# TerrainLogic Nutrient Validation: Protocol and Findings

**Public scientific record | August 13, 2026**

## Current claim

TerrainLogic provides **planning-level, mass-balanced nutrient accounting for conservation screening and management-scenario comparison within documented applicability domains**. It is not a fertilizer recommendation, certified nutrient-management plan, compliance determination, or universal field-scale nutrient-fate model.

## What was tested

The validation program separates:

1. numerical mass conservation and multiyear state continuity;
2. conformance with cited published science;
3. predicted-versus-observed field performance; and
4. management-treatment direction.

Passing an equation or conservation test does not automatically establish field accuracy. Each claim is assigned its actual evidence tier: independent observational validation, reference/equation verification, implementation verification, diagnostic comparison, or unsupported.

## Anti-leakage protocol

- Entire studies, sites, years, or experiments are held out where needed.
- Random-row splitting is prohibited when rows share site or treatment structure.
- Validation targets are not reused as inputs or substituted for missing drivers.
- Sources and eligibility decisions are frozen before scoring.
- Missing chronology, units, or required drivers cause exclusion—not outcome-derived repair.
- Passes, failures, exclusions, and applicability limits are all reported.

## Acceptance criteria

Strict numerical endpoints use predeclared planning gates: |PBIAS| no greater than 20%, NSE at least 0.50, KGE at least 0.50 where supported, and at least 75% correct direction for paired management comparisons. Reports also include N, mean bias error (predicted minus observed), MAE, RMSE, Pearson correlation, R-squared where appropriate, and Spearman correlation.

## Principal findings

### Verified accounting

Mass-conservation, boundary, regression, and state-continuity tests pass. An authored ending state can become the next cycle's starting state without creating or deleting accounted nutrient mass. This establishes a coherent ledger, not independent field accuracy.

### Crop nitrogen removal

- Transforming Drainage strict paired corn/soybean grain-N subset: N=112, PBIAS +6.60%, RMSE 19.53 kg N/ha, Spearman 0.850, NSE 0.790.
- Global maize grain-N: N=341, PBIAS +8.22%, Spearman 0.903, NSE 0.697.
- Global wheat grain-N: N=340, PBIAS -5.44%, Spearman 0.886, NSE 0.649.
- Global rice grain-N: N=111, PBIAS -13.45%, Spearman 0.881, NSE 0.637.
- U.S. corn whole-plant N: N=116, PBIAS +10.88%, RMSE 24.64 kg N/ha, Spearman 0.917, NSE 0.780.

These endpoints support planning estimates within the tested crop, biomass, geography, and management domains.

### Crop phosphorus removal

- Global maize grain-P: N=192, PBIAS -10.61%, RMSE 7.93 kg P/ha, Spearman 0.792, NSE 0.558 (pass).
- Wheat grain-P: N=104, PBIAS +5.70%, Spearman 0.610, NSE 0.113 (bias passes; efficiency fails).
- Rice grain-P: N=72, PBIAS -9.10%, NSE 0.363 (bias passes; efficiency fails).
- Whole-shoot corn P: N=116, PBIAS +5.38%, RMSE 6.26 kg P/ha, NSE 0.260 (bias passes; efficiency fails).

### Residue return

Residue nitrogen is better supported than residue phosphorus. Rice residue-N passes both screens (+2.39% PBIAS, NSE 0.684); wheat has acceptable magnitude but modest efficiency (-6.62%, NSE 0.298). Global maize and soybean residue-N narrowly miss the 20% magnitude gate. Organ-confounded residue-P datasets do not support one universal correction, so residue-derived P is reported with greater uncertainty.

### Phosphorus transport

The independent soil-derived MANAGE total-P subset passes at PBIAS -12.61%, correlation 0.860, and NSE 0.580. Broader total-P and particulate-P collections do not all pass. Applied-P comparisons remain ineligible where runoff, erosion, soil-test method, placement, timing, weather, or manure water-extractable P are missing.

### Tile nitrate and leaching

One input-documented eight-row diagnostic gives PBIAS -6.32%, RMSE 4.25 kg N/ha, Spearman 0.786, and NSE 0.575. A separate 55 plot-year multisite screen fails (PBIAS -58.32%, RMSE 22.88 kg N/ha, NSE -0.583). TerrainLogic therefore does not claim generalized field-scale leaching validation.

### Ammonia volatilization

Independent surface-versus-injection comparisons reproduce the observed direction in 10/10 environments. An unfitted 18-site-year surface-urea planning endpoint gives 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. The result supports bounded planning use but narrowly misses the NSE gate and is concentrated in winter-wheat conditions.

### Fixation and denitrification

Independent soybean and clover evidence supports expected fixation response direction, but available inputs do not support a universal field-scale fixed-N magnitude claim. Denitrification has equation and nitrogen-conservation verification only; N2O observations cannot validate total denitrified N. Total denitrification remains a screening estimate.

## Failed and unsupported claims

The release claim excludes universal nitrate-leaching magnitude, total denitrification magnitude, universal residue-P return, complete applied-P budgets without required drivers, fertilizer recommendations, regulatory use, and predictions outside represented crops, soils, climates, and application methods.

Alternative published formulations were independently evaluated and did not improve performance sufficiently to replace the production method. Detailed counterfactuals and implementation experiments are confidential engineering records.

## Public scientific sources

Sources include USDA-ARS APLE and MANAGE resources; USDA-NRCS National Agronomy Manual and CPS 590; Transforming Drainage, LTAR, and drainage-network datasets; Morel et al. (2020), DOI `10.1016/j.eja.2020.126174`; Djaman and Irmak (2018), DOI `10.4081/ija.2017.958`; King and Torbert (2007), DOI `10.1017/S0021859607006946`; Sigdel et al. (2024), DOI `10.1002/agj2.21644`; and Hussain et al. (2025), DOI `10.1002/jeq2.70000`.

Exact production architecture, coefficients, state/pool routing, fallback behavior, internal artifacts, debugging history, and failed implementation recipes are intentionally excluded from this public report.