Fortunes go to those who price risk before the world does.Fortunes go to those who price risk before the world does.

What verifiable rewards solved

Frontier LMs scaled general reasoning through RL with verifiable rewards - it works elegantly when the answer is deterministic and checkable e.g. math, coding, physics.

The real world withholds information

Real-world decisions add another layer of complexity: no single actor ever has the complete information about their world. e.g. in a negotiation, a sale, or a market, each participant sees only part of the environment and knows things the others don't.

Therefore, a model now needs to estimate two things - what it can't see about the world, and what the other participants already know (poker is the simplest version of this problem).

The ceiling of human-derived reward

But there's no verifiable answer to these estimates - a decision is only proven right or wrong once it plays out. So, today's best approaches can only reward it with a human-derived signal, and intelligence stays bound by human understanding.

Reward from real-world outcomes

We propose rewarding models directly from real-world outcomes instead, letting them explore causal relationships invisible to human reasoning.

Continual learning labs (e.g. Applied Compute) have adopted a similar approach, but they determine the reward with an LM judge. Even the best LM judge is a derivative of human reasoning - it can make models cheaper and more specific to each workflow, but it can't push them past human reasoning.

Forecasting first

We start with the cleanest case, forecasting: the model is rewarded by how its forecasts actually turned out, so it learns from the specific parts of the reasoning trace that helped or hurt the forecast. As they scale, models learn decision procedures (which evidence matters and how to weigh them against each other) instead of imitating human reasoning.

On ForecastBench, our system ranks second on market questions - the harder half of the benchmark - beating several Google DeepMind and xAI submissions, and is one of the first two teams to reach superforecaster parity there: a Brier index (higher is better) of 77.0 vs the superforecaster median's 75.6 on 57 resolved market questions, and 64.1 vs 63.6 on the dataset half.

The simulator trap

AI-for-science labs (e.g. Periodic Labs) break past the human ceiling by pulling a real, deterministic reward straight from experiments. But even with the reward solved, it doesn't scale: you can't run millions of real experiments, so you need a simulator.

Simulators can only approximate a physical experiment. RL exploits the sim2real gap: its best solutions ace the simulation and fail in the real world.

Markets as the honest simulator

To solve this, we pick financial markets - the simulator is a clean backtest that deterministically replays past data. Any reward-hacking behaviour is eliminated by the market's adversaries, which we model rather than replay - so the model faces opponents that react to it, and can't overfit to one recorded history.

This is already running live: in its first 48 days trading Polymarket, the system made $18K in profits after fees ($20K before) - positive on 45 of 48 days, with a 33-day winning streak, an annualized Sharpe of 38, and a max drawdown of ~$350. A live market is a more rigorous eval than any benchmark - every mistake is charged to the account.

The road ahead

By Mar 2027, we will have a 300B model that consistently beats superforecasters on Metaculus and ForecastBench - this intelligence will be used to trade financial & prediction markets, with RL optimizing the decision layer.

Our goal is to have this system trade without any human intervention and get through 12 months without a losing week. This is the most rigorous benchmark for a model optimised on decision making.

In ten years, our models will run real companies to top-decile growth and profit, outperforming most human operators in the hardest environment for decision-making.