From living complexity
to testable intelligence.
Soil is a dynamic system. A credible model must represent uncertainty, transfer carefully, and be validated against outcomes—not just historical correlations.
Six interdependent dimensions
Biology
Microbial communities, roots, decomposition and nutrient cycling. Genetic potential is not the same as measured activity.
Soil intelligence must preserve context across all six dimensions.Four research programs
Low-cost biological inference
Can inexpensive spectroscopy, field measurements and environmental context estimate biological functions that currently require costly assays?
Transferable representations
Can multimodal models learn useful structure across crops, soil types and regions without hiding domain shift?
Causal intervention science
Can controlled field experiments distinguish the effect of management practices from weather and site differences?
Soil → food → health
Can we establish the links from agricultural intervention to crop composition, dietary exposure and clinically meaningful outcomes?
Soil health has no universal single indicator. Microbial DNA is not equivalent to measured function, satellite imagery does not directly reveal subsurface biology, and better soil does not automatically establish better human health.
Reference starting points
For independent scientific grounding, consult USDA NRCS soil health guidance ↗, FAO's soils portal ↗, and peer-reviewed research on soil spectroscopy, microbiomes and causal agronomy. This website presents hypotheses, not validated product claims.
The soil is an ecosystem, not an ingredient list.
Texture, pores, organisms, roots, moisture and chemistry interact; an isolated indicator cannot describe the system’s behavior.
- 01
Management changes the physical environment and the availability of resources.
- 02
Water, oxygen and roots constrain microbial processes and nutrient transport.
- 03
Organisms transform organic matter and nutrients at different rates.
- 04
Plants respond through growth, stress and yield under local weather conditions.
What would count as evidence?
- Paired chemical, physical and biological assays with documented methods
- Repeated samples across depth, season and management history
- Yield and economic outcomes tied to the same geospatial units
What could break the thesis?
- Measurements are too variable or costly to change practice
- Conventional agronomy explains the outcome just as well
Which soil functions are both measurable and actionable?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.
What sampling resolution changes a real decision?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.
Which properties remain stable across seasons?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.
The measurement problem comes before the model.
A sophisticated prediction cannot repair unknown sampling bias, incompatible laboratory protocols or missing provenance.
- 01
Define a soil unit, depth, timestamp and collection protocol.
- 02
Collect independent observations and record uncertainty.
- 03
Normalize across instruments and laboratory methods without erasing differences.
- 04
Test repeatability before using observations as model labels.
What would count as evidence?
- Replicate and blinded sampling
- Inter-lab comparison and calibration records
- Explicit missing-data and measurement-error models
What could break the thesis?
- Measurement noise overwhelms treatment effect
- Costs exceed incremental decision value
What is the minimum trustworthy observation package?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.
Who certifies the method and its uncertainty?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.
Can low-cost proxies replace expensive assays?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.
A soil foundation model must earn its generality.
A model trained in one region may fail when climate, soil texture, crops or management regimes change.
- 01
Fuse imagery, weather, terrain, management and laboratory evidence.
- 02
Pretrain representations while preserving location and time context.
- 03
Test against mechanistic and conventional agronomic baselines.
- 04
Measure uncertainty and out-of-distribution failure on withheld geographies.
What would count as evidence?
- Geographically and temporally held-out evaluation
- Ablations for every expensive data modality
- Calibration and error analyses for minority soil types
What could break the thesis?
- Local models outperform the foundation model at lower cost
- Uncertainty is too high for high-stakes recommendations
Which tasks genuinely benefit from shared representations?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.
Where must mechanistic science constrain machine learning?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.
How will the model admit ignorance?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.
Food Is Medicine is a chain of separate hypotheses.
Changes in soil do not automatically imply clinically meaningful changes in human health.
- 01
Test whether agricultural intervention changes soil function.
- 02
Measure whether that change alters edible crop composition.
- 03
Establish actual consumption, bioavailability and exposure.
- 04
Evaluate clinical outcomes with appropriate study designs and confounder controls.
What would count as evidence?
- Controlled agronomic comparisons
- Independent compositional and dietary assessments
- Human studies designed for meaningful clinical endpoints
What could break the thesis?
- Crop differences disappear after storage and preparation
- Observed nutrient changes do not translate to health outcomes
Where is the weakest causal link?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.
Which outcome is worth testing first?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.
How can claims avoid outrunning evidence?
Do not assume the answer. Identify the relevant stakeholders, competing explanations, baseline measurement, and a test that could change the decision.