Research
Research is the argument. Infrastructure is the proof.
Regnant is a Sovereign Artificial Intelligence research and infrastructure company. The research programme exists to make sovereignty technically true rather than rhetorically true. Every position below either ships in a system or gets retired.
No fabricated benchmarks. No paper mill. Three programmes and four theses: the positions we actually hold.
Language is infrastructure: set one sort at a time.
R.01
LUGHA
Sovereign language models
Models built for the languages institutions actually operate in. A decision-maker briefed by a system that cannot reliably work in their working language is constrained by the tool, so linguistic fluency is treated as a first-order engineering requirement rather than a translation layer added afterwards. The programme is early: what we have published so far is a small adapted model, not a foundation model.
Research Preview
KW5-Lite →- A 74 MB LoRA adapter over Qwen2.5-1.5B-Instruct, not a model trained from scratch
- Holds a short exchange in standard-register Kiswahili on modest hardware
- Degrades after 7 to 10 turns, and is not for consequential work
- Open weights, Apache 2.0, so the claims can be checked
R.02
UAMUZI
Decision simulation
Deterministic systems that let institutions rehearse consequential decisions and delegate execution safely. Infrastructure, procurement and public-health decisions unfold over decades, so the systems that guide them have to produce the same outputs from the same inputs: auditable, reproducible, explainable. Rehearsal without determinism is theatre. Delegation without constraints is reckless.
Decision Intelligence
Feeds Wallgarden
- Same inputs produce the same outputs: determinism as a product principle
- Decision simulation: what happens if, answered before money moves
- Organizational memory: outcomes compound instead of being relearned
R.03
MAANA
Semantic grounding
Meaning treated as infrastructure. One machine-verifiable, governed definition of every concept, so systems integrate cleanly and AI stays grounded in fact instead of hallucinating across mismatched definitions. Every consequential output must be explainable back to source facts and rules, with defined behavior under disagreement.
Semantic Governance
Feeds Matta · Cordon
- OWL 2 DL and SHACL with strict, non-overlapping mandates
- A deterministic truth hierarchy and a temporal graph
- AI components consume only authoritative facts; probabilistic ones are labeled
Four working theses: the backbone
T.01
Sovereignty before convenience
When ownership is harder to deliver than access, we choose ownership.
When ownership is harder to deliver than access, we choose ownership. Most of the field optimizes the other way; that asymmetry is the research opportunity.
T.02
Language is infrastructure, not a feature
Linguistic fluency is a first-order requirement, engineered from the outset.
Fluency is not a feature you add after pre-training. It is a constraint you accept before it.
T.03
Systems that govern must explain themselves
Every consequential decision must be reconstructable, auditable, defensible.
Reconstructable, auditable, defensible: or not deployed. Explanation is an engineering deliverable, not documentation.
T.04
We ship only what we can account for
A smaller portfolio we own completely, over a broad one dependent on infrastructure we don't control.
A smaller portfolio we own completely, over a broad one that quietly depends on infrastructure we don't control. The portfolio is the proof of the thesis.
The register: live
What's on the bench.
Maintained by the team, published when real: this section reads from the working register, not from marketing copy.
ACTIVE
Sovereign language models
Models built for the languages institutions actually operate in precision and accountability, not approximation. Current artifact: KW5-Lite, a Kiswahili-first model in research preview with open weights under Apache 2.0.
KW5-Lite · Kiswahili · open weights
ACTIVE
UAMUZI
Deterministic systems that let institutions rehearse consequential decisions before committing. Same inputs, same outputs, auditable, reproducible, explainable. Feeds Wallgarden's decision simulation.
determinism · Wallgarden
ACTIVE
Semantic grounding — MAANA
Meaning treated as infrastructure: one machine-verifiable, governed definition per concept, built on OWL 2 DL and SHACL with strict, non-overlapping mandates. Feeds Matta, and grounds Cordon's inference in verified facts.
OWL 2 DL · SHACL · Matta
Published notes
Nothing published yet: and nothing will be listed here until it is real. No paper mill.
The empirical constraints
Positions are cheap; standards are not. Everything above is built against these.
- PDPA 2022Data Protection, Tanzania
- e-GAe-Government Interoperability
- BoT Cyber RiskFinancial Data Isolation
- Ed25519 · MerkleTamper-Evident Logging
- OWL 2 DL · SHACLSemantic Governance
- Apache 2.0Open Weights: KW5-Lite
What we release openly
Apache 2.0, open weights
KW5-Lite →
Our first step in Kiswahili: a small conversational fine-tune, published with its limits written down so the next one has something to beat.
MIT: open source
Milkshake →
The education ecosystem built on that thesis. Kairos for the teacher, Weave for the student. Free for schools, NGOs, and government education projects.
Interested in the programmes themselves? We are hiring slowly.