Field notebook · v0.1 · 2026
SoilFlux
Vision

Healthier soil under
more of the world's farmland.

The soil under a farm is the most chemically and biologically complex centimetre of any production system on Earth. Treating it like the living, place-specific organism it is — at scale, with credible models — is how we get out of the cycle that got us here, and how biological products finally get adopted on evidence rather than hope.

The AlphaFold of soil.

One global model of the soil — its state and its dynamics — queryable anywhere on Earth from public data, sharpened by cheap field sensing. The same shift AlphaFold brought to biology, brought to the ground beneath agriculture.

The analogy where it counts

AlphaFoldSoilFlux

Unified a fragmented field — one model, not one experiment per protein

Unifies fragmented soil science — one model, not one predictor per property

Sequence → 3D structure, globally

Coordinates (+ sensors) → full soil state + dynamics, globally

Foundation model across all proteins

Foundation model across modalities, regions, and time

Per-residue confidence (pLDDT)

Per-prediction spatial-CV confidence bands

Changed the field to predict-then-verify

Cheap field sensing closes the loop — predict-then-verify in your hand

Harder than protein folding, in one specific way

Soil is not a static structure to be predicted once. It is a living system that changes with climate, management and time — and the hardest part is not any single property. It is what happens when a native microbial community meets an introduced one. Those interactions produce behaviour that none of the parts predict on their own: emergence. A model that only ever sees a snapshot cannot see them at all. Capturing the dynamics well enough to anticipate them is the bar we are building to, and we are a long way from clearing it.

Four questions, one model

Today, each question needs a different tool. The bet is that state, dynamics, decision and measurement — unified in one representation — is worth more than the sum of four siloed predictors. Each constrains and sharpens the others.

  1. 01State

    What is this soil?

    Chemistry, biology, physical structure, microbial community, pathogen pressure — at any coordinate on land.

  2. 02Dynamics

    How does it change?

    Soil carbon, nutrient cycling, microbial succession under climate and management. Forecasts, not just snapshots.

  3. 03Decision

    What should I do?

    Inoculant fit, crop suitability, intervention timing — with calibrated confidence and an evidence chain you can defend.

  4. 04Measurement

    What is actually here?

    Refine the global prior with measured spectra, microscopy, and sensor streams from the site itself. The local truth completes the picture.

The engine is the soil. The product is the lifecycle.

The same global soil model becomes biological lifecycle intelligence: one system that walks a product from a candidate strain to a field result to a registered market — asking, at each step, will it establish, persist, and transfer here, with calibrated confidence and an evidence chain.

  1. 01Screen

    What do I test, and where?

    Rank candidate strains by where they should establish, persist, and fit the target — before committing the trial spend.

  2. 02Expand

    Where do I register, and in what order?

    Climatic and soil comparability to a validated region, fused with register status — the mutual-recognition question, answered as evidence.

  3. 03Explain

    What can I claim?

    Every result with its evidence and its uncertainty, in plain language — a spokesperson for the honest model, never inventing confidence.

  4. 04Validate

    Did it hold?

    Predict, trial, observe, update — each real outcome sharpens the model, and only where the data is representative.

Staged, causal, probabilistic

Efficacy is predictable in principle — but it is causal, not correlational, and it earns its confidence one controlled stage at a time. Asked to vouch for a new crop or country, we don't guess: we score how comparable the new environment is — climate and soil, separately — and say how far the strain's behaviour should transfer, with the minimal trial that would confirm it. Transferability is not efficacy, and we say so.

Same vision. Two lenses.

If you're taking a biological from lab to field

The evidence layer for where your biological should establish and hold — defensible to a buyer, a registrar, or your own field walk.

Today you triangulate across a lab panel, last year's data, a consultant's instinct, and a vendor's product sheet — and still can't say whether the strain will hold up in the field. One model of the soil itself collapses that triangulation, queryable at any coordinate, with the confidence on every claim spelled out — because the barrier to adoption was never the prediction, it was trust in the claim.

Efficacy gets a biological into the lab; fit — regulatory, agronomic, application, economic — is what gets it adopted. Concretely: a developer with a candidate strain sees where it should establish, persist, and be worth trialling before committing the spend — and how far that result should carry to the next crop, and the next market.

If you back foundation models in deep verticals

A foundation-model bet on the ground beneath the food system — defensible by the law of conservation.

The bet has the shape of Harvey for legal or Hippocratic for medicine — a foundation model that owns one critical interface in a vertical that has historically been served by fragmented, siloed predictors. The differentiation is structural: physics-grounded constraints (conservation laws) bound what the learned model is allowed to believe, which means accuracy compounds rather than drifts as data scales.

The wedge is the upstream biological developer — the team running field-trial programmes it cannot afford to repeat, and that can defend a €0–40k co-development pilot, and €25–90k a year thereafter, to de-risk which strain, which site, which market is worth the spend. The same model carries the sharpest expansion: regulatory and lifecycle intelligence for biological inputs — soil persistence and climatic comparability, the evidence that informs where a product can persist, register, and expand. Agronomy and grower-tier field sensing stay long-horizon, and carbon quantification stays optionality.

Request investor materials

§ 06 · In closing

“A single dynamic model of the soil, queryable anywhere on Earth from public data and sharpened by cheap field sensing. What your soil is, how it will change, and what to do about it — with calibrated confidence and an evidence chain.”

The SoilFlux team · 2026

Work with usMeet the team