Time machine: leak-free RL environments, graded by what actually happened

Every AI lab now wants reinforcement-learning environments: places where a model practises a task, tries things, and is scored on how it did, thousands of times over. The hard part is never the practice. It is the score: somebody has to know the right answer for every attempt, without being fooled.

An environment needs three things: tasks, a way for the model to act, and a score it cannot cheat. Code is the favourite because tests pass or fail. Most real work has no tests, so environment builders pay experts to grade, or ask another model to, and both can be fooled.

A forecast is different. “Will this company's next annual report show X?” has an answer that arrives on its own, from the world, with a date on it. And forecasting is no toy: a team at Berkeley got a search-and-reason system to the level of a human forecasting crowd, and a team at Lightning Rod Labs trained a 14-billion-parameter model on nothing but how Polymarket questions turned out, and it matched far larger models (sources at the end).

The trouble is that a real forecast takes months to score. So everyone backtests: pretend it is the past, ask about what happened next, score at once. That is exactly what a time machine does, and it only works if the past stays the past.

Traders have a name for a backtest that can see the future: look-ahead bias, and it makes every strategy look brilliant. AI forecasting has the same problem in new forms. Researchers at ETH Zürich's SPY Lab showed that search engines' “before this date” filters let later pages through, and that models know things from after their stated training cutoff. A Northwestern study this August found the standard check for that kind of leak flagging four of five models on questions they could not have seen.

The careful answer so far has been to freeze a news archive and search only that. It works, but news is only part of what an agent reads. It also looks up filings, prices, trial records and drug approvals, and each one is a door the future can come back through.

hedwigai ask does all of its looking up through one shell: a locked-down command line where each data source is a program, such as biotech for SEC filings, FDA approvals and clinical trials. One door is the point. If every source comes through the shell, one clock on the shell covers every source.

The everyday shell knows only now. Its one clock function is time, and it reads the real date. The clock lives in a separate place, the gym: a shell session opened at a past date, where time says that date and two programs exist that exist nowhere else. bet records a probability that something will happen, and fast-forward turns the clock on, resolving every bet whose answer has been published by then. Every other program follows one rule there: answer only from what had been published by the gym's date. SEC filings make it easy, because every one carries the day it was filed. Add a program for share prices, as printed on the day, and an agent can reason about a company's value the way an investor could have then.

Set the clock to 2025-10-10 and ask: will Vertex Pharmaceuticals' next annual report show a public float (the market value of the shares the public holds) above $120bn? With the clock set, the agent sees 3 years of it, climbing from $71.8bn to $121.8bn. The trend says yes.

Here is what the two annual reports actually say, word for word.

The report filed on 2026-02-13 said $113.4bn. The answer is no. An agent that followed the line loses points; one that hedged keeps more. Nobody had to grade it: the answer is a number in a filing.

In reinforcement learning, an episode is one go at the task, from start to score. With a time machine, it is simply a shell session. The agent enters a shell whose clock reads 2025-10-10, does whatever research it likes, places its bets, and leaves. Then the clock turns forward, one filing at a time. Each time a bet's answer is published, the bet resolves and pays out. When the last one resolves, the session closes on a result card.

Try it below as the agent: 4 real biotechs, each asking whether the next annual report shows a higher public float than the last one filed before the clock. The answers are what the companies actually filed.

Before building anything, we simulated the design with the figures EDGAR already publishes: every US filer's public float for five years. One episode per year, with the clock at 1 April and one bet per company: will its next float be higher than its last? That made 15,277 bets over 4 years, every one graded by a later filing.

The RL players learned from payouts alone: bet, fast-forward, collect. Each was trained on three years and scored on the fourth, which it never saw, for each year in turn, and five times over. Scores are the average payout per bet, from 0 to 1; the coin always scores 0.75.