Implementation verification
Equations, units, numerical behavior, deterministic examples, and mass conservation.
Evidence record · updated August 13, 2026
Science should be inspectable.
TerrainLogic reports what has been verified, what agrees with a reference model, what has been tested against independent observations, and what remains uncertain. Those are different claims, and this page keeps them separate.
Reference-model parity
Agreement with RUSLE2, APLE, APEX, NRCS, USGS, or an analytical reference.
Independent validation
Predictions compared with observations not used to choose equations or coefficients.
01 / Erosion modeling
RUSLE2 parity is strong.
Field validation is a separate next step.
TerrainLogic’s erosion implementation is benchmarked against annual soil-loss results produced from NRCS RUSLE2 management, vegetation, operation, soil, climate, and site records. This tests compatibility with the reference model; it does not prove that either model exactly predicts measured erosion on every field.
National raw scalePearson r0.9993
National raw scaleR²0.9986
National benchmarkMBE+0.475ton/ac/yr
National log1p sensitivityPearson r0.9985
National log1p sensitivityR²0.9971
Locked release regression
199-case stability suite
A fixed, deeply investigated case set is rerun after erosion-engine changes to detect cases that move from passing to failing. It complements the broader national workbook; it is not merged into it.
- Strict within 20%
- 176 / 199 88.4%
- Practical tolerance
- 195 / 199 98.0%
- Outside both limits
- 4 material regressions
- Mean bias error
- -0.433 ton/ac/yr
- Pearson r
- 0.9973 raw annual A
- R²
- 0.9946 raw annual A
- Log1p sensitivity
- 0.9981 / 0.9962 Pearson r / R²
What is compared
TerrainLogic A versus RUSLE2 A
The 749-case national benchmark spans states, climate zones, crop and forage systems, construction templates, tillage, residue, grazing, manure, orchards, vineyards, and very low-loss systems. The separate 199-case locked regression set guards against release-to-release breakage.
- Same location and management identity
- Legacy RUSLE2 climate/event mode
- Annual soil loss in ton/ac/year
- Daily intermediate tables retained for diagnosis
Reference science
USDA RUSLE2 process documentation
The erosion engine follows the published RUSLE2 science for erosivity, erodibility, slope length and steepness, cover-management subfactors, residue, consolidation, roughness, and support practices, using NRCS database records where licensed for the installed product.
- USDA-ARS RUSLE2 Science Documentation
- USDA-NRCS RUSLE2 User Reference Guide
- NRCS management, vegetation, and operation records
- SSURGO soil properties and USGS elevation data
Important limitation
This is not measured-plot validation
A monitored runoff-plot soil-loss dataset held out by site is still required for an independent predictive erosion claim. Known parity differences include representative-event timing in some construction cases, reference-model slope-length limits, incomplete source inputs, and percent inflation where expected erosion is near zero.
Why two pass rates?
The strict rate reports cases within 20%. The planning tolerance also accepts a case when its absolute difference is no more than 0.25 ton/ac/year. This prevents tiny expected values from being called materially wrong solely because a very small absolute difference creates a large percentage. Both rates and the full denominator are reported.
What correlation and bias mean
Pearson r measures linear association; R² is from an ordinary least-squares fit with an intercept. They do not measure agreement with the 1:1 line and are not pass criteria—a consistently biased model can still correlate almost perfectly. Erosion MBE is the mean of TerrainLogic minus RUSLE2 in ton/ac/year: positive means average overprediction and negative means average underprediction. Raw and log1p correlations are therefore shown beside MBE and the absolute and percentage error buckets.
02 / Nutrient accounting
A process-by-process validation mosaic.
No single field dataset observes every nutrient pathway. TerrainLogic therefore uses independently sourced datasets for crop removal, phosphorus budgets, runoff, leaching, fixation, volatilization, and multiyear continuity, with eligibility decisions and exclusions preserved.
Crop nutrient removal
Independent harvest endpoints
Strict paired Transforming Drainage grain-N rows produce +9.33 kg N/ha mean bias error (MBE), 6.60% PBIAS, 19.53 kg N/ha RMSE, Spearman 0.850, and NSE 0.790. Additional LDMI, KBS, USDA, and open global crop datasets test N/P concentration, partitioning, and removal across crops and stages.
Phosphorus
Mass balance and observed loss
APLE equation conformance passes deterministic reference calculations. After harmonizing currently available study-period precipitation, the independent 15-record soil-derived MANAGE endpoint remains a pass at -12.61% PBIAS, correlation 0.860, and NSE 0.580 for total P. Applied-P field validation remains incomplete where required weather, runoff, placement, timing, or manure WEP inputs are absent.
Surface urea NH₃
Publication-aligned site-year test
A published process formulation was evaluated without fitting to the withheld observations. Across 18 complete site-years: MBE +0.52 kg N/ha, PBIAS +3.77%, MAE 6.01 kg N/ha, RMSE 7.14 kg N/ha, Pearson 0.732, Spearman 0.761, and NSE 0.485.
Mean bias error convention
MBE is the arithmetic mean of predicted minus observed values and is reported in the endpoint's physical units. Positive MBE means average overprediction; negative MBE means average underprediction. MBE is shown with MAE and RMSE because positive and negative errors can cancel.
ProcessEvidence statusRelease interpretation
Ledger / pool conservationVerifiedNumerical closure and state continuity
Crop N/P uptake and removalDomain validatedStrong crop-specific endpoints; not universal
Soil-derived total PDomain validatedIndependent no-application subset
Urea volatilizationCandidateInput-limited winter-wheat site-years
Runoff N/P and leachingIn progressProcess tests exist; national magnitude claim withheld
Total denitrificationInsufficient evidenceNo adequate direct field-magnitude holdout yet
Current claim: nutrient results are planning-level, mass-balanced estimates for screening and scenario comparison. They are not fertilizer recommendations, certified nutrient-management plans, compliance determinations, or regulatory validations.
03 / Engineering
Transparent preliminary design—not construction certification.
A frozen, independently calculated numerical suite now verifies the portions that can be validated without a constructed project. Site-specific authorization still requires the applicable state Field Office Technical Guide, documented hydrology, field survey, outlet design, utilities review, and the responsible professional’s approval.
Grassed waterways
NRCS CPS 412 / NEH 650
Six hydraulic cases from 5–100 cfs and 0.3–2.0% grade independently verify parabolic geometry, vegetation-dependent Manning capacity, boundary stress, freeboard geometry, and numerical convergence. Published NEH Chapter 7 example 2 is reproduced at 1.179 ft depth and 37.061 ft top width versus its rounded 1.2-ft and 37-ft values.
Terraces
NRCS CPS 600 planning workflow
Three graded and three level cases independently verify vertical interval, runoff-volume conversion, trapezoidal geometry, Manning capacity, level storage, and cut/fill conservation. The largest terrace equation difference is 0.11%, including reporting precision.
Boundary of use
Professional review remains required
TerrainLogic does not replace survey-grade elevations, state criteria, hydrologic documentation, permits, utility locating, construction staking responsibility, or licensed review where required.
What has not been validated
The numerical suite does not validate peak-discharge hydrology, DEM-derived drainage area or alignment, outlet adequacy, state-specific FOTG compliance, utilities, constructability, as-built performance, or long-term maintenance. Those items remain explicit REVIEW requirements and cannot inherit a PASS from the equation checks.
04 / Technical record
Methods, findings, and limitations.
The technical documents are versioned with the model and updated when an equation, source, dataset, or validation conclusion changes.
Erosion validation
Erosion Validation Protocol & Findings
National parity, locked regression results, pass rules, source science, known differences, and the independent field-validation gap.
Open document ↗ Master protocolTerrainLogic Scientific Validation Master Plan
Claim levels, anti-leakage rules, evidence matrix, acceptance metrics, and current conclusions.
Open document ↗ Process scienceNutrient Process Science
Public overview of represented processes, scientific provenance, verification boundary, and the reason implementation details remain confidential.
Open document ↗ Public validation recordNutrient Validation Protocol & Findings
Evidence hierarchy, holdout rules, acceptance gates, endpoint scores, failures, applicability domains, and source literature—without production architecture.
Open document ↗ EngineeringGrassed Waterway Verification
Preliminary-design checks, criteria, test process, limitations, and required field review.
Open document ↗ EngineeringTerrace Verification
Planning calculations, analytical checks, survey requirements, and certification boundary.
Open document ↗