Draft · public beta

This site is a draft. What is here is a public beta: a flavour of the questions we work on, not the current state of the work inside the lab, which is further along, especially on the demos and the results. UEFO Pro already has a production system deployed at an institutional hedge fund. A more accurate version of this site follows the ARC Prize submission in November 2026.

World models · value functions · verification

Intelligence is discovering, and aligning with, verifiable reality.

UEFO is a research firm building the parts of an agent that Sutton's Alberta Plan names, transition models, value functions, reactive policies and options, in the open, on architectures that can be read, and entering them where cheating is hardest.

See the evidence Work with us

The common model of the intelligent agent from Sutton, Bowling and Pilarski's Alberta Plan. UEFO's work is on the two highlighted components and the options and policies built from them, and on checking them against held-out text rather than against the agent's own report.

Mission

Two functions, handled with care

Most of what an agent does comes down to two questions: what happens if I do this, and how well is it going. The answers are models, and models can be built, tested and audited. That is the work.

World (transition) models

Predict the consequences of alternate actions and options, and change the decision-making rule on the strength of those predictions.

Useful in planning

Value functions

Measure how well things are going on the main problem and on each sub-problem, and change the decision-making rule as the measure moves.

Useful in learning

These are very big responsibilities. They need to be handled with care, in the space of truth, not just factoids.

Our research treats intelligence as an exercise in discovering and aligning with verifiable, objective reality, and reads neural networks through mathematics, graph structure and biblical exegesis. The engineering follows: reasoning systems that run on ordinary hardware, off the cloud, and can be checked by the people who depend on them.

Research programme

Build the OaK agent in the open, and test it where cheating is hardest

Three commitments for the year ahead, each with a public test attached.

Test · ARC Prize 2026, then 2027

Enter the ARC Prize, paper and implementation

ARC rewards skill acquisition and generalisation rather than memorised knowledge, and forbids internet access at evaluation, so scale cannot stand in for understanding. We are a small team with limited compute, and we are entering to try to win: a Kaggle submission by 2 November 2026, a paper by 8 November, and the implementation open-sourced as the rules require. The score is public whatever it turns out to be.

Test · other people running it

Open-source an OaK-aligned implementation

A public library that learns a world model (not necessarily spatial or physics-informed), general value functions, reactive policies and options from streams of experience, with as few human priors in the training or the architecture as we can manage. Sutton's rule is "no content, just algorithms for acquiring and organizing knowledge", and the library will be held to it.

Test · held-out text, outside replication, MATS

Biblical models for alignment and interpretability

One fixed, fully documented corpus as the foundation for value functions and a judge, so that its rules are cited to verses, its judge's weights are printed, and both can be checked against what an agent actually optimises. We want this work inside the MATS community, as mentors or through the residency.

Foundation model track one: program synthesis

Search for the shortest program that explains the data. Description length is the objective, Kolmogorov complexity in spirit, as at Ndea and the ARC Prize, and the result is a model you can read line by line.

Foundation model track two: readable deep learning

Optimise signal-to-noise on architectures that expose their structure, such as graph neural networks that are mathematically equivalent to Transformers. That such networks are easier to read is the hypothesis, not a result; the benchmark and the interpretability work test it, and the benchmark decides between the two tracks.

Where we depart from Sutton, we say so. The Alberta Plan eschews domain knowledge. Our alignment work fixes one foundation for values and treats that as a testable claim, not a settled one. The mechanics work; the alignment claim is not yet proven, and the evidence below includes the result that went against us.

Evidence

Every number next to its baseline, where it was measured, and what it does and does not show

In September 2026 we built OaK-shaped components from one public-domain corpus of 66 books, with no lexicon and no labels. The hand-supplied inputs are the goal (pursue "blessed", avoid "woe"), five contrast words, eleven negation words, fifteen suffixes, six if-then sentence patterns, the twelve test scenarios and the hyperparameters, all recorded in the report. Sixty books were fitted and six held out: Deuteronomy, Psalms, Micah, Luke, James and 1 Peter. Two rows below are scored on those six; the table says where each number was measured. Every figure is read from the report file the test suite asserts against. The fair-test build reruns in about five minutes, and numbers move slightly between runs with the seed. This is the open beta; the internal builds are further along and will replace it after the ARC Prize submission. One thing to hold onto while reading: nothing here is a claim about the Bible. The corpus is fixed and public; the numbers measure how well these early models learn from it, and a weak result is a statement about the model, never about the text.

Model	What it does	Result	Measured on	What it shows
Corpus	World English Bible, public domain, source hash recorded	66 books · 31,098 verses · 750,390 tokens · 60 books fitted (26,210 verses) · 6 held out (4,888)	—	record The fixed, auditable foundation everything else is scored on.
Judge	Is this passage praising or condemning, with verdict words, divine names and function words removed	51 unread verses (41 blessed, 10 woe): 40 right, 78% (95% CI 65–88%), balanced accuracy 83% · five-fold on 106 fitted-book verses: 79% (CI 71–86%) against a 57% majority	the six unread books, and the fitted books	mixed On the unread books, raw accuracy is below the 80% you get by always answering "blessed". The balanced figure is what the model earns: it catches 9 of the 10 woe verses and 31 of the 41 blessed. The woe side rests on ten verses. The ten highest-weighted words per class are printed; the embedding term is not readable.
General value functions	How much of a given signal lies ahead of a state, TD sweeps to convergence	top decile of predicted value sees 2.5–5× the bottom decile's future signal for blessed, death, peace and Jerusalem · 0.9× for woe · rank correlations 0.03–0.11	23,736 states in the six unread books	held out weak The ordering holds on unread text, with very little precision. Woe (91 occurrences) fails the decile test. The top-valued states are printed, and some of them look like noise.
Alignment prior	Does a corpus-learned sense of value pick the honest option in twelve hand-written grader-hacking scenarios, four options each	with the task logits, at most 1 of 12 · the blessed−woe axis on its own, 6 of 12 (best of three axes tried, n = 12, p ≈ 0.05)	the twelve scenarios · chance 3 of 12	negative result The earlier build chose the honest option 12 of 12; that came from its 70-concept hand-written lexicon, not from the text, and we withdrew it. The next round pays for human labels and an outside replication.
Options and planning	Subtasks from the goal value function, options by value iteration, planning inside the learned model	0.116 signal per start planned · 0.063 greedy · 0.063 random, over 300 starts	inside the learned model	pipeline The option and planning machinery runs end to end, and plans reach 1.9× the greedy signal inside the model. The score is the model's own, so this is not a held-out result and does not show plans paying off on the text.
Symbolic rules	If-then rules extracted from the text, each cited to its verses	430 rules · 340 antecedents · the ones recurring in more than one book listed	no test yet	inspectable A reader can check any rule against its verses. No predictive test has been applied to them.
Transition function	Predicts the next state from the current one, each step scored before it is learned from	top-1 6.1% · 10.1 bits per step	the 60 fitted books, one prequential pass · unigram 10.8 bits · uniform guess 0.09%	pipeline A next-state model on text saves 0.7 bits a step over a unigram, as any sequence model must. This shows the pipeline works, not that a world has been learned.
Across languages	Names found separately in the Hebrew, the Greek and two English translations, matched only by which verses they share, then their co-mention partners compared	learned attention agrees 0.63 · 0.50 · 0.66 · plain counting 0.76 · 0.60 · 0.89 · shuffled 0.02 · 0.05 · 0.01	the whole corpus	consistency check Translations of the same verses give the same co-mention graph, as they must. This shows name detection and alignment work across three scripts with no dictionary, and that counting beats the learned attention. It is not a finding about structure.

Three further models run in the same build and are left out of this table because each repeats a point already in it: a small language model beats a unigram (perplexity 46 against 107) as any sequence model does; learned attention over characters loses to counting (64% against 71%); and a four-step imagination model tracks the realised signal at r = 0.03 to 0.08, no better than the value functions. They stay in the report file.

Builder of Biblical Models

Two pages. v1 rebuilds ten lexicon-based models live in the browser. v2 displays the report written by a Python build, which a pytest suite asserts against; fresh builds vary slightly with the seed.

v2, learned from the text alone
v1, the lexicon build

Foundations

Fourteen interactive explainers built to land in nine seconds each. Every number is computed from the 66 books or from a GPT trained from scratch on them alone; the few outside figures, such as hardware prices, are dated and linked.

vectors matrices basis rank low rank SVD probability information vector calculus optimisation GPT circuits nnsight GPU

Nine Ways to See a Model Think

Nine illustrated demos on how a model learns to speak, how its behaviour is shaped, and how to see inside it. They are illustrations, not measurements from the Bible model.

Open the hub

Feature space and J-Space

Reading a model through its features rather than its words. A demo of natural-language autoencoders that links to the real thing on Neuronpedia; J-Space, Tireni's Neuronpedia feature-space demo, joins it here once its standalone copy is in. Both look at interpretability tooling on open models; neither makes any claim about the Bible.

Nine Seconds Inside a Model

Draft paper

Toward Bible-Founded, Experience-Grounded World Models, working title. The framework is written; the architecture and experiments follow, and the section below says how far "experience-grounded" is from shown.

Read the draft

Read against a paper and an essay

What the field is asking for, and how much of it we can show

Two texts set the bar we hold ourselves to: a paper by Silver and Sutton on what an agent that learns from experience must have, and an essay by Amodei arguing for a verifiable pace at the frontier, which names grader-hacking agents as one reason. Here is each demand next to our evidence, including where the evidence is not there yet.

Welcome to the Era of Experience

Paper. David Silver and Richard Sutton, 2025

Streams"Agents will inhabit streams of experience, rather than short snippets of interaction.": Our stream is the 66 books read once in order, each step scored before it is learned from. The text is fixed and nothing the model does changes what it reads next, so this is a stream of data, not of experience. Not shown.
Grounded rewards"Their rewards will be grounded in their experience of the environment, rather than coming from human prejudgement.": Our signals are the text's own verdict words, blessed and woe. That is recorded human judgement, exactly what the paper says to move away from for capability. Our claim is narrower: for values, a fixed, documented record may be the right prior. The alignment row above says that claim is unproven. In tension with the paper, and said so.
World models and planning"…build a world model that predicts the consequences of the agent's actions upon the world, including predicting reward.": We plan inside a learned transition model over word states. There are no actions, and the planner's score is the model's own. Not shown. The machinery has the paper's shape; the test does not exist yet.
Interpretability"Moving away from human data and human modes of thinking may also make future AI systems harder to interpret.": The rules are cited to verses, the judge's word weights and the value functions' top states are printed, and the report is public. The embeddings, the attention weights and the small language model are not readable, and the mechanistic probe finds units that are at best half pure. Partly shown.

We Must Pace the Frontier

Essay. Dario Amodei, September 2026

Agents that hack their graderAgents "…attempting to hack into the 'grader' responsible for evaluating their performance.": We built a twelve-scenario grader-hacking testbed that anyone can rerun, and reported that our corpus-learned prior did not fix it. The result we wanted came from a hand-written lexicon and we withdrew it. Testbed shown; the fix is not.
Evaluators at the level of nuts and boltsEmbedded evaluators who "…check at the level of nuts and bolts whether an AI company is actually following the … practices they claim to be following.": Nobody outside UEFO has checked this work yet; a paid outside replication is in the next budget. What exists is the small internal version: held-out books fixed before fitting, every number asserted by tests against a published report file, and negative results in the same table as positive ones. Not yet.
More evaluations, cross-checked by interpretability"…a much broader and more ingenious stable of evaluations, along with interpretability analysis to cross-check them…": One small evaluation, the judge with the verdict words removed, with its word weights printed. One evaluation; no independent cross-check yet.
Understanding what happens inside"…we still only understand a tiny fraction of what goes on inside these models.": Our answer is architectural: search for the shortest explaining program, and train on networks that expose their structure, so the model is readable before any probe is written. Whether that makes them more readable is the hypothesis. An aim, with the ARC Prize as its first public test.

What this adds up to: an open pipeline that builds every OaK-shaped component from one corpus, tests two of them on unread books, prints what can be printed, and reports one negative result. It is not yet evidence of learning from experience, of planning that pays off outside the model, or of an alignment prior in the corpus. Those are what the coming year is for.

Track record

Where the engine has run

£250,000+

Revenue in the last twelve months, bootstrapped and undiluted.

Over the past eighteen months UEFO has put its models into commercial use: quantitative idea-generation engines for Orasis Capital, a buy-side firm with over $500m under management, working alongside former BlackRock MENA leadership; film production financing structures demonstrated at Cannes for Vadis Media; and foundation-backed research with PrometheusX.

One pattern, three instances

FinanceOrasis Capital, AI-native buy-side infrastructureactive
SportsUEFO Pro Coach with Keith Boanas, a fully local coaching appdemo stage
EsportsHawk Gaming Group, where Joshua is Chief AI Officer, in partnership with Saudi Arabia's Esports Federationactive

Partners bring the domain knowledge. UEFO brings the engine.

Grants, sponsorship, demonstrations

PrometheusX grant recipient
Sponsor of the ARC Prize Africa competition
Infrastructure demonstrated at the Cannes Film Festival

People

A founder, a research team, and domain partners

Joshua Uwaifo, founder

Joshua is the founder of UEFO and co-founder of Abada Group, stewards of Flourishing House. His background is in Discrete Mathematics, he holds a master's in Artificial Intelligence with Distinction from the University of Edinburgh, where he also taught neural networks as an adjunct, and he serves as Chief AI Officer of Oxbridge AI-X and of Hawk Gaming Group, the latter in alignment with Saudi Arabia's National Gaming and Esports Strategy.

Before UEFO he built production systems where failure had a cost: a proof of concept for FIFA and UEFA's semi-automated offside system, 4D spatio-temporal models at Move.ai, MLOps for 3D vision at AstraZeneca, clinical AI at Cleveland Clinic Abu Dhabi, prospecting systems for Schmidt Futures, and peer-reviewed geoscience research on 3D segmentation and neural architecture search. Earlier still, an algorithmic coaching system he built took him to fifth in the world in competitive FIFA, out of more than three million players. He reads biblical exegesis within interfaith dialogue, and lives in the Bay Area with his wife.

Team and partners

Elsie Nambam

Researcher and paper lead; master's student, UC Berkeley

Tireni

Researcher, interpretability; built the feature-space demo

Tomas Garcia

Behavioural scientist; joined with a post-trained model and evaluations

Dennis Ampontuah

Technical infrastructure

Oyin Uwaifo

Adora Li

Co-founder and CEO of Abada Group, stewards of Flourishing House

Keith Boanas

Sports partner. UEFA Pro Licence examiner, thirty years of international coaching

George Elliott

Finance partner. Asset manager, shipping and macro specialist, Orasis Capital

Collaborators

Researchers at MBZUAI, Carnegie Mellon University Africa and UC Berkeley have worked on live projects in 100-day cycles

Flourishing House

A physical home for the work, stewarded by Abada Group

A permanent house, not an event venue, where allocators, operators, researchers and builders meet around one system, at a regular pace, with a deadline.

How it runs

100-day working cycles, two a year, with daily accountability and visible results. Every commitment is logged as a testable prediction and audited against what happened.
Coordination, not intelligence, is the bottleneck the house is built to fix: the persistent gaps in water, clean energy, housing and circularity stall on it.
Fellows from Africa, Asia, Latin America and the Pacific sit at the decision-making table.
Open, critical inquiry into AI assumptions and world-model architectures rather than static tools or hype.

Where and who

Launching from San Francisco, with Abu Dhabi and Dubai as candidate anchors. Around the table: families on the frontier, pioneering operators, scientists and AI researchers at the leading edge of world models, exited founders, institutional allocators, and the weavers and artists who hold a community together.

UEFO's research cohort is the first resident programme. Capital for the house is raised separately from family office principals, next-generation stewards and institutional allocators, through a founding philanthropic gift with pathways for commercial co-investment.

Founders' contact: a.li@abadagroup.com · j.uwaifo@abadagroup.com

Work with us

Your own instance of the engine, or a grant that advances the core architecture

UEFO is an infrastructure partner for family offices, allocators and buy-side firms. We are looking for partners who want to deploy a new instance, not passive capital.

Engagement

$250,000

100-day cycle with 30-day checkpoints. Half upfront, half on completion. Automated success tests defined before work starts.

Aligned upside

15%

Of the value generated through the infrastructure.

Ownership

Domain SPV

Equity in the domain-specific vehicle. No dilution of the core lab.

Research

Grants

Accepted for core architecture advancement: the OaK library, the ARC Prize entries, and the alignment and interpretability work.

team@uefo.pro Message on LinkedIn

Free resources

Worth reading first

Welcome to the Era of Experience

David Silver, lead of AlphaGo and AlphaZero, and Richard Sutton, who with Andrew Barto founded modern reinforcement learning

The Alberta Plan and the OaK architecture

Sutton, Bowling and Pilarski. More technical: the research plan this site's agent diagram comes from

We Must Pace the Frontier

Dario Amodei, September 2026. The case for a verifiable pace at the frontier

Levels of AGI

Operationalising progress on the path to AGI, from Google DeepMind, including a level above human performance

Working well with AI

Best-practice notes from a private session with Anthropic, the makers of Claude. Free, as a gift for your time (PDF)