AI trading strategies: the 2 dates that make one checkable

AI trading strategies have a trust problem: how do you know the results were not picked after the fact? Finforge's answer, in the Sophon-3 paper, Version 1.0, from July 2026, is oddly concrete. In an appendix it prints two dates, its reproducibility anchors, that give you a real way to check.

An AI trading strategy that tells you when it froze

The first date is the configuration data boundary: 31 December 2025. Everything about the system, every choice of data and tuning, was made using information from 2025 only. The second is the architecture freeze: 21 April 2026. On that day Finforge locked the evaluation procedure and the design, anchored by configuration and checkpoint manifests timestamped at or before that date. After those dates, none of the system was picked with knowledge of what would happen next.

The windows you can measure against

Once the design is frozen, the point is that you can slice the results by when they happened. The full ablation window runs from early 2023 to 1 July 2026, roughly 855 to 870 trading days per flow. The post-freeze slice runs from 21 April to 1 July 2026, about 49 trading days. The real-money window runs from 27 April to 24 June 2026, about 41 trading days. The distinction matters: only what happened after the freeze is a fair out-of-sample test of a locked design.

Why the benchmark is set the way it is

The paper also fixes how it compares itself. Three live agents carry forecasts, Apex and Core on the Nasdaq-100 and Surge on the S&P 500, while a fourth live agent is screener-only. Benchmarks get a symmetric dividend treatment: simulated backtests, which do not credit dividends, are compared to price indices, while the live account, which does receive dividends, is compared to the total-return index. Same rule on both sides, so the comparison is fair rather than flattering.

What the freeze dates actually buy you

The real reason an AI trading strategy publishes its anchors is that it removes the wiggle room. If a design was fixed on 21 April 2026, then any result after that date is a test, not a selection. The companion papers disclose the gritty detail, the selection rule, the search budget, and how many candidates were considered and rejected. Finforge runs four Sophon agents in public and publishes what they do, wins and losses, since April 21, 2026, on founder capital in Alpaca paper accounts. No customer money is traded before launch.

A freeze date is not a promise of future returns. It is a promise about process: that the results you are reading were not arranged in hindsight. That is a small, checkable thing, and it is the whole point.

Questions people ask

What is a reproducibility anchor in an AI trading strategy?

An anchor is a dated event. Finforge's Sophon-3 paper uses two: the configuration data boundary on 31 December 2025 and the architecture freeze on 21 April 2026. They mark the point past which results count as a real out-of-sample test.

Why do the two dates matter?

Because everything chosen before them could have been picked using past data, and everything after them was decided blind. Only the post-freeze slice, about 49 trading days from 21 April to 1 July 2026, is a fair test of a locked design.

How long is the real-money window?

About 41 trading days, from 27 April to 24 June 2026. The paper keeps it short on purpose and says so.

Check the live results, benchmarks and drawdowns for every Sophon agent and read the Sophon-3 research summary for the full method. Past performance is not a guide to future returns.