# TerrainLogic Erosion Validation: Protocol and Findings

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

## Purpose and claim boundary

TerrainLogic estimates conservation-planning sheet and rill erosion using published RUSLE2 science and authoritative site, climate, soil, and management data. The present evidence establishes numerical verification, release stability, and strong agreement with RUSLE2 reference outputs. It does **not** establish independent predictive accuracy against measured runoff-plot soil loss.

The evidence is reported in three separate tiers:

1. **Implementation verification:** units, numerical invariants, conservation checks, and deterministic reference examples.
2. **Reference-model parity:** annual TerrainLogic soil loss compared with annual RUSLE2 soil loss for equivalent cases.
3. **Independent observational validation:** predictions compared with measured field soil loss from sites not used to select or adjust the method. This tier remains open.

Agreement with RUSLE2 is evidence of RUSLE2 compatibility—not proof that either estimate equals measured erosion.

## Scientific and data basis

The public scientific basis includes USDA-ARS RUSLE2 Science Documentation, the USDA-NRCS RUSLE2 User Reference Guide and database conventions, Agriculture Handbook 703, authoritative NRCS soils and management records, and published erosion-factor science.

TerrainLogic preserves sufficient factor-level audit information for a reviewer to identify whether a difference arises from climate, soil erodibility, topography, cover-management, or support practice. Exact production architecture, state handling, database adapters, fallback logic, internal artifacts, and correction history are maintained as confidential engineering records.

## National parity benchmark

The completed national benchmark contains 749 eligible cases with a RUSLE2 expected annual soil-loss value and the inputs required by TerrainLogic.

| Measure | Result |
|---|---:|
| Cases within 20% | 600 / 749 (80.1%) |
| Additional cases outside 20% but within 0.25 ton/ac/year | 107 |
| Practical planning tolerance (within 20% **or** 0.25 ton/ac/year) | 707 / 749 (94.4%) |
| Material failures outside both limits | 42 / 749 (5.6%) |
| Mean bias error, TerrainLogic minus RUSLE2 | +0.475 ton/ac/year |
| Pearson correlation, raw annual A | 0.9993 |
| R-squared, raw annual A | 0.9986 |
| Pearson correlation, log1p sensitivity | 0.9985 |
| R-squared, log1p sensitivity | 0.9971 |

Mean bias error is the mean of TerrainLogic minus RUSLE2 and retains units of ton/ac/year. Positive MBE indicates average overprediction. Correlation and R-squared describe association and must not be interpreted as agreement with the one-to-one line; they are therefore reported with bias and practical error buckets rather than used as standalone pass criteria.

## Locked regression suite

The separate locked 199-case suite measures release stability. It is not combined with the national denominator because cases overlap and the two suites answer different questions.

| Measure | Result |
|---|---:|
| Cases within 20% | 176 / 199 (88.4%) |
| Additional cases outside 20% but within 0.25 ton/ac/year | 19 |
| Practical planning tolerance | 195 / 199 (98.0%) |
| Material failures outside both limits | 4 / 199 (2.0%) |
| Mean bias error, TerrainLogic minus RUSLE2 | -0.433 ton/ac/year |
| Pearson correlation, raw annual A | 0.9973 |
| R-squared, raw annual A | 0.9946 |
| Pearson correlation, log1p sensitivity | 0.9981 |
| R-squared, log1p sensitivity | 0.9962 |

The regression suite is rerun whenever an erosion-model change is proposed. Release acceptance requires preserving established passing behavior while evaluating any corrected cases.

## What the parity evidence covers

The benchmark spans a national range of climates, soils, slopes, management families, vegetation, residue conditions, tillage and field operations. It tests whether TerrainLogic reproduces the combined annual effects represented by the RUSLE2 reference cases.

The evidence supports a planning-level compatibility claim over the tested domain. It does not establish accuracy for concentrated-flow or gully erosion, wind erosion, streambank erosion, mass wasting, construction sites outside the represented managements, or conditions beyond the data and science domain.

## Known sources of parity difference

- Some legacy construction cases use representative erosion-event timing while TerrainLogic resolves legacy climate information differently. Annual factors may agree while timing-weighted annual loss differs.
- Reference-interface limits can prevent an identical entry for unusually long slope profiles.
- Extremely small expected values can produce large percentage differences despite negligible absolute differences.
- Workbook reference values have been corrected when independent review found an erroneous climate, slope, management, or reported result. These are reference-data corrections, not model improvements.
- Source records can be incomplete or ambiguous. Such cases are retained and classified rather than silently removed from the denominator.
- Reference-model parity can conceal limitations shared with the reference science.

## Independent observed-erosion validation still required

A future independent field-validation package must freeze monitored runoff-plot or field observations before scoring; hold out entire sites or studies; preserve rainfall, erosivity, soil, slope, management, support-practice, and uncertainty information; prohibit calibration rows from being relabeled as validation; and publish row-level predicted and observed values.

Required statistics include N, PBIAS, MBE, MAE, RMSE, median absolute error, Pearson correlation, R-squared, Spearman correlation, NSE, KGE, log-scale sensitivity, uncertainty coverage where available, and stratification by soil, slope, climate, management, and observed-loss magnitude.

## Current public claim

**TerrainLogic demonstrates strong planning-level agreement with RUSLE2 across national and locked regression benchmarks.** The result is reference-model parity and software-verification evidence, not independent field-measured erosion validation.