Physics is
the specification.
Code is the proof.

Diffusion MRI can see inside living tissue without a scalpel. It can infer the diameter of an axon, the density of cell membranes, the permeability of a soma wall — all from the way water molecules move. This is one of the most remarkable capabilities in modern medicine.

And the field has been working without the infrastructure it deserves.

Not for lack of trying. The scientists who built this field are rigorous, inventive, and deeply committed to getting it right. Validation has happened — but bespoke, each group building its own simulation, its own ground truth, its own interpretation of what the model should recover. Brilliant work, done in isolation, with tools that were difficult to build and harder to share. The supporting code is often unpublished, the methodology unreproducible not by choice but by circumstance. What survives in the literature is the claim. What does not survive is the chain of evidence behind it.

The result is a field where the same model can be validated six times in six different ways, with six different simulators, producing six subtly different answers — and no shared language to reconcile them. Where scope boundaries exist in the author's intuition but not in the publication. Where a protocol validated for one tissue type migrates quietly into another without anyone having established whether it should.

This is not a failure of the people. It is a failure of infrastructure. And infrastructure is exactly what has become possible to build.

The specification exists.

Physics governs water diffusion in biological tissue completely and without ambiguity. The Bloch-Torrey equation is not a hypothesis. Brownian motion in restricted geometries is not a conjecture. These are the specification. Everything else — every model, every protocol, every clinical claim — is an implementation that either satisfies the specification or does not.

Monte Carlo simulation is the compiler. Given a substrate geometry, a permeability, a relaxation time, and an acquisition waveform, simulation produces ground truth signal with no approximation beyond the physical assumptions themselves. It does not negotiate with analytical convenience. It does not care about publication deadlines.

A claim that cannot survive Monte Carlo validation within its stated scope is not a validated claim. It is a hypothesis.

dmrai-lab is building the infrastructure to make this distinction enforceable.

dmipy implements three nodes of one closed loop: simulation — Monte Carlo substrate generation and signal synthesis, grounded in biologically cited parameter distributions; fitting — JAX-native model inference with explicit scope boundaries and bias characterization against ground truth; and design — acquisition waveform optimization that incorporates both analytical sensitivity and Monte Carlo-derived bias simultaneously, closing the loop between what we can measure and what we claim to measure.

This is not three tools. It is one pipeline. The same substrate that generates training data also validates the model that fits it and constrains the waveform that acquires it.

The compendium is the output.

Every validated result produced by this pipeline becomes a node in a public, machine-readable knowledge graph. Each node carries its physical assumptions, its scope boundaries, its Monte Carlo evidence, and its dependency on every node above it in the chain. A reader can trace any claim backward to the physics that supports it — or discover that no such chain exists.

This is not a database. It is a formal epistemological framework. The first one this field has had.

This is what the AI era makes possible.

A single researcher with the right stack can now implement, simulate, validate, and publish results that previously required a laboratory, a cluster, and a team. The bottleneck was never intelligence. It was infrastructure.

dmrai-lab is one researcher and a physics engine. The physics does not care about affiliation. We are building in the open. The code is public. The compendium is public. The validation pipeline is reproducible by anyone with a GPU and a pip install.

If your model survives the compiler, it survives. If it does not, the compiler tells you exactly why, and you can fix it. That is how software has worked for fifty years. It is time for science to catch up.

The hallucination problem has a physical solution.

dmipy is built agent-first: the majority of implementation work is done by AI agents operating autonomously across simulation, fitting, and design. This makes extraordinary velocity possible — but it introduces a failure mode that cannot be ignored. An agent can write code that is internally consistent, passes its own tests, and is completely wrong. The fix is not better prompts. The fix is physics.

Every feature in the dmipy ecosystem must satisfy success conditions stated entirely in physical terms — conditions that can be verified from two completely independent directions. Agreement between those directions is not a unit test. It is a proof.

Direction one: the analytical equation.

For every physical phenomenon modelled in dmipy, the field has derived closed-form solutions under well-defined assumptions. The Kärger model gives the exchange signal for permeable membranes as a function of permeability κ and diffusion time. The van Gelderen / Neuman solutions give the restricted diffusion signal inside cylinders and spheres as a function of radius, intrinsic diffusivity, and the acquisition parameters δ, Δ, and q derived from the gradient waveform. These equations encode decades of physical intuition. They are exact — within the regime for which they were derived.

An agent implementing, say, the cylinder restriction model receives the analytical signal equation as its specification. The implementation must reproduce it. If it does not, the agent has hallucinated something — regardless of how plausible the code looks.

Direction two: Monte Carlo from G(t).

The second direction makes no approximations beyond the physical substrate itself. Given a geometry — a packed array of cylinders, a sphere, a permeable membrane — and a free gradient waveform G(t), Monte Carlo simulation propagates thousands of spins through that geometry via Brownian motion, accumulates their phase under the applied field, and reads out the echo signal. No analytical shortcuts. No assumption about pulse shape. G(t) is the input in full.

This is a completely different computational path. It does not know the analytical equation exists. It does not care whether δ and Δ are well-defined for the waveform being tested. It just runs the physics.

Convergence is correctness. Divergence is knowledge.

When the analytical equation and Monte Carlo agree — across substrate geometries, diffusion times, b-values, and waveform shapes — the implementation is validated. Not just tested: validated against physical ground truth from two independent directions. Any AI-introduced error that survives one direction is extremely unlikely to survive both.

When they diverge, something more valuable than a passing test emerges: a precise characterisation of the analytical equation's scope boundary. The Kärger model breaks down at short exchange times. The Neuman cylinder solution breaks down at high b-values and non-PGSE waveforms. Monte Carlo does not break down — it simply reflects the physics of that regime. Mapping where the analytical solution departs from Monte Carlo ground truth is not a validation failure. It is a scientific result: the exact envelope within which a published model is trustworthy, encoded directly into the compendium.

This is the architecture that makes agent development safe.

Each feature ships with a physics-grounded benchmark: a set of parameter regimes where analytical and Monte Carlo must agree to within a stated tolerance, and a set of departure curves that characterise the boundary beyond it. An agent that hallucinates an implementation will fail the benchmark. An agent that ships a correct implementation extends the compendium. The physics enforces the standard. The agents do the work.

In practice, the heavy lifting across the entire dmipy ecosystem — simulation engine, fitting framework, acquisition design, tractography, preprocessing, and compendium pipeline — has been done by agent teams running Anthropic's Claude Sonnet 4.6. The role of the human researcher is to set direction, state the physics, and hold the agents accountable to it. This is not a workflow where AI assists. It is a workflow where AI builds and physics audits.