RESEARCH · ALL CLUSTERS · ALL THREADS

Four research clusters

Privacy, safety, accessibility, and open model development.

Each cluster targets open problems with no settled solution. Preprints, datasets, tooling, and negative results are published as they become available.

04
research_clusters
16
active_threads
Est. 2026
lab_founded
100%
open_access

FEATURED.RESEARCH

Three open problems.

A.03 · CRYPTOGRAPHY

Cryptographic Trust Chains

Generalising the Wyvern trust chain into a framework any on-device model can adopt. Formal verification of model provenance and inference integrity, proving what ran, where it ran, and what it accessed.

Can inference integrity be verified without access to weights?
2-3 year horizon · Cluster A · Privacy-Preserving AI
B.03 · POLICY

EU AI Act Alignment

Translating technical safety properties into compliance-ready frameworks engineers can build against. Mapping model transparency, explainability, and red-teaming results onto EU AI Act requirements.

How do you translate safety research into buildable frameworks?
1-2 year horizon · Cluster B · AI Safety & Governance
D.03 · GOVERNANCE

Decentralised Governance

How communities govern AI systems that affect them. This thread runs on the protocols it studies. Contributors earn a vote on cluster direction.

How can research communities maintain aligned incentives as they scale?
Ongoing · Cluster D · Open Model Development

ALL.CLUSTERS

Four research clusters — sixteen active threads

cluster_a.privacy[ long-term ]
// CLUSTER_A

Privacy-Preserving AI Systems

Privacy-preserving machine learning techniques that keep data and computation on-device. Covers federated learning with differential privacy, quantised transformer inference, verifiable computation via zero-knowledge proofs, and privacy-safe synthetic data generation. Wyvern serves as the reference on-device engine.
Key techniques
Federated learning · Quantisation · zk-SNARKs · Differential privacy · Synthetic data
Related product   Wyvern, on-device engine
A.01 Federated Learning at Scale

Training ML models across distributed nodes without centralising raw data. Open protocols for cross-organisational federated training with differential privacy.

A.02 On-Device Transformer Inference

Quantisation, pruning, and distillation research aimed at transformer-scale models that run privately on consumer hardware. Wyvern Engine is the reference implementation.

A.03 Cryptographic Trust Chains

Formal verification of model provenance and inference integrity. Generalising the Wyvern trust chain into a framework any ondevice model can adopt.

A.04 Synthetic Data Generation

Privacy-safe training datasets using generative models. Realistic data without exposing real users.

cluster_b.safety[ policy-critical ]
// CLUSTER_B

AI Safety & Governance

Developing interpretability and safety evaluation methods that translate into actionable engineering practice. Work spans explainability frameworks (LIME, SHAP), fairness auditing across languages and cultures, compliance mapping for regulations like the EU AI Act, and open red-teaming protocols for pre-deployment testing.
Key techniques
LIME · SHAP · Fairness metrics · Compliance mapping · Adversarial testing
Related product   Guardrail, moderation API
B.01 Model Transparency & Explainability

Making neural networks interpretable to the humans affected by their decisions. Layered explanation frameworks that work for non-experts, not just ML researchers.

B.02 Bias Detection & Fairness Metrics

Systematic measurement of discriminatory outcomes in model outputs. Current fairness benchmarks largely cover English-language and Western contexts; this thread develops evaluations for underrepresented languages and cultural settings.

B.03 EU AI Act Alignment

Translating technical safety properties into compliance-ready frameworks engineers can actually build against. Mapping transparency, explainability, and red-teaming onto EU AI Act requirements.

B.04 Red-Teaming Methodology

Open adversarial-testing protocols any lab can run against their own systems before deployment. Published methodologies for red-teaming that cover common attack surfaces and failure modes.

cluster_c.access[ high-impact ]
// CLUSTER_C

Accessible AI Integration

Making AI systems usable where compute, connectivity, and linguistic coverage are constrained. Work includes lightweight deployment pipelines for healthcare, education, agriculture, and legal aid in low-income contexts; low-resource NLP extending transformers to underrepresented languages; and infrastructure tooling that reduces the compute budget required for competitive model development.
Key techniques
Low-resource NLP · Transfer learning · Workflow analysis · Open infrastructure
C.01 AI for Under-Resourced Industries

Lightweight deployment pipelines for healthcare, education, agriculture and legal aid in low-income contexts.

C.02 Multilingual & Low-Resource NLP

Extending transformers to languages underrepresented in training data. English-only excludes people. Native-language annotators and new methodology required.

C.03 Human-AI Workflow Integration

How AI integrates into daily work across real task environments. Studies what breaks, what helps, and what creates dependency in production deployments rather than controlled settings.

C.04 Open Training Infrastructure

Tooling and frameworks letting a two-person lab compete on model quality, not compute budget. Innovation should not require a datacenter.

cluster_d.open[ community-led ]
// CLUSTER_D

Open Model Development

Community-directed research where contributors propose, second, and scope research threads through a transparent governance process. The lab holds one vote, equal to any other participant. Current threads cover open-weight reproducibility, ethical fine-tuning with auditable preference data, decentralised governance protocol design, and community-maintained evaluation benchmarks.
Key techniques
RLHF · DPO · Constitutional AI · Community governance · Open benchmarks

Cluster D is governed by contributors. Threads are proposed, seconded, and scoped by the community. The foundation has one vote, same as anyone else. We're building the governance model now, in public.

D.01 Open Weights & Reproducibility

Publishing weights, training configs, and evaluation results openly. Reproducible science is a standard. Community-maintained checklists and CI tooling for every release.

D.02 Ethical Fine-Tuning

RLHF, DPO and Constitutional AI with open datasets and preference models the community can audit and extend. Looking for contributors with multilingual annotation experience.

D.03 Decentralised Governance

Communities govern AI systems that affect them, not corporations. This thread itself runs on the protocols it studies. Sustained contributors earn a vote on cluster direction.

D.04 Benchmark Integrity

Next-generation evaluation frameworks that measure what actually matters. Not what's easiest to game. Community-proposed benchmarks with full methodology published.

GET.INVOLVED

Contribute

Research threads are open for contribution — propose a new thread, contribute to an existing one, or share datasets and methods.

// Coming soon

OPEN.PROBLEMS

Here are some problems
we are trying to solve.

Each is an open research question with no settled answer. If you have a lead, a dataset, a method, or just an opinion, these are the problems where your input matters most.

A.03 · CRYPTOGRAPHY
Can inference integrity be verified without access to weights?
ZK-based provenance works for training. We haven't cracked deployment-time verification at scale.
B.02 · FAIRNESS
How do you measure harm in languages with no established benchmark?
Western harm taxonomies don't translate. We need native-language annotators and new methods.
C.04 · INFRASTRUCTURE
What is the minimum viable compute for competitive fine-tuning?
We're tracking the floor as hardware and methods improve. Community data welcome.
D.01 · GOVERNANCE
What are viable quorum thresholds for early-stage research communities?
Too low and votes lack legitimacy. Too high and nothing passes. We need data on real participation patterns.
B.04 · FAIRNESS
How do you align harm taxonomies across languages and cultures?
A directly harmful statement in one language may be ambiguous in another. Static category systems aren't enough.
A.05 · CRYPTOGRAPHY
Can federated training be verifiable without a central coordinator?
Existing protocols assume a trusted aggregator. We want protocols where no single party holds the keys.

PUBLICATIONS.FEED

preprints · datasets · workshop papers · software releases, openly licensed

View all papers
publications.feed[ 1 paper ]
2025-07-01 · B.01 · CONTENT MODERATION
Efficient Content Moderation via Contextual Curriculum Learning
Fine-tuned transformer models consistently outperform a prompt-engineered Mistral-7B (not fine-tuned) on content moderation, using up to 244x less energy while processing content 167x faster. Curriculum learning improves contextual pattern recognition by 15 percentage points on average.
Content Moderation Curriculum Learning NLP
Read paper →