1 · Executive Summary

Earthflow is a Physics AI™ environmental-intelligence platform purpose-built for utility-scale solar PV reinsurance underwriting. At its core sits the Physics AI™ engine — an explainable, physics-first modeling stack that quantifies hazard intensity, asset vulnerability, and stochastic loss from first principles rather than from historical claims regressions. This document specifies the modeling framework, data sources, computational pipelines, and outputs that underwriters, cedants, brokers, and capital providers can rely on to price, structure, and aggregate solar P&C and parametric covers.

1.1 Table of Contents

2 · Why Earthflow — Market Position & Advantages

The solar reinsurance analytics market sits at the intersection of three product categories: (a) generic property cat-modeling platforms (RMS, AIR Worldwide, KCC, Moody's RMS), (b) climate-risk specialists (Jupiter Intelligence, Climate-X, Sust Global, First Street), and (c) solar-specific analytics (kWh Analytics, GCube, GIA, distributed-monitoring SCADA vendors). Each addresses a slice of the underwriting workflow but none integrate the full picture in a single physics-informed pipeline calibrated to utility-scale PV.

Earthflow is purpose-built for that integration. It combines cat-modeling-grade hazard rigor with solar-asset-specific vulnerability functions and real-time IoT condition-monitoring, then surfaces the outputs through a reinsurer-fluent interface (loss curves, parametric triggers, ACORD-aligned bordereaux). This chapter introduces the Physics AI™ engine that powers it, compares Earthflow's positioning against the three competitor categories, and lists the differentiators.

2.1 The Physics AI™ Engine — Earthflow's Core IP

Every output in this methodology is produced by Earthflow's Physics AI™ engine — a hybrid stack that fuses three distinct technical traditions traditionally kept separate in the catastrophe-modeling industry:

The first layer — physics-based hazard — derives every peril intensity from first principles using federally curated physical sensor networks (NEXRAD reflectivity → MESH hail size; NREL Wind Toolkit → 50-year 3-second gust; USGS ASCE 7-16 design maps → PGA; NHC HURDAT2 + SLOSH → hurricane wind and surge). Nothing is fit to historical claims.

The second layer — mechanics-based vulnerability — translates hazard intensity into damage using laboratory test data (ASTM E1038 ice-ball impact studies for hail), engineering standards (ASCE 7-16 wind loads, IEC 61730 PV module stress envelopes), and manufacturer-published MTBF / derating curves for inverters and balance-of-system. Damage functions are parameterized by asset spec (glass thickness, frame type, racking clamp, stow capability) and adjusted by mitigation factors that multiply the baseline curve — not by black-box adjustment factors fit to a claims book.

The third layer — AI inference — sits on top of the physics-and-mechanics layers and handles the tasks where machine learning is genuinely useful: Earthflow Detect for satellite-imagery panel / substation / vegetation / construction-phase recognition; Cirra for natural-language agent reasoning across 27 tools; and continuous calibration that nudges parameter values when sensor evidence diverges from physical model predictions. AI is the inference layer, not the science layer — a critical separation for explainability.

What Makes Physics AI™ Different from Conventional Modeling

2.2 Competitive Landscape

2.3 Twelve Core Advantages

2.4 Earthflow vs. Each Competitor Category

3 · Framework Overview — The Four Stages

Earthflow's methodology is organized around the canonical four-stage cat-modeling workflow used industry-wide. Each stage takes the prior stage's output as input, and each stage's output is independently inspectable in the Earthflow dashboard. This is critical for underwriter trust — the model is auditable end-to-end.

3.1 What Each Stage Produces

3.2 Why This Order Matters

The four-stage workflow is not arbitrary. Separating hazard from vulnerability lets you swap underlying assumptions independently — if NOAA releases a revised hail climatology, only Stage 1 needs to update; if a new module hits the market with different ASTM E1038 ratings, only Stage 2 updates. Separating loss from financial lets you re-price quickly as policy terms change without re-running the catastrophe simulation. Reinsurers use this separation every day when running treaty-placement scenarios.

3.3 Why Physics AI™ — and Why Purely Statistical Models Fall Short

Most commercial cat models rely on historical-claim regressions to fit hazard frequencies and damage functions. This approach has three material weaknesses for solar PV: (a) the historical record is short (utility-scale PV is largely post-2015), (b) the climate is non-stationary (hail belts are migrating, hurricane intensities are scaling, wildfire seasons are lengthening), and (c) operational perils unique to PV (hot-spot contagion, BESS thermal runaway, inverter MTBF aging) have effectively zero historical baseline. Statistical fits trained on a 10-year claims window quietly assume all three conditions are stable. None of them are.

Earthflow's Physics AI™ engine sidesteps each weakness by working from physical hazard intensities and mechanical damage functions:

• Hazard — NEXRAD reflectivity drives hail size · NREL Wind Toolkit drives gust exceedance · USGS ASCE 7-16 maps drive PGA / S_s / S₁ · NHC HURDAT2 + SLOSH drive hurricane wind & surge · NLDN drives lightning density · CMIP6 ensembles drive forward-looking climate horizons. Each is a published, peer-reviewed physical dataset.
• Vulnerability — ASTM E1038 ice-ball impact studies define hail fracture probability · IEC 61730 envelopes define module stress limits · ASCE 7-16 defines wind uplift load on racking · published manufacturer MTBF + derating curves drive inverter aging. Damage curves are mechanical, not actuarial.
• Mitigation — Each factor (stow @ 75°, mid-clamp racking, 3.2 → 4.0 mm glass, BMS isolation valves) is a multiplier on the physical damage curve. The mitigation impact is the physics of the mitigation, not a discount applied after the fact.

The result is a model that extrapolates correctly to perils that fall outside the historical claims window, responds correctly to climate-driven shifts in hazard frequency, and — crucially for reinsurance — produces an audit trail from every score back to a published physical equation and external data source. This is what "Physics AI™" means in practice: AI is used for inference layers (satellite detection, agent reasoning, sensor-evidence calibration) but the underlying hazard and vulnerability science is physics, not pattern-matching.

3.4 How the Stages Map to the Dashboard

If you have explored the interactive demo (EarthflowRE), every panel maps to a specific stage of this framework:

• 12-peril decomposition grid → Stage 1 (Hazard) intensity scores × Stage 2 (Vulnerability) damage ratios summarized as a 0–100 score and AAL contribution per peril
• Exceedance Probability curve + TVaR breakdown → Stage 3 (Loss) output
• Net-of-mitigation overlay (stow / vegetation / glass) → Stage 2 toggle re-running Stage 3
• Premium guidance + deductible matrix → Stage 4 (Financial)
• Parametric triggers panel → Stage 4 (Financial) parametric pricing engine
• IoT / condition-monitoring panel → Stage 2 vulnerability inputs in real time
• Equipment pedigree panel → Stage 2 mitigation factors
• Construction-phase dashboard → All four stages reshaped for CAR/EAR + DSU framing
• Climate stress test → Stage 1 climate overlay across 7 horizons

4 · Stage 1 — Hazard Assessment

Stage 1 quantifies how often each peril occurs at each site and how intense it is when it does. The output is a frequency-severity distribution per peril per site, with climate-change overlays for forward-looking horizons. Hazard is computed by the Physics AI™ engine directly from physical data — radar reflectivity, fault maps, gauge networks, satellite imagery, climate ensembles — not from claims-data regressions.

4.1 The 12-Peril Taxonomy

4.2 Severe Convective Storms (Hail · Wind · Tornado · Lightning)

SCS is the single largest loss driver in solar P&C — hail alone accounts for ~73% of solar loss costs (GCube 2024) with an average claim size of $58.4M. Earthflow's SCS hazard treatment combines four distinct hazards (hail, straight-line wind, tornado, lightning) that share atmospheric drivers but require different vulnerability functions.

Hail Frequency / Severity

Wind / Tornado

Lightning Strike Density

NLDN cloud-to-ground strike density per square mile per year, 10-year climatology. Florida tops the US at 31.3 strikes/mi²/yr; arid Southwest sits near 1.5/mi²/yr. Lightning hazard scales the equipment-breakdown and module-degradation perils via secondary-effect coupling (surge events, DC arc-fault risk).

4.3 Tropical Cyclone (Hurricane + Storm Surge)

Hurricane hazard is treated as a frequency-of-direct-hit problem. Earthflow uses NHC HURDAT2 historical tracks (1851–present) Monte-Carlo-resampled with intensity scaling for climate-change horizons. Storm-surge inundation depth is derived from NOAA SLOSH model envelopes by Saffir-Simpson category.

4.4 Wildfire / WUI

Wildfire is treated as a propagation problem rather than an ignition problem — the question is not "will a fire start nearby?" but "will a fire perimeter reach the array?". The Physics AI™ Fire Risk Module (3,800+ lines of code) integrates 19+ satellite and ground-truth datasets (MODIS, VIIRS, GridMET, LANDFIRE, Sentinel-1 SAR, NIFC, Cal-Fire, USGS DEM) through six parallel Google Earth Engine queries to produce 49+ quantitative output fields in under 1 second per site.

Surface Fire Composite Score (0–100)

Subsurface (Peat) Fire — USDA Soil Taxonomy Thresholds

Peat fires smolder underground for weeks and represent a tail-risk class generally absent from conventional cat models. Earthflow classifies peat per USDA Soil Taxonomy and applies volumetric-water-content (VWC) ignition thresholds:

4.5 Flood

Earthflow combines four flood-data sources to handle the well-known gaps in FEMA NFHL (which omits pluvial / urban surface flooding):

• FEMA NFHL — official flood zones (A, AE, AH, AO, V, X, X-shaded)
• USGS NHDPlus + stream gauges — flow record · drainage area
• JRC Copernicus Global Surface Water — 38-year history of water occurrence
• MERIT-HAND — Height-Above-Nearest-Drainage at 90 m global

The composite flood score reflects all four sources, with elevation above BFE (Base Flood Elevation) as the primary mitigating factor. Distance to nearest stream and HAND values catch pluvial risk that FEMA misses.

4.6 Seismic

USGS ASCE 7-16 design maps yield peak ground acceleration (PGA), spectral accelerations S_s and S₁, and the Seismic Design Category (SDC: A through F). SDC drives foundation design loads, which in turn drive racking pull-test requirements.

4.7 Equipment / Operational Perils

Equipment-breakdown, module-defect / hot-spot, BESS thermal runaway, and cyber risks are hazards specific to operating solar assets — they have no analog in traditional property cat models. Earthflow treats each as a first-class peril with site-specific frequency:

• Inverter MTBF — manufacturer-published mean-time-between-failures, derated by ambient temperature and lightning exposure. Observed on-site MTBF tracked via SCADA fault log.
• Module hot-spot incidence — EL imaging at commissioning and annual intervals; trend analysis flags serial-defect contagion. Industry-wide hot-spot rate jumped 3.4× in 2024 (kWh Analytics).
• BESS thermal runaway — cell-level temperature differential (ΔT) monitored via BMS; isolation-valve test cadence + chemistry (LFP outdoor vs NMC indoor) determines the tail.
• Cyber / SCADA — NIST CSF audit findings + MFA + segmentation status; ERCOT-style grid-control exposure scored separately.

4.8 Erosion / Geotech (RUSLE 5-Factor)

Soil erosion uses the full 5-factor RUSLE equation with solar-specific innovations. The Physics AI™ Erosion Module (1,126 lines of code) anchors each factor to USDA Agriculture Handbook 703 baselines and validates against field observations from 650+ projects.

R-Factor (Rainfall Erosivity) — sub-linear with precipitation

K-Factor (Soil Erodibility) — USDA nomograph

LS-Factor (Slope Length × Steepness)

4.9 Climate-Change Overlay (CMIP6)

Every hazard frequency is recomputed under RCP 4.5 and RCP 8.5 scenarios at seven horizons (today, 2030, 2035, 2040, 2045, 2050, 2055). CMIP6 ensemble downscaling propagates climate-driven shifts in hail frequency, hurricane intensity, wildfire season length, and sea-level rise into the same modeling pipeline.

5 · Stage 2 — Vulnerability Quantification

Stage 2 translates hazard intensity into damage. For each peril and each asset class, a damage function maps the hazard intensity to a mean damage ratio (MDR — fraction of TIV lost in a single event). The damage function is parameterized by asset spec and mitigation factors — what makes solar PV vulnerability fundamentally different from generic property cat modeling.

5.1 Hail Damage Function (ASTM E1038)

The hail-damage curve is the most consequential single function in solar reinsurance. Earthflow uses lab-tested damage probabilities from the ASTM E1038 ice-ball standard, cross-referenced to NEXRAD-derived stone-size distributions in the field.

5.2 Solar-PV Mitigation Factors

5.3 BESS Vulnerability — Chemistry Matters

Battery Energy Storage Systems are increasingly co-located with solar (Tesla Megapack, Sungrow PowerTitan, etc.). BESS vulnerability is dominated by thermal runaway tail risk and is overwhelmingly driven by chemistry choice and physical configuration.

EPRI's 2024 industry survey found 18% of inspected BESS units had thermal-management defects. Earthflow's BESS hazard score reflects BMS certification status, cell-level ΔT monitoring, isolation-valve test cadence, and thermal-anomaly event history — not just chemistry.

5.4 Inverter MTBF Aging

Inverter failure causes 40% of solar downtime and 59% of lost energy industry-wide (kWh Analytics 2024). Earthflow tracks observed on-site MTBF vs manufacturer-rated spec and applies an aging-curve derating to vulnerability per peril (high-SCS sites accelerate inverter wear; lightning-heavy sites push DC arc-fault risk).

5.5 Solar-Specific Erosion C-Factor — Operational vs Construction

RUSLE soil-loss vulnerability is dominated by the C-factor (cover-management). For solar PV, the C-factor differs materially between construction and operational phases — a Physics AI™ innovation calibrated against field observations from 650+ utility-scale projects.

The implication for CAR/EAR pricing: erosion exposure is materially front-loaded into the construction phase. Sites with a long grading-to-energization gap (especially through monsoon or wet season) compound the risk; sites that grade-and-install in a single dry window minimize it. The phase tracker in Stage 1 catches these schedules.

5.6 Construction-Phase Vulnerability

During construction, the vulnerability profile shifts. Modules in transit / staging are vulnerable to hail damage at staging areas (covered tarps required); RUSLE erosion C-factor jumps from 0.04 (operational vegetation cover) to 0.85 (cleared / grading); inverters in commissioning have infant-mortality failure rates 3-5× operational baseline. Earthflow's Phase Analyzer verifies physical progress weekly via Sentinel-2; vulnerability parameters reflect the active phase.

6 · Stage 3 — Loss Quantification

Stage 3 aggregates per-peril damage ratios into portfolio loss metrics. Earthflow uses Monte Carlo stochastic simulation — 10,000 to 1,000,000 synthetic years per site — to construct the full loss distribution, from which Annual Average Loss (AAL), Probable Maximum Loss (PML), Tail Value-at-Risk (TVaR), and Exceedance Probability curve are extracted. Reinsurers can audit any quantile.

6.1 Monte Carlo Aggregation

6.2 Loss Metrics Produced

6.3 Exceedance Probability Curve Construction

The EP curve is the canonical loss-distribution visualization. For each simulated loss ℓ, plot the probability that annual loss equals or exceeds ℓ:

6.4 Net-of-Mitigation Overlay

Each mitigation factor toggled in the dashboard (stow / vegetation / glass / etc.) re-runs Stage 2 vulnerability with the factor applied and re-aggregates to a new EP curve. The visual overlay of "with mitigation" vs "baseline" curves quantifies the underwriting impact of each lever — typically the most persuasive interaction for a binding decision.

7 · Stage 4 — Financial Impact

Stage 4 applies policy and treaty terms — deductibles, limits, layers, reinstatement premiums, sub-limits, aggregates — to the Stage 3 loss distribution. The output is what the reinsurer cares about: ceded loss, attached loss, treaty-net AAL, indicated rate, and parametric trigger pricing.

7.1 Per-Event Payout Function

7.2 Indicated Technical Rate

7.3 Parametric Trigger Pricing

Parametric covers pay a pre-agreed amount when an objective trigger is breached — no claims adjudication, settle in 5 business days. Earthflow prices parametric layers using the same hazard distribution that drives the indemnity model, ensuring consistency.

Common Solar Parametric Triggers

7.4 Treaty Fit Metrics

For each site added to a cedant's existing treaty, Earthflow computes four treaty-fit scores against the cedant's incumbent book:

• Aggregate concentration — site PML as % of treaty PML budget
• Correlation to top-5 sites — peril-driver correlation matrix
• Diversification benefit — geographic + asset-class spread credit
• Layer attachment fit — whether the site PML falls within the treaty's attachment band

8 · Data Source Catalog

Earthflow ingests data from 70+ external sources spanning federal hazard datasets, satellite imagery archives, soil and topography surveys, weather reanalyses, and climate-projection ensembles. Every value displayed in the dashboard cites its source and last refresh — no black-box scores. This chapter catalogs the major sources by category.

8.1 Hazard & Climatology

8.2 Satellite Imagery & Detection

8.3 Soils, Terrain, Vegetation, Infrastructure

8.4 Solar Resource & Weather

8.5 Climate Projections

8.6 Data Provenance — 7-Level Fallback Architecture

Every computed field in Earthflow carries provenance metadata: source dataset, confidence level (high / medium / low), fallback level used (1–7), all sources attempted, and retrieval timestamp. The 7-level fallback chain ensures that no analysis fails for lack of data — the engine reports both the value and the level of evidence behind it. Example chain for the RUSLE K-factor (soil erodibility):

For a reinsurer, this matters because every score is auditable to its evidence level. A composite risk grade carrying mostly L1 inputs is materially more bindable than the same grade carrying mostly L5–L7 inputs. Earthflow reports the input-level distribution alongside the score itself.

8.7 Industry Loss Data & Benchmarks

• GCube Insurance Solar Loss Reports — annual claims data, hail share of solar loss costs, average claim size
• kWh Analytics Solar Risk Assessment — annual industry report (2024 edition: hot-spot 3.4× rise, inverter 59% of lost energy)
• EPRI BESS Failure Database — BMS defect rates, thermal-management findings
• NIST CSF — Cyber Security Framework for SCADA / OT audit baseline

9 · Earthflow Module Reference

The Earthflow platform consists of three computational layers: a Cloud Run analytics orchestrator, 19 Physics AI™ modules, and 8 satellite-detection models (the Earthflow Detect family). All outputs flow into Firestore and are exposed via Firebase Cloud Functions to the dashboard and to 25 Cirra agent tools.

9.1 Cloud Run Analytics Orchestrator

9.2 Risk Modules (14)

9.3 Earthflow Detect — Satellite Computer-Vision Family

Earthflow Detect applies computer-vision segmentation to Sentinel-2 (10 m) and NAIP aerial imagery (~1 m) to identify solar panels, substations, inverter buildings, transmission corridors, and vegetation encroachment. Detection has been validated across 650+ utility-scale solar installations. Grid analysis runs at 30 m default resolution; drone ingest pushes the floor to ≤10 ft. Monthly change detection feeds construction-progress tracking (Stage 1 hazard) and vegetation encroachment alerts (Stage 2 vulnerability, where 5–15% output reduction is typical in affected areas).

9.4 Cirra Agent — 27 Tools Across 7 Functional Categories

Cirra is the Physics AI™ autonomous agent that wraps the entire platform with a natural-language interface. The agent orchestrates 25 specialized tools across 7 functional categories (site assessment · weather & climate · detection & monitoring · portfolio & comparison · reporting & export · financial & grid · utilities), supports up to 7 autonomous tool-calling rounds per query, and queries a live Firestore database covering 1,200+ sites with 1,180+ fields per site. Site resolution uses Levenshtein fuzzy matching to bridge user-facing names and database keys. Underwriter-relevant tools include:

10 · Outputs for Underwriters

Every dashboard panel in the Earthflow Reinsurance Edition maps to a specific underwriting decision. This chapter walks the dashboard left-to-right and explains how each output is intended to be used.

10.1 Dashboard Panel → Underwriting Decision Map

10.2 Verdict Decision Tree

The composite grade and verdict are derived from a deterministic decision tree applied after the Monte Carlo aggregation:

10.3 ACORD-Aligned Submission Packet

The "Underwriter Submission (PDF)" export produces a packet that maps directly to ACORD reinsurance submission fields:

• Cover page — insured asset, capacity, COD, cedent, TIV, layer, period
• Composite risk grade + verdict + rationale
• 12-peril decomposition table (score, band, AAL, TVaR contribution)
• Loss economics (AAL, PML 1-in-100/250, TVaR 99%)
• Equipment pedigree (module, inverter, racking — make/model/Tier/cert)
• Recommended subjectivities (ordered by priority)
• Mitigation factor inventory
• Parametric trigger sheet (if structured)
• Treaty fit metrics (concentration, correlation, diversification, attachment fit)
• Methodology + data-source citations appendix

The "ACORD Bordereau (CSV)" export produces a per-site row matching ACORD's standard solar reinsurance schema, suitable for direct ingestion into treaty-placement workflows.

11 · Reinsurer Workflows

Different reinsurer profiles prioritize different sections of this methodology. Earthflow's dashboard exposes three lens-aware presentation modes (Equipment & Reliability, Catastrophe & Parametric, Full Risk View) — same underlying model, three vocabularies and three feature orderings.

11.1 Equipment & Reliability Lens — Workflow

For reinsurers whose underwriting DNA is condition-monitoring, equipment-breakdown, and energy-shortfall insurance. The workflow emphasizes operational perils, IoT signals, and predictive maintenance.

Cirra questions tuned to this lens

• "Walk me through the inverter health profile."
• "What's the MTBF on this site's equipment?"
• "How would energy-shortfall insurance respond to last year's underperformance?"
• "What predictive-maintenance signals are firing today?"
• "Is the BESS thermal management within spec?"

11.2 Catastrophe & Parametric Lens — Workflow

For reinsurers whose underwriting DNA is natural-catastrophe modeling, parametric trigger structuring, and PV warranty backing. The workflow emphasizes hail/SCS deep-dive, parametric pricing, climate-creep, and PV warranty status.

Cirra questions tuned to this lens

• "Why is hail the #1 peril here?"
• "Price me a $5M xs $5M hail layer at 1-in-50."
• "What stow protocol would shift this to bind?"
• "How does PV warranty backing change the rating?"
• "Show the SCS climate-creep trajectory through 2055."

11.3 Full Risk View — Workflow

For first-look prospects, captive analysts, and integrated reinsurers wanting all 12 perils in their natural priority order. No suppression — every panel visible at full prominence.

11.4 Bordereau Aggregation

For cedant-level rollup across multi-site treaties, Earthflow runs a single Monte Carlo across the entire portfolio rather than summing per-site outputs. This preserves correlation structure (e.g., a single hurricane track impacts multiple FL sites in the same simulated year) and produces materially different layer-attachment recommendations from per-site aggregation.

12 · Validation & Calibration

A cat model is only as good as its back-testing record. Earthflow's calibration program follows a four-step cadence — annual model recalibration, quarterly data refresh, per-event post-loss validation, and continuous sensitivity analysis.

12.1 Back-Testing Methodology

For each historical loss event with available perimeter/footprint data, the model is re-run with information available prior to the event and the resulting damage prediction is compared against observed claims:

• Wildfire — NIFC perimeter shapefiles · 10-year window · matches CA Cal-Fire and AZ MNF perimeters to within ±200 m typical
• Hurricane — NHC HURDAT2 · post-1990 · matches Saffir-Simpson category and surge depth at 50+ historical landfalls
• Hail — NOAA Storm Events + NEXRAD MESH · 15-year window · matches max-stone-size to within ±0.25 inch
• Flood — FEMA disaster declarations + NWS forecast vs observed gauges · 20-year window
• Equipment — manufacturer-disclosed failure data · cross-referenced to SCADA aggregator output

12.2 Industry-Loss Benchmarks

Model outputs are sanity-checked against published industry loss data to ensure pricing remains within market bands:

12.3 Calibration Cadence

• Quarterly — external data refresh (NOAA, NREL, FEMA, NIFC, NLDN, satellite imagery)
• Per-event — major-event post-loss validation pass within 30 days
• Annual — full model recalibration · vulnerability functions updated against new ASTM / IEC publications · CMIP6 climate ensemble refreshed
• Continuous — sensitivity analysis · automated regression tests on a held-out validation set

12.4 Sensitivity Analysis

For each composite score, Earthflow computes the partial derivative of composite with respect to each input — letting underwriters answer "what would it take to move this from C to B?". Common findings:

• Adding mid-clamp racking: typically +3 to +6 composite points
• Adding stow protocol: typically +5 to +12 composite points in hail-exposed zones
• BMS recertification: typically +2 to +4 composite points for BESS-co-located sites
• Quarterly EL imaging: typically +1 to +3 composite points for older sites
• 30 m defensible buffer (wildfire): typically +6 to +10 composite points in WUI zones

13 · Known Limitations

Every cat model has boundaries. Earthflow documents its known limitations openly — black-box vendors that don't publish theirs deserve more scrutiny, not less. The following limits apply as of this methodology version.

13.1 Geographic Scope

• Hazard modules cover CONUS + select Caribbean at full fidelity. International coverage (LATAM, EMEA, APAC) uses ERA5 + CMIP6 fallback at lower spatial resolution.
• FEMA flood zones are US-only. International flood hazard relies on JRC + HAND.
• NEXRAD MESH is US-only. International hail uses ERA5 hail-proxy + sparse ground reports.

13.2 Asset-Class Scope

• Primary calibration: utility-scale ground-mount PV (≥ 5 MW). Distributed / rooftop / floating PV are partially covered with reduced confidence.
• BESS support: outdoor LFP (full), indoor LFP (full), outdoor NMC (limited), indoor NMC (flagged as high-uncertainty).
• Hybrid solar + wind co-located not yet modeled (wind turbine vulnerability functions on roadmap).

13.3 Detection Model Limitations

• Earthflow Detect (panel) is calibrated for c-Si arrays. CdTe thin-film detection accuracy is lower (~75% recall) and flagged where applicable.
• Bifacial vs monofacial detection is not separated at present.
• Roof-mount and floating PV detection performance is lower than ground-mount.

13.4 Statistical Limits

• Monte Carlo standard error scales as 1/√N. Default N = 100,000 simulated years provides ±1-2% AAL precision and ±5-8% PML 1-in-250 precision. Tighter tolerance available via N = 1,000,000 at runtime cost.
• Peril correlation matrices are calibrated against the 50-year observational record. Climate-non-stationarity may shift correlation structure beyond the historical training window.
• BESS thermal-runaway tail distribution has limited historical data (industry is < 10 years old). Estimates rely on engineering models more than statistical fits.

13.5 Update Cadence

• Methodology document: revised annually (next: May 2027) and after any major model recalibration.
• Industry-benchmark calibration: refreshed against the annual kWh Analytics Solar Risk Assessment and the annual GCube renewables loss report.
• External data sources: refreshed per individual source cadence (see Chapter 8).