts.
2025-08-01 · CC BY 4.0 · 1.5 MB
MLNT v0.0.4 — Molecular Folding Revolution in Semantic AI
Extends MLNT with four innovations: Molecular Folding Layer ℳ=(F,U) achieving 1,536× compression at 94.7% quality; Bootstrap Convergence Theory (ρ≈0.87); Morphological Layer Enhancement across 12+ language families (+22.9% accuracy); Quantized Optimization (192× speed).
2025-07-31 · CC BY 4.0 · 282 KB
Theoretical foundation × Patent coverage
Multi-Layer Network Theory is the theoretical foundation underlying the AQEA substrate. All mechanisms described in MLNT v0.0.3 and v0.0.4 are covered by the 25 USPTO patent filings across Tracks A–F. Published under CC BY 4.0 — cite, extend, and build on freely.
Papers 01 and 02 were originally Zenodo-registered (10.5281/zenodo.18138436 + 10.5281/zenodo.18152431). Records withdrawn (HTTP 410) — now self-hosted by nextX AG. DOI references in PDFs preserved for citation integrity.
NDA-Light — available on request
| Whitepaper | Topic | Access |
|---|---|---|
| WP-Series (12 Whitepapers) | aqea-science deep-dives | Request access → |
| MLNT-Series (14 Versions, v0.0.3–v0.0.16) | AQEA Core mathematical proofs | Request access → |
| External Advisor Briefing v0.1 | Investor / Strategic-Partner briefing | Request access → |
All public whitepapers are released under CC BY 4.0 — open access by design. Copy, adapt, cite.
Discipline
Pre-registered hypotheses. Evidence classes A/B/C/D/LIVE.
Every numeric claim on this site is marked with an Evidence Class — so investors, customers and researchers know exactly what kind of backing each number has.
| Class | Definition | Use |
|---|---|---|
| A | Runnable code + tests + documented benchmarks | Supports valuation |
| B | Documented benchmark with artifacts/logs; packaging pending | Supports with DD-discount |
| C | Whitepaper / strategy claim without full reproducibility pack | Narrative use only |
| D | Pre-registered future experiment | Catalyst, not current value |
| LIVE | Run completed but not yet investor-grade frozen | Internal use |
Collaborations
Academic and industrial partnerships.
We work with academic researchers, industrial labs, and standardisation bodies. If you've got an interesting hypothesis on information geometry, deterministic AI, or substrate-class representations — let's talk.