1 · Why Validation Matters — And What We Tested

When an underwriter, asset owner, lender, or EPC bids on a solar site, the accuracy of the underlying hazard, soil, hydrology, and irradiance data is the difference between a profitable underwrite and a catastrophic loss. Earthflow’s foundational claim is that every numeric value the platform emits is rooted in a citeable, authoritative public dataset — NOAA, USGS, FEMA, NASA, NREL, USDA, ASCE — not in opaque statistical regressions.

1.1 Four Classes of Independent Ground Truth

1.2 Table of Contents

2 · What Is Being Validated — The Physics AI™ Pipeline

Earthflow is a Physics AI™ platform — environmental intelligence that starts from established physical equations (RUSLE, KBDI, HAND, ASCE 7-22, FEMA HAZUS, IEC 62305, Heymsfield-Wright, Faiman, Kimber, NSHM 2023) and uses Physics AI™ only where equations alone cannot answer the question. Every numeric value the platform emits traces to either a citeable equation or a named authoritative public dataset (NOAA, USGS, FEMA, NASA, NREL, USDA, ASCE). This document validates that the production pipeline implements those equations correctly and that the resulting outputs match independent ground truth across geographies.

2.1 The Physics AI™ Pipeline at a Glance

For every coordinate the platform receives, 19 parallel Physics AI™ modules ingest data from 70+ authoritative sources, run their respective physics equations, and emit an average of 1,165 fields per site — all in roughly 49 seconds end-to-end. Each module is independent: a single upstream source outage never blocks the others, and every output carries a provenance tag identifying its source dataset and version.

This architecture is what makes the platform defensible: every output is auditable to a published formula and a named data source, with no opaque ML weights driving headline numbers. It is also what makes the pipeline validatable in the way this report demonstrates — against published reference values from the same authorities that publish the underlying equations.

2.2 The Validation Loop

Each validation site has a structured golden-record JSON file declaring the expected value for each field, the citeable source (e.g. “NOAA SPC Storm Events DB 50-mi catchment”), the acceptable range, the expected units, and a rationale. The comparison engine iterates every entry, locates the matching Earthflow output, and classifies the comparison into one of five buckets:

2.3 Portfolio-Scale Sanity Checks

A field-by-field check passes a single site, but only a portfolio-scale view confirms that the platform behaves correctly across geographies. Three independent sanity tests run on every portfolio comparison:

3 · The Seven Golden-Record Anchor Sites

These seven sites form the validation backbone — chosen because each carries published, citeable ground truth from an independent source (an insurance loss report, an instrument-grade reference station, a federal monitoring network).

Table 3.1 · The seven golden-record anchor sites (six SURFRAD / SRRL reference stations + two catastrophic-event retro-prediction sites). The fuller validation framework also runs against an 8th University-of-Oregon site plus 31 USPVDB stratified solar plants — covered in Chapter 6.

4 · Catastrophic Hail Event Retro-Prediction

The single most consequential question a solar underwriter asks: “Would your model have flagged this site as high-risk before the loss?” We tested Earthflow against the two largest utility-scale solar hail-loss events on record. Both pass.

4.1 How the Pre-Event Classification Was Produced

The Physics AI™ pipeline doesn’t just count storms — it interprets them through layered physical reasoning. The NOAA Storm Prediction Center (SPC) hail catalog is raw observational data: decades of discrete event records scattered across millions of square miles, with no per-site climatology, no impact physics, and no loss outlook attached. Reading the catalog at a coordinate doesn’t answer “how risky is this site” — it just lists nearby events.

Earthflow’s Physics AI™ Hail Module bridges that gap. It converts scattered catalog observations into a per-site climatological signature, runs each maximum-stone-size record through universal atmospheric impact physics to derive terminal velocity and kinetic energy, maps the resulting impact-energy distribution onto module-fragility curves rooted in FM Global Property Loss Prevention standards, and assembles the chain into both a categorical risk tier and a defensible $/MW/yr expected annual loss. The interpretation is where the value sits — not in the raw catalog and not in any single equation, but in the orchestrated chain that turns historical observations into a per-site engineering decision. Critically: no machine learning, no statistical regression, no per-site tuning. Every coordinate in the country runs through exactly the same transparent Physics AI™ chain — universality is the validation.

4.2 Confidence Boundary — Temporal Honesty

4.3 Reinsurance-Grade Hail Physics

Beyond the categorical risk label, the Physics AI™ pipeline emits a full hail kinetic-energy + module-fragility loss curve at every site — built from published terminal-velocity physics and FM Global module-fragility curves. At Fighting Jays, the catalog-maximum 4.5-inch stone produces:

This isn’t a categorical label dressed up as a number; it’s an auditable, universal physics chain. Reinsurance underwriters receive a loss curve, not a tier name. Every step traces to a published authority — impact physics from peer-reviewed atmospheric-sciences literature, module-fragility curves from FM Global, and replacement-cost benchmarks from the reinsurance industry.

5 · NREL SRRL — Instrument-Grade Reference Comparison

The NREL Solar Radiation Research Laboratory Baseline Measurement System (SRRL BMS) operates 80+ ISO 9060 Class A pyranometers and pyrheliometers at 1-minute cadence on South Table Mountain in Golden, Colorado — a continuous BSRN-quality record running since July 1981. It is the single highest-fidelity multi-domain reference station in North America. Validating Earthflow against SRRL is the strongest single-site test of the continuous-resource modules.

5.1 Detailed Field-Level Results at SRRL

6 · CONUS Portfolio — 44 Sites Across 25+ States

A handful of golden-record sites tests accuracy. A 44-site CONUS portfolio tests generalization — whether the same Physics AI™ pipeline behaves correctly in the Mojave Desert and the Mississippi Delta, in Florida hurricane country and the Northern Plains, without any per-site tuning.

6.1 Geographic Coverage

The portfolio combines six NOAA SURFRAD + two NREL MIDC reference stations with 31 stratified USPVDB v4.0 utility-scale solar plants (each >10 MW). States represented:

States represented (overlapping): AL · AZ · CA · CO · FL · GA · IL · IN · MA · MI · MN · MS · MT · NC · NM · NV · NY · OH · OK · OR · PA · SC · SD · TN · TX · UT · VA · WA · WI.

6.2 Pipeline Performance Across the Portfolio

6.3 Hail-Belt Geographic-Gradient Test

The hail-risk classifier passes the geographic-gradient sanity test — Hail Alley sites all cluster Very High, non-Plains sites distribute realistically across all five tiers:

6.4 Per-Site Physics Spot-Check — Eight Climate-Diverse Sites

To verify the per-site physics outputs (hail EAL, ASCE 7-22 wind §27 and tornado §32, IEC 62305 lightning, HAZUS-MH 2.1 seismic, foundation classifier, OSM road access, PV performance derate, permitting layers), eight sites were captured directly from the production /analyze endpoint — each chosen to exercise a specific physics output at a regional extreme.

7 · Classifier Sanity & Tier-Distribution Health

A classifier that emits the same tier at 77% of sites isn’t a classifier — it’s a constant dressed up as one. Earthflow validates every risk-classifier output on portfolio-wide tier distribution before any production deploy. Below are the post-validation distributions across the 44-site portfolio for the five primary peril classifiers.

7.1 What Healthy Distributions Look Like

Each donut shows full tier diversity — no peril classifier is stuck on a single tier across the portfolio. The hail and tornado classifiers concentrate (correctly) on the high end because the portfolio overrepresents Hail Alley by design. The erosion classifier concentrates (correctly) on Low because most utility-scale solar plants are sited on flat, low-erosion-risk terrain by definition.

7.2 Realistic Distributions Across the Portfolio

The three classifiers below show the realistic variation customers see across a continent-wide portfolio — each value sourced directly from the cited public authority, not from a regional default:

8 · Reinsurance-Grade Physics Outputs

Categorical risk labels (Very High / High / Moderate / Low) are useful for portfolio screening but inadequate for pricing. Earthflow emits per-site physics-equation outputs for each of the four primary catastrophe perils — hail, wind, seismic, lightning — sourced from published engineering standards (ASCE 7-22, FEMA HAZUS-MH 2.1, IEC 62305-2, FM Global PRG 18-1, NASA LIS/OTD).

8.1 Why Equations Beat Regressions

9 · Authoritative Data-Source Chain

Earthflow’s foundational claim — every value traces to an authoritative public dataset — is verifiable. Every one of the 19 production modules below lists its primary data source and the federal/international authority that publishes it. There is no model weight, no statistical regression, no ML training data quietly providing the underlying numbers.

10 · What This Means for You

10.1 Confidence Summary

10.2 Appropriate Use

10.3 The Business Value, in One View