AI trading strategies: what a release refuses to claim yet

AI trading strategies live or die on trust, yet the way most of them are sold hides the part you should see most: what the maker is not claiming yet. Finforge's Sophon-3 paper, Version 1.0, published in July 2026, puts that in its own section. It limits the release to what has been computed and verified: the formalism, the three-flow ablation suite with paired statistics, the populated per-agent results, and the real-money proof-of-concept. Everything else is named, dated and parked.

Why scope matters for AI trading strategies

A scope section is a promise about what a claim does not cover, and it is the cheapest way to tell a serious team from a hype team. Anyone can publish a rising chart. Very few will print a list of the checks they have not finished yet. Version 1.0 does exactly that. It draws a firm line between results it is ready to defend and diagnostics it is still running.

The checks the August 2026 revision holds back

The paper saves its overdue homework for one dated release rather than scattering promises. The August 2026 revision adds it all in one place. It quantifies what the no-future-bleed rule removes, the leakage a careless design would let in. It stresses costs under Interactive Brokers' commissioned schedule, not just commission-free Alpaca. It runs cadence-matched ablations to isolate the Remove-Forecast residual, and it adds a table of turnover, execution and realized weight drift. It brings forecast-ranking diagnostics: rank IC, NDCG at K, coverage and effective ensemble size, plus directional and magnitude tests.

How Finforge locks a freeze date to the outside world

Then come the exhibits a reader can audit by eye. The live account gains a daily series: an active-return path, a window information ratio with a confidence interval, and an active drawdown. The live panel gains rolling information-ratio and correlation exhibits. A holdings-versus-index dividend-yield differential shows whether the book's yield gap is real. A public hash commitment anchors the freeze on 21 April, so the date cannot be quietly moved. And two external heuristic baselines, a momentum top-K and a random-screener placebo matched for breadth and turnover, give the numbers something naive to beat.

What a companion paper means

Two companion papers are in preparation, one on live-versus-backtest parity and one on screener configuration and pre-registered agent selection. A scope section tells you these are coming, so a gap in Version 1.0 reads as planned, not as an accident. An information ratio, the risk-adjusted edge per unit of volatility, only means something against a stated benchmark and window, and Version 1.0 puts those in place. The discipline is the point: name what is verified, name what is next, and let a reader check both.

What Finforge publishes today

Finforge runs four Sophon agents in public and publishes what they do, wins and losses, since 21 April 2026, on founder capital in Alpaca paper accounts. No customer money is traded before launch. 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.

Questions people ask

What does a scope section in an AI trading strategy paper do?

It names what a release will not claim. Version 1.0 of Finforge's Sophon-3 paper limits itself to the formalism, the three-flow ablation suite with paired statistics, the populated per-agent results and the real-money proof-of-concept, and parks the rest in a dated August 2026 revision.

Why does Finforge freeze its design on 21 April?

A dated freeze fixes the evaluation rules before results come in, so the numbers cannot be tuned after the fact. A public hash commitment anchors that freeze, making the date externally provable rather than asserted.

What are the external heuristic baselines?

A momentum top-K and a random-screener placebo, both matched for breadth and turnover. They give the model's results a simple and a random yardstick to beat before the revision publishes them.

Published as a research paper. Coming to your phone.