Winnex Tracer Series: Mathematical-Proof Retrieval for Regulated Domains — Government, Legal, Healthcare, and Finance, Built on the Winnex Madhava Engine (Product Family of the Winnex AI Stack)

This document is the pre-patent technical specification for the Winnex Tracer Series — a product family of the Winnex AI stack. Four source-available products (Tracer-GOV, Tracer-JUS, Tracer-MED, Tracer-GAP) built on the Winnex Madhava deterministic vector-search engine and the shared Winnex tracer-gov audit framework. Each product provides soundness-guaranteed retrieval (per-document Cauchy-Schwarz exclusion bounds, zero bound violations), a tamper-evident SHA3-256 evidence chain, and compliance-ready self-assessment reports for its regulated domain (government audit, legal e-Discovery, clinical triage, financial gap identification).

The Winnex Madhava engine is published on PyPI as winnex-madhava (native C++20 core). The published benchmarks demonstrate correctness and auditability: real-data Kaggle results across all four products show recall@10 = 1.0000 with zero bound violations. Honest scope: at the published corpus sizes (187–2,000 documents) the system evaluates every document per query (bound pairs = |corpus| × queries), so the benchmarks demonstrate correctness and auditability, not sublinear speed; efficiency at production scale and float32 eps-slack rigor are planned future work.

All products are licensed under Business Source License 1.1 (BSL 1.1) — source-available (not OSI open-source), free for Brazilian government agencies. This document covers the algorithm, the products, benchmark results, use cases, and the engineering rationale — as a pre-patent disclosure. Part of the Winnex AI stack (related: Winnex Madhava Cascade 10.5281/zenodo.20970487; Winnex AI Inference Stack 10.5281/zenodo.19630736).

Version 4 (2026-08-06): Winnex Tracer Series — renamed and repositioned as a product family of the Winnex AI stack; honest scope section added (benchmarks demonstrate correctness/auditability, not sublinear speed; WORM = tamper-evident; BSL 1.1 = source-available, not OSI open-source; float32 requires eps-slack).