Real recorded data Synthetic tests labelled inside

Klein Biodefense System — Compiled Benchmark Report

Paul Samuel Guarino — Independent Researcher, Guarino Laboratory, East Northport, New York Compiled: August 26, 2026 · System version v23.3 Patent family: US 63/834,399 (provisional, June 2025 priority) · US 19/700,318 (non-provisional) · US 63/937,225 (provisional, November 7, 2025 priority) · US 64/028,112 (provisional)

1. What It Is

A deterministic, multi-pathogen outbreak and contamination detection engine. The system processes time-series and event data from public health and environmental feeds and flags the onset, escalation, and severe phases of biological, chemical, and environmental threats. It is designed for early warning, not clinical diagnosis.

• Deterministic: identical input produces byte-identical output on every run. No neural network, no model training, no weights to drift, no black-box probability.
• Air-gapped capable: runs on local data with no network, cloud, GPU, or external model dependency.
• Auditable: tamper-evident trail suitable for regulatory and public-sector review.
• Cross-domain: validated on viral, bacterial, toxin, chemical (nerve agent), and environmental contamination time series.

The detection layer measures coherence breakdowns in surveillance time series against a fixed mathematical reference (41.176 Hz, derived as 700/17) and layers five detection tiers, from absolute-limit violation through statistical spike, wave detection, harmonic (phase-locking) disruption, and coherence lock.

2. Headline Results

• Zero false negatives across every benchmark ever run — the system has never missed a crisis event, on any dataset, at any threshold setting.
• Published aggregate: 13 real-world and synthetic tests, 14,891 records, 100% recall, F1 = 1.0000 (pre-NDA redacted brief). The portfolio has since grown; current figures below.
• Early warning lead time: detection of the 2011 Yap dengue outbreak 85 days before peak with the full system, and 70 days before peak with the threshold layer removed entirely. Detection of the second wave of the 1918 Geneva influenza pandemic 22 days before its peak (p = 1.82×10⁻⁸).

3. Real-World Outbreak Benchmark (current portfolio)

All datasets processed through the same engine with the same fixed parameters. The only per-dataset input is the absolute response threshold, set to each threat's real-world action level (e.g., WHO emergency threshold).

Coverage: 4 continents; bacterial, viral, toxin, and chemical nerve-agent threats; single-wave, multi-wave, gradual-onset, and instantaneous mass-casualty event structures.

Synthetic stress tests

Eight synthetic stress tests push the engine beyond real data: borderline oscillation, intermittent spikes, rapid oscillation, post-incident decay, and poisoned baselines. Five of eight achieved perfect F1 = 1.000. The remaining three, under extreme noise, produced false positives while maintaining 100% recall — the engine errs on the side of public safety by design. Across all 13 tests and 14,891 total records: zero false negatives.

Cross-domain environmental validation

Four environmental datasets from the original Klein's Pollution Solutions validation, all F1 = 1.0000: Flint lead crisis (EPA/Michigan DEQ records), Bethpage TCE plume (US Navy NAVFAC), East Palestine derailment (EPA), and daily asbestos events (OSHA/NIOSH, 3,499 records).

4. Early Warning: Lead Time on Real Outbreaks

The capability that matters operationally: not confirming that a crisis crossed a threshold, but detecting escalation before it does.

4.1 Yap Island dengue, 2011 — isolated-island natural experiment

Real daily surveillance counts (RECON outbreaks R package, published Yap records; 185 daily observations, 978 total cases, population ~11,000). Nothing modeled or simulated.

• Full system: first detection 1 September 2011 — the day of the first case, 85 days before the epidemic peak (25 November 2011), with 1 of the outbreak's 978 cases recorded. Zero missed days across the 185-day epidemic. At the zero-tolerance setting: recall 1.0000, precision 0.9933, F1 = 0.9966, specificity 0.9730 (one false positive in 185 days — a harmonic-tier firing on a no-case day, evidence the harmonic layer operates independently rather than shadowing the count-based tiers).
• Threshold removed entirely (robustness ablation): with the absolute-limit tier switched off, the statistical and harmonic tiers alone still identified the epidemic on 16 September 2011 — 70 days before peak — from 7 cases, with 99.3% of the epidemic's cases still in the future. The early warning does not rest on the threshold rule.

Baseline was established from the first 15 observations, which precede the outbreak — a genuinely naive baseline. The engine was never shown the epidemic before being asked to find it.

4.2 Geneva 1918 influenza — multi-wave structure

Real historical record (Chowell et al. 2006), a two-wave epidemic with a prolonged inter-wave trough — the hardest detection structure in the portfolio.

• Crisis onset, decline, resurgence, and resolution all detected with zero false positives.
• Second wave detected 22 days before its peak (onset September 25; peak October 17), at statistical confidence p = 1.82×10⁻⁸.

4.3 What this shows, and what it does not

It shows same-day detection of real outbreak onset and statistically significant early warning of escalation on real, named historical events, surviving removal of the threshold layer.

It does not show prediction. Detection is same-day, not prior. It is retrospective — an analysis of historical outbreaks, and a reason to run the system prospectively, not a substitute for having done so. Prospective validation on a live feed is the natural next step and the intended deployment test.

5. Algorithm Configuration

Fixed parameters, identical across all datasets: anchor frequency 41.176 Hz (700/17), chamber 110.0 Hz, pulse 7.83 Hz, FFT size 256, rolling window 20, sensitivity 2.0, baseline window 15, significance α = 0.05, minimum effect size Cohen's d = 0.8, PLV threshold 0.50. Per-dataset input: the absolute response threshold only.

Five detection tiers: (1) absolute-limit violation — the benchmark metric; (2) statistical spike (>4σ over baseline) — early warning; (3) wave detection — early warning; (4) harmonic disruption — PLV collapse against the anchor — early warning; (5) coherence lock — confirmation.

7. Data Provenance

All real-world datasets derive from published, citable sources: WHO global COVID-19 data; WHO Ebola situation reports (2014–2016, Guinea/Liberia/Sierra Leone); the EpiEstim SARS2003 dataset (Cori et al. 2009, PLoS Comput Biol); the Geneva 1918 record published with Chowell et al. 2006 (J. Theoretical Biology); Tokyo sarin casualty reconstructions from Okumura et al. 1996 (Annals of Emergency Medicine) with Tokyo Metropolitan Police incident records; Yap dengue from the RECON outbreaks R package. Environmental datasets from EPA, Michigan DEQ, US Navy NAVFAC, and OSHA/NIOSH public records. The three attack-scenario datasets (anthrax, plague, ricin) are synthetic, modeled from documented events, and are labeled as such throughout.

Complete per-dataset citations, thresholds, and file inventory: MASTER_EVIDENCE_RECORD.md.

8. Availability

Available for licensing or government partnership. A pre-NDA redacted capability brief is available without restriction; extended technical briefing, source-code walk-through, and patent details are available under NDA.

Contact: Paul Samuel Guarino — Guarino Laboratory, East Northport, New York

This compiled report consolidates the master evidence record (August 12, 2026), the Yap dengue detection report (August 15, 2026), and the published redacted capability brief. All figures trace to those source documents.