Like an army of analysts who judge every stock by the same rules — then read each one’s situation differently, and correctly.

A fixed group of 100 large, liquid US companies, all of them eligible for the DARWIN platform. It doesn't roam the entire market chasing tips; it works the same defined universe every day, which keeps its behaviour consistent and measurable.

One deliberate exclusion: the "Magnificent 7" mega-caps (the handful of dominant names like Apple, Microsoft, Nvidia and their peers). The reason is about signal quality. These companies generate an enormous volume of news and attention every single day, most of which is noise rather than genuinely price-relevant. Because the model judges every stock relative to the others, letting a few names carry that outsized, noisy information load would distort the comparison. Removing them keeps every stock on a comparable footing — so the model's relative judgements stay meaningful rather than being dominated by whichever mega-cap is in the headlines that day.

It's deliberately not supervised prediction — there's no model trying to answer "will this stock go up?" It's reinforcement learning: the model learns a policy for managing the whole portfolio, looking at the entire universe at once and judging each share against the rest.

It's rewarded for after-cost returns. Real trading costs are inside what the model is trained to maximise, so it can't win by racking up trades that fees would eat in a live account.

And it's deliberately small. A few years of daily data is not much to learn from, and a large model memorises the past instead of learning from it. The current generation was rebuilt from a basic model and grown only where the evidence on unseen data said so.

By trial and error across the earlier years of market history — the model proposes portfolio decisions, sees how they would have played out after costs, and gradually improves the policy that produced them. Several independently trained copies are averaged into the model that is judged.

The specifics of the architecture, the exact inputs, and the training and validation regime are proprietary and deliberately not detailed here — but the principles that keep it honest, especially the out-of-sample discipline and the fair benchmark, are described throughout this section.

Strictly in time order. The model trains on the earliest part of the history; every choice is made on two later validation periods, with the most recent counting double; and a final period is locked in code — any attempt to read it before every choice is frozen stops the run. It is opened once, and the results are reported on it.

The deployment bar is intentionally severe: a model must hold up in its worst out-of-sample stretch, not just look good on average, and must survive higher trading costs than it was tested with. Average competence isn't enough — resilience is the test.

The benchmark isn't cash. It's buying and holding the same 100 stocks in equal amounts, so the model only earns credit for doing something genuinely smarter than holding them all. Every input is built only from information available before that morning's open, and the trading simulator itself uses only past prices — the lesson of the audit that retired the previous generation.

Daily prices and volume, company results and events, analyst views, and news — from an institutional-grade market-data provider, which avoids a whole class of mismatch problems that come from stitching vendors together.

News is not a positive/negative score. Each morning, before the open, a language model reads the recent articles on every share and estimates how likely a big move is, how big, and when the catalyst lands. It deliberately does not judge direction: language models have read the history they would be scored on, and direction scores that looked strong on past data collapsed on periods after the model's knowledge ended. Size and timing held up. Direction is left to the trading model, which weighs the news against everything else — § 06 shows it doing so, trade by trade.

Not DCF, and not analyst-style valuation. Fundamentals enter as features drawn from each company's latest results and analyst views, and as cross-sectional ranks — where each stock sits relative to the rest of the universe on each dimension, each day. The model learns how much weight those deserve in each market regime, rather than following weights hard-coded by hand.

The cross-sectional ranking turned out to matter a great deal. In the Feb–May 2026 semiconductor selloff, the names that recovered were separable on analyst ratings and earnings revisions — but not on price momentum alone. A model looking only at price would have missed the distinction the fundamentals made visible.

These are the traps that quietly wreck backtests. Handled one by one, including the honest limitations:

Survivorship

The honest answer: a fixed group of currently-listed large US companies, so universe selection does contain hindsight. Two mitigations. The benchmark is buy-and-hold of the same group, so universe-level survivorship largely cancels out of the alpha measurement — the model only gets credit for beating a basket carrying the identical bias. And these are mega/large caps, where delisting risk over the window is low. A known limitation, not a hidden one.

Reporting lags

Features are timestamped to when information was actually available, not to the fiscal period it describes — so the model is never fed a figure before the market would have had it. Automated checks throughout the pipeline are designed specifically to catch this kind of look-ahead.

Changing news bias

The subtlest trap with language models: they have read the history they are tested on. So the news scorer is never asked which way a stock will move — only how big a move is likely and when — and that measure was checked for any fall-off after the language model's knowledge cut-off. It showed none; direction scores did, and were dropped.

Corporate actions

Prices are split- and dividend-adjusted at source, and the fixed large-cap universe is checked for symbol-breaking events rather than assuming them away.

Symbology

Single primary data vendor, single broker, one fixed set of US tickers — so there's no cross-vendor mapping problem to get wrong. A deliberate scope decision, not an oversight.