Robo-advisor vs AI trading: what actually keeps learning

Robo-advisor vs AI trading comes down to one question: after the thing is switched on, what is allowed to change? A robo-advisor sets its rules in advance and reviews them on a schedule. Finforge's Sophon-3 paper, version 1.0, published in July 2026, does the reverse in one specific way. It freezes the rulebook and lets the numbers inside keep updating while the market runs.

That sounds like a technicality. It is the entire argument.

Robo-advisor vs AI trading: which part is locked?

Section 4 of the paper, titled Formal Problem Statement, names what gets locked. The paper calls it the procedure and lists nine things inside it: the graph, the data contract, the universe rule, the benchmark, the portfolio and cost constraints, the horizons, the metrics, the search spaces, and the seed and state rules. That is the rulebook. Every decision the system is allowed to make has to come out of it.

What the paper does not freeze is the parameters. In its own words, Sophon-3 does not freeze parameters during evaluation. The states inside each step keep moving, but only through the update functions the rulebook already specified. So the machine can change its mind about a stock. It cannot change the rules it used to get there.

The freeze also has a date on it. The paper says that dated freeze is what makes a clean out-of-sample window possible. Out-of-sample means the stretch of time after the rules were locked, so nobody can tune a strategy to answers it has already seen. Lock first. Trade second. Publish either way.

What AI trading strategies are allowed to know on the day

Before it defines anything else, the paper defines what counts as information at the moment a decision is made. The list is short and very literal: closed price bars, available fundamentals, published news, index membership, corporate actions, the outputs of earlier steps, and the system's own saved state. Closed bars. Not a price that is still moving.

A forecast, a shortlist of stocks or a set of portfolio weights is admissible only if it was built from that list and nothing else. Admissible is the paper's word for allowed. Everything outside it is peeking, and peeking is how a strategy ends up brilliant on a chart and ordinary with real money.

No lookahead, written as a rule a computer can check

Sophon-3 is built as a graph of steps, and the paper calls each step a causal operator, meaning it can only use what already existed when it ran. The graph has to be a DAG, short for directed acyclic graph: arrows point forward, nothing loops back into itself. The paper says the graph is admissible when every step's output at a given timestamp uses only information available at that timestamp, and calls this the mathematical form of no lookahead.

The rule does not care about the shape. Several steps feeding one step is fine, which the paper calls fan-in. One step feeding several is fine, fan-out. The only structural requirement is that the graph stays a causal DAG.

Two details are worth holding on to. Feedback across days is allowed: a forecast made today can feed tomorrow's stock selection. A loop inside the same timestamp is not, because the information would end up depending on itself. Circular reasoning, with a ticker attached.

How the result gets scored, and why the benchmark is in the equation

The objective in the paper is benchmark-relative. Active return is the portfolio's net return after modeled costs minus the benchmark's return over the same period, measured on the same dividend basis. A good-looking year is not a result on its own. It is a result once you know what the benchmark did over the same weeks.

Alongside that, the paper reports two risk-adjusted active measures, the information ratio and active drawdown, and two forecast diagnostics. The rank information coefficient, rank IC, asks whether the order of the predictions matched the order of what actually happened. NDCG@K is a top-K ranking-quality score, so it asks the same question about the handful of names at the top of the list. Getting the direction right is not enough. The order has to hold up too.

Questions people ask

What is the difference between a robo-advisor and AI trading? A robo-advisor allocates your money using rules decided up front and revisited on a schedule. An AI trading strategy of the Sophon-3 kind keeps updating its own internal state between decisions, inside a rulebook that was locked and dated before the run started. One holds its view. The other revises it.

Does freezing the procedure stop the system improving? No. It stops the procedure being quietly rewritten halfway through a test. The states keep updating through the update functions the procedure already described, and changing the procedure itself means starting a new dated window.

Why measure against a benchmark at all? Because markets move everyone at the same time. Active return takes out the part you would have had by simply owning the index, and shows what the decisions added, or cost.

Finforge runs four Sophon agents in public and publishes what they do, wins and losses. They have traded since April 21, 2026, on founder capital in Alpaca paper accounts, so no customer money is being traded before launch. You can check the live results, benchmarks and drawdowns for every Sophon agent, and the Sophon-3 research summary sets out the full problem statement. Past performance is not a guide to future returns.

The rulebook was locked with a date on it. The results are public either way.