We are building an engine that predicts. Sport is where we prove it.
ClutchTip is an AI research and development lab, not a tipping service. Our engine learns patterns of behaviour and calls what happens next, before it happens, from public information alone. Sport is the proving ground because every call can be lodged in advance and graded in public. On that record: the most accurate published NRL predictor of 2026, 73.6% across every NBA game of 2025–26, and a closing-line profit in two football codes, without the engine ever seeing a price.
Built by Jack, for a bigger problem than sport
ClutchTip was founded, and its engine built, by Jack: an Australian solicitor admitted to the Supreme Court of Queensland and the High Court of Australia, with a Master of Laws, a Bachelor of Laws and a Bachelor of Commerce in accounting. Before law he worked in insolvency accounting, a career spent reading the numbers for the moment a pattern breaks.
The engine began somewhere personal. Jack lives with a brain tumour, and living with it raised a question no tipping service asks: could a machine learn one person's patterns closely enough to notice change early, and help them think, decide and remember? That is the long goal. An assistant for the brain.
The first layer of any such assistant has to be a prediction engine that is accurate, honest about its uncertainty, and versatile enough to work on problems it was never shaped for. Sport is the hardest public test we could find for that layer: thousands of outcomes, markets of millions setting the price, and a result every night that cannot be argued with. So every call is lodged before the game and graded in public, wins and losses alike, and the same engine is already being pointed beyond sport.
ClutchTip is that first layer, running in the open. The sports record is the evidence. The engine is the work.
How a call is made
Every prediction comes out of the same proprietary computation procedure. What follows is the shape of it, enough to show the rigour, not the recipe.
Lineups, availability, form windows, schedule load, travel and venue, built only from information that exists before kick-off. Nothing in-game, ever.
Machine-learning models trained on years of game and player history produce independent probability estimates. In the NBA, each game is simulated 20,000 times.
The estimates pass through ClutchTip's own computation procedure, which weighs what the numbers can and can't see. This is the part we keep to ourselves.
Each call gets a confidence tier and a margin, is lodged before the game, and is graded in public on the final score. No edits after lodgement.
The engine never sees betting odds. Every pick is formed without reference to the market, which is why it is free to disagree with it.
The four tiers
Every call carries a tier, and every tier counts in the record. Across sports the tiers land in order: at the 2026 World Cup, PREMIUM went 87.5%, STRONG 75.0% and EDGE 66.7%.
Highest conviction. 81.7% across the 2025–26 NBA season.
Clear edge, good agreement, acceptable variance.
A directional lean with limited confidence.
The engine is telling you it could go either way. Still lodged, still counted.
Every sport, every call counted
| Competition | Record | What it means |
|---|---|---|
| NRL 2026 | 141 / 204 · 69.1% | The most accurate published NRL predictor of 2026: ahead of every verifiable public record we found, the Channel 9 panel's best (139) included. Grand Final called: Roosters, won 19–18. |
| AFL 2026 | 143.5 / 199 · 72.1% | Every game, misses included. 77.8% from Round 18. Grand Final called: Brisbane by 1–12, won by 7. |
| NBA 2025–26 | 73.6% (425–152) · PREMIUM (now referred to as ELITE) 81.7% | Every game played. Season three tips off 20 October. |
| NFL 2026 | first season | A brand-new code. The fitted model joined from Week 3. |
| World Cup 2026 | 72% | A cold start. Called Spain in the final while the Opta supercomputer had moved to France. |
| Super Rugby 2026 | 45 / 57 · 79% | Rugby union, full slate. |
| Furlong · racing | 25.6% top pick | 440 of 1,728 graded races won by the top pick, from form alone, without a price. |
It never sees the odds. It beat them anyway.
Because the engine forms every call without the market, its record can be measured against the market honestly. We joined every 2026 pick to the closing odds:
| Engine | Backing every favourite | |
|---|---|---|
| NRL · flat stake at the close | +9.0% | −10.3% |
| NRL · accuracy | 69.0% | 62.6% |
| AFL · flat stake at the close | +5.7% | +1.9% |
An accuracy audit for public information, not betting advice.
What started as a question is becoming a full analysis
The first goal was simple: show that outcomes could be called. That is now settled on the public record. The engine is branching into complete pre-game analysis: margins, player projections, the reasons a call could fail, and how each game sits against the market. One engine, every sport, and the record keeps growing.
- March 2024Day zero
One developer, one workstation, manual NBA tracking. Every outcome logged.
- October 2024First model live
NBA 2024–25: 478 tracked predictions at 61.5%, with no confidence filtering.
- January – March 2025100 straight
100 consecutive correct NBA game winners, and a 3,700:1 multi on 11 March, every leg from the model.
- December 2025The rebuild
New architecture, the confidence tiers, live injury filtering, 4,497 games of multi-season validation.
- 2025–26NBA season two
73.6% across every game played (425–152), 81.7% on PREMIUM (now referred to as ELITE).
- March – September 2026Into the football codes
NRL and AFL called every round, lodged before the first kick-off. NRL finishes as the most accurate published predictor of the year.
- June – July 2026World Cup, cold
Pointed at a tournament with no preparation: 72%, tiers in order, and the final called.
- July – August 2026A new path
What the football codes and the World Cup taught us about how outcomes are reasoned led us to test a new path. It is now the engine.
- September 2026The market audits
Every NRL and AFL pick joined to closing odds and published. NFL season one begins.
- October 2026Season three
Furlong rates every Australian race. The NBA, our flagship, is back.
- September 2026A wider question
The programme's purpose is set down: an assistant for the brain. Sport stays the proving ground, because it grades every call by morning.
- October 2026Racing at the top level
Furlong's top pick wins The Everest (17 October) and the Cox Plate (24 October). The engine never saw a price.
- October 2026Feeling as architecture
Research into value and affect as the building blocks of the assistant: like and dislike as primitives, everything else derived.
- 26 October 2026Furlong v1.1
A new racing model goes live after three weeks run side by side with the old one on every race.
- November 2026The spring four
The Big Dance and the Melbourne Cup complete four feature races, four top picks first past the post. The Rugby League World Cup pool stage closes 13 from 13 lodged games.
- NowThe lab
ClutchTip is an AI research and development lab. NBA, NFL, NHL, Premier League and racing in Australia, the UK and Ireland run every day; cricket is next. The working paper below sets out where it is going.
Watch the engine work.
Every call lodged before tip-off, every result graded in public.
See the NBA card The full recordClutchTip is a sports prediction and analytics service. It is not a gambling or wagering service, does not accept bets and receives no commission from sportsbooks or gambling operators. Predictions are pre-game only and do not account for changes after lodgement. Past accuracy is not a guarantee of future performance.
From a prediction engine to an assistant for the brain
Abstract
This paper sets out the research programme behind ClutchTip. Its long-term aim is an assistant for the human brain: a system that learns one person's patterns well enough to support how they think, decide and remember. Its first component, and the subject of the public record, is a general prediction engine. We argue that prediction is the right first layer, explain why sport was chosen as the test environment, report the engine's lodged and graded record to November 2026, and set out the philosophical and legal questions the later layers raise. Method detail is withheld. Outcomes are not.
1. The question
The programme began with a narrow question: can the behaviour of a complex system be called before it happens, from public information alone? Sport supplied a version of that question that could be answered honestly, because every game ends in a result nobody can argue with. The broader question sits behind it. If a machine can learn the patterns of a team, a horse or a market well enough to anticipate them, can it learn the patterns of one person well enough to help them?
2. Prediction as the shape of a mind
When we considered how the engine should develop, the first thing to settle was what a prediction engine is ultimately for. The idea that perception is a form of prediction is older than computing. Helmholtz described seeing as unconscious inference in the 1860s1. The modern predictive-processing programme treats the brain as a system that continually forecasts its own inputs and learns from the error2–5. On this view prediction is not one faculty among many; it is the common currency of perception, action and, on some accounts, conscious experience itself5.
We take a practical position from that literature rather than a metaphysical one. If prediction is close to the core of what a mind does, then an assistant for a mind needs a prediction engine at its base, and that engine should be general: able to work on problems it was never shaped for. Our ideas were generally reached from the problem side and found their precedents afterwards; the references are acknowledged as prior work, not as the source of the design.
3. Sport as the proving ground
Sport was chosen for three properties. It produces thousands of outcomes a year with unambiguous ground truth. It is priced continuously by betting markets that aggregate the judgement of very large numbers of people, which gives a demanding benchmark. And it is public, so every claim can be checked by anyone.
Three rules govern the record. Every call is lodged before the event. Every call is graded in public, wins and losses alike. And the engine never sees a price: market data is excluded from everything that answers who wins, and is used only afterwards, to audit the calls against the marketa. The last rule is a cost as well as a principle; a model that sees the price inherits much of its accuracy. It is kept because an engine that learns the sport, rather than the market, is the only kind that can transfer to problems without a market at all.
4. A new path
As the record grew, the engine needed further evolution. The first models were conventional, and running them across the football codes and the 2026 World Cup showed us where conventional forecasting stops: it describes the average case well and the particular case badly. So we considered a different approach, and that led us to test a new path. It is now the engine. Its design is proprietary and is not described here; its record is published in full.
5. Results to date
| Domain | Record | Note |
|---|---|---|
| NBA 2025–26 | 425–152 · 73.6% | Every game played; 81.7% on PREMIUM (now referred to as ELITE) |
| NRL 2026 | 141 / 204 · 69.1% | Most accurate published NRL predictor of the season |
| AFL 2026 | 143.5 / 199 · 72.1% | Profitable at the closing line without seeing it |
| Rugby League World Cup | 13 / 13 | Every lodged pool game |
| Racing (Furlong) | ≈ 1 in 4 | Top pick wins about one race in four across thousands; four feature races, four top picks first past the post |
Live seasons (NBA 2026–27, NFL, NHL, Premier League) are graded nightly on their own pages. The full record and citation page carry every figure with its denominator.
6. Generalisation
Once the new path held in one sport, the next question was whether it would travel. A forecaster tuned to one sport proves little about prediction in general. The same engine has now lodged calls in NRL, AFL, Super Rugby, rugby union Tests, the NBA, the NFL, the NHL, MLB, five European football leagues, the FIFA World Cup, the Rugby League World Cup, a professional boxing card and thoroughbred racing in three countries, and it has been applied to one domain outside sport that is not yet published. Several of those were cold starts: the engine was pointed at a competition with no preparation and its first calls were lodged and graded like any other. Generalisation across domains with different structures, scoring and noise is the claim this programme most needs to establish, and the one a forthcoming peer-reviewed paper will address.
7. Feeling, value and the individual
A general engine still only answers what will happen. To evolve toward an assistant it needed a sense of what matters, because an assistant that only predicts is a calculator with a good memory. So we considered value and feeling. Here the programme draws on a long line of work treating emotion as a form of judgement rather than its opposite. Spinoza derived the passions from a small set of primitives, defining hope and fear in terms of an uncertain future6; Hume placed the passions ahead of reason in motivation7; Bentham built an ethics on pleasure and pain8; James and, later, Damasio located feeling in the body's own state and argued that decision without it fails9,10; Minsky treated emotions as different ways of thinking11.
The working design takes like and dislike as the only primitives and derives the rest: fear as a predicted dislike, love as a present and enduring like, and so on outward to empathy and grief. A related claim, close to Simondon's account of individuation12, is that feeling is what makes a point of view someone's: two systems with the same knowledge are distinguished by what they care about. We set tests in advance for when such a layer is doing more than describing feeling, rather than claiming it does.
8. Other minds
Building toward feeling raises a question that had to be considered before going further: what, if anything, is such a system? No part of this programme claims machine consciousness. The problem of other minds is as old as Descartes13, and the modern debate (Turing's behavioural test14, Searle's Chinese Room15, Dennett's deflationary account16 and Chalmers' hard problem17) has not settled it. Two observations shape our approach. Most of the processing in a human brain is not available to awareness, so introspection is a weak instrument in either kind of system. And a question that cannot be settled by argument can still be approached by measurement. The programme therefore adopts an agnostic, behavioural standard: it names what a system would have to do, records whether it does it, and leaves the metaphysics open.
9. Law and the directing mind
A system that acts for a person needs someone to answer for it, so we next considered how the law already treats minds that are not human. Law has attributed a mind to a non-human entity for more than a century. A company acts and intends through those who are its "directing mind and will"18,19, and the courts have since framed this as a question of the rules of attribution appropriate to the purpose20. Hobbes had the artificial person in 165121; Kelsen treated legal personality as a construct of the legal order rather than a fact about the world22.
Current AI regulation is product-shaped: it governs systems as things that are made, sold and used. It does not yet ask whose mind is acting when an autonomous system decides. Corporations law already holds a worked answer to a closely analogous question, and applying that framework to AI governance is the subject of the founder's planned doctoral research.
10. Towards an assistant for the brain
Each of those steps points to the same destination. The familiar metaphor for personal computing is a bicycle for the mind: a tool that amplifies a capable user. Our aim is closer to a wheelchair for the brain: a support for a mind whose own capacities are limited, by illness, injury, age or load. The prediction engine is the first layer. Later layers add memory of a person's own history, a model of what they value, and the means to act for them under limits they set.
Four design commitments follow. The assistant is owned by its user and runs on their own hardware where it can. It does not depend on any single AI provider. Its values are taught as a code in plain language rather than locked into rules. And its security is defensive only: it detects, deflects and records, and never retaliates.
11. Limitations and open questions
Sport is a narrow world, and success there does not establish success in a person's life. Several samples are small, and some headline figures will move as seasons complete. The engine's method is withheld, which limits independent replication; the public, pre-lodged record is offered in its place, and a peer-reviewed paper is planned. In racing, a model that cannot see the price trails the market favourite on raw win strike, which is the cost of the market-blind rule stated plainly. It also opens the next question, now under forward test: whether the price-blind model's read, combined with market data in a separate layer downstream of the engine, can outperform both the model and the market alone. Early results put the combination level with the market; whether it moves ahead is not yet established. The later layers described in sections 7 and 10 are in development and are not claimed as results.
The timeline above records each step as it happened.
References
- Helmholtz, H. von (1867). Handbuch der physiologischen Optik. Leipzig: Voss.
- Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.
- Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204.
- Hohwy, J. (2013). The Predictive Mind. Oxford University Press.
- Seth, A. (2021). Being You: A New Science of Consciousness. London: Faber & Faber.
- Spinoza, B. (1677). Ethics, Part III: Of the Origin and Nature of the Affects.
- Hume, D. (1739–40). A Treatise of Human Nature, Book II.
- Bentham, J. (1789). An Introduction to the Principles of Morals and Legislation.
- James, W. (1884). What is an emotion? Mind, 9(34), 188–205.
- Damasio, A. (1994). Descartes' Error: Emotion, Reason, and the Human Brain. New York: Putnam.
- Minsky, M. (2006). The Emotion Machine. New York: Simon & Schuster.
- Simondon, G. (1958; publ. 1964, 2005). L'individuation à la lumière des notions de forme et d'information.
- Descartes, R. (1641). Meditations on First Philosophy.
- Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460.
- Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417–457.
- Dennett, D. C. (1991). Consciousness Explained. Boston: Little, Brown.
- Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200–219.
- Lennard's Carrying Co Ltd v Asiatic Petroleum Co Ltd [1915] AC 705.
- Tesco Supermarkets Ltd v Nattrass [1972] AC 153.
- Meridian Global Funds Management Asia Ltd v Securities Commission [1995] 2 AC 500.
- Hobbes, T. (1651). Leviathan, ch. 16: Of Persons, Authors, and Things Personated.
- Kelsen, H. (1934). Reine Rechtslehre (Pure Theory of Law). Vienna: Deuticke.