Research Programmes, Papers and Reports

Choose the audience that matches your question

The page contains three selectable readings of the same subject.

Engineers turn the proven idea into real code. The same tests that proved the research now prove the product. You get the benefit without ever reading the paper.

Before research becomes a feature, we publish it so others can check our method, our data, and our conclusions. Open scrutiny catches mistakes and makes the work stronger.

We build prototypes, run experiments, and measure the results against real data. If something does not work, we write that down too. A negative result is still a result.

Search is too slow. Answers are inconsistent. Energy bills are too high. We turn the thing people complain about into a precise question we can study.

Replay is a testable property under a pinned build

Proposed 60Hz timestep, seeded randomness, and a typed input-log path

Candidate kernels checked against bounded obligations

SMT checks with Z3 and CVC5 plus a Cranelift codegen path; validation pending

Candidate example: a three-multiply route for complex multiplication

Multiplication-count comparison target (not a published benchmark)

Sparse dispatch is proposed to skip inactive cells; per-tick behaviour requires validation

CPU-saving comparison target (not a published result)

Replay conditions are testable; full reproducibility remains to validate

Declared rules are intended to support tick-by-tick replay; validation is pending

Each figure should trace to the source run and measurement record; the evidence status is part of the review.

Forge's candidate backend scope includes x86-64, RISC-V, WebAssembly, and GPU; coverage and quality require validation.

Jacquard source is private and available by request. Forge remains research and is not open.

Claims, evidence, and the number behind each.

Reproducibility is a design goal for these research tracks. Each row should pair a candidate claim with the stream, source, and measurement record that produced it. Read it as an evidence sheet with validation status, not as a brochure.

Numbers shown here are report fields or comparison targets; check the source record and access terms before treating one as a measured result.

AI is everywhere now, but most people do not know how it works, where their data goes, or whether they can trust the answers. Our research exists to fix that. We work on three problems that matter to everyone who uses technology.

Three questions that guide everything we build.

We study how to make AI explain its reasoning, give the same answer every time, and say so when it is unsure. You should never have to trust a black box. You should be able to see why it said something and decide for yourself.

Running AI can use a lot of power. We look for ways to get the same answer using far less. When AI uses less electricity, it costs less and pollutes less, so smaller organisations and public services can afford it too.

We study how to build AI that stores and handles your information inside European borders, under European law. That means it is protected by GDPR, not handed to foreign surveillance, and looked after by people you can hold to account.

Each stream publishes openly so others can check the work behind the claim.

Dweve research is organised around four strategic areas that map directly to product capabilities. Each is the public research layer that supports a commercial claim. Hover a stream to see what it underpins.

Four streams. One mission: sovereign, efficient, trustworthy AI.

Research into data-residency frameworks and transfer mechanisms that satisfy Schrems II. This underpins our deployment guarantees that data stays inside European borders, under European law.

Research into homomorphic encryption, secure aggregation for federated training, and zero-knowledge proofs for model verification. This is the foundation under the federated learning inside Mesh, our circular compute product.

Cost control and sustainability reporting

Research into quantisation and hardware-specific kernel work that cuts the power a model needs without degrading the output. Less electricity means lower running cost and a smaller footprint.

Research into fixed-point arithmetic, seeded randomness, and platform-independent execution so a model gives the identical answer on every machine. This underpins the reproducibility you need for an audit.

We publish intermediate results, failed experiments, and revised hypotheses. Research as an open engineering process.

Jacquard remains a private research track. Any transfer to Fabric, Nexus, or Aura requires separate validation.

Jacquard replay is a research property, Mycelia replay is a test condition, and Forge checks bounded obligations in an experimental programme.

Each stream carries an evidence status and a boundary around its current claims.

model scopes: perception, chemistry, evolution

A digital-world model with explicit local rules and energy costs. The report supports deterministic replay as a testable property; broader behaviour claims remain future validation work.

Forge explores candidate kernel implementations through program synthesis. Its design uses Cranelift-based measurement and SMT checks for bounded obligations; the 2025 report does not establish production readiness or benchmark performance.

Private source repository, access by request

A browser-based world engine that agents can build through a typed MCP surface. Fixed timestep, seeded randomness, and replay from a seed and input log are research properties to evaluate. The source repository is private.

Our research programmes turn difficult systems questions into inspectable apparatus: deterministic worlds, synthesis pipelines, and artificial-life laboratories. The output is not a demo alone, but a route another engineer can reproduce and challenge.

Jacquard studies deterministic world semantics, Forge studies program synthesis, and Mycelia studies artificial life. Each programme has its own apparatus, result type, and failure boundary.

Open the implementation story behind each programme

Each programme owns its state and evidence

A typed world engine with fixed-step execution, explicit transitions, and replayable agent traces.

A synthesis pipeline that searches candidate programs, lowers them through Cranelift, measures them, and checks bounded correctness obligations.

A deterministic artificial-life laboratory for local rules, resource pressure, lineage, and emergent structure. Its runs remain research evidence, not a scientific conclusion by themselves.

A programme starts with an observable claim, controlled variables, perturbations, and a stop condition. The specification determines the state we retain and the counterexample that can reopen the result.

The experiment contract precedes the implementation

Unspecified state cannot support a reproducible claim

Does the mechanism preserve its invariant across controlled perturbations

The invariant, perturbation family, and observation boundary are committed before the run.

Typed entities, transitions, clocks, and seeds.

Cases vary independently against a fixed baseline.

Invariants and failure thresholds run as checks.

No inference outside the versioned state space

Question, apparatus, run, challenge, and transfer are connected by versioned inputs and outputs. Select a gate to inspect the contract at that boundary.

Define the invariant, observation surface, controlled state, perturbation family, and stop rule.

Implement the minimal deterministic apparatus with explicit clocks, seeds, schemas, and event boundaries.

Execute the protocol while retaining inputs, transitions, timings, outputs, and environment identity.

Generate counterexamples across seeds, edge cases, scale, and altered assumptions; attach failures to the result.

Export the surviving mechanism, API boundary, reproduction packet, and operational constraints as one handoff.

The receipt binds source revision, build identity, protocol, state, observations, checks, and limits. A second engineer can recover the route without guessing which environment made it work.

A result without its environment is incomplete

revision + seed + protocol → identical trace hash

A research result crosses into product engineering only with an explicit interface, reproducible behaviour, benchmark context, and a failure mode the destination can operate.

The ledger distinguishes a result from an integration

Capability, contract, evidence, envelope

Each programme carries its own architecture, apparatus, and result shape. The paper shelf provides the longer theoretical and empirical context around those systems.

Follow code-shaped questions into their programme

Programme details first, cross-cutting papers beside them

Inspect typed worlds, deterministic ticks, transitions, and replay traces.

Inspect candidate generation, lowering, measurement, and solver-backed constraints.

Inspect local rules, lineage, scarcity, and deterministic emergence.

Read the cross-cutting technical arguments

Long-form work connects deterministic execution, neurosymbolic methods, representation, efficiency, and evidence architecture.