Is AI trading transparent? Sophon-3's claim-by-claim audit

Is AI trading transparent? Only if the vendor gives you a way to check it. Finforge's Sophon-3 paper, Version 1.0 published July 2026, is built around that check. It scores every claim in the introduction with an evidentiary status, prints a column listing what is not yet established, and says in the open that its live record is short and not yet statistically conclusive. That is a transparency test most trading models would fail.

The paper starts from an uncomfortable fact. The central problem in systematic investing is not prediction, it is prediction while the game keeps changing. New data-generating mechanisms appear, observability is incomplete, markets feed back on their own signals, and evidence about whether a signal helped arrives late. A strategy that performs well in a static backtest can fail when the regime changes, the eligible universe drifts, costs rise, or live execution diverges from research.

Is AI trading transparent about how it was built?

Finforge's answer is a causal, continually learning decision graph. A frozen predictor is a model whose parameters are set after an estimation period and then deployed unchanged. A static rule is a fixed heuristic with constant thresholds. Both are common in quant practice. The paper argues that adaptation should reach every component, so universe selection, neural forecasts, ensemble weights and portfolio sizing all keep adjusting. The whole graph stays auditable under a frozen evaluation procedure and under point-in-time data, the data as it stood on the day rather than the tidied version in a vendor database now.

How Sophon-3 makes itself checkable

The core trick is where the freeze happens. The paper defines leakage-free evaluation as freezing the procedure, not freezing the model state. The forecast ensemble adapts, but its weights depend only on past realized losses, what the paper calls no-future-bleed, so the blend is settled before the results it is judged on arrive. The evidence framework dates the architecture freeze, separates results by provenance, and includes a real-money out-of-sample window. Each claim is tested at the final tradable-portfolio level, not on a convenient subcomponent.

What Sophon-3 refuses to claim

This is where transparency gets loud. The paper does not assert that any model beats all markets in all regimes, that a positive backtest proves future profitability, or that continual learning eliminates market risk. The scope of claims is a list of refusals, and every claim in the introduction carries a status: Strong, Strongest empirical, Moderate, Not yet proven, or a precise Null. Each row in the claim hierarchy names what is not yet established, plus what still depends on outside work such as a companion parity paper.

Where the checks land

The ablations put numbers on the audit. An ablation removes or freezes one part of an agent, and effects are paired daily differences of active return, agent minus variant, averaged across the three agents for the pooled figure. Pooled, removing the Screener cost about 45% of active return a year and was significant in all three agents. Forcing equal-weight sizing instead of sizing by forecast magnitude cost about 36% a year. Letting the Screener re-search its rules every month added about 23% a year over rules fixed once at inception. Two checks came back empty. Continual forecast retraining measured about +4% a year with a 95% interval spanning zero, from -5% to +13%, so the paper does not claim it. Adaptive ensemble weighting measured -0.0% give or take 0.6%, a precise null kept as low-cost regime insurance. Active return is portfolio return minus the return of the benchmark universe the agent trades.

Questions people ask

Can you check the results yourself?

Yes. The paper includes a real-money out-of-sample window, and Finforge runs its agents in public with wins and losses shown. Four Sophon agents have traded since 21 April 2026 on founder capital in Alpaca paper accounts, so no customer money is traded before launch. Check the live results, benchmarks and drawdowns for every Sophon agent, and the Sophon-3 research summary carries the full claim hierarchy.

What does not yet statistically conclusive mean?

It means the live window is too short to tell whether a result is edge or luck. Finforge's paper says so on its own: the live track record is short and not yet statistically conclusive. Short windows are noisy, so the record is treated as evidence the system operates as designed, not as a verdict on future performance.

Finforge Research wrote the Sophon-3 paper and runs its trading agents in public, wins and losses included. These are simulated comparisons over the same sample, not an account statement. Past performance is not a guide to future returns.