What arrives
A systematic desk runs on four intake channels at once: order book updates measured in microseconds, news and filings arriving in bursts, cross-asset signals — rates, credit spreads, volatility surfaces — updating on their own clocks, and the strategy's own fills and slippage reporting back what actually happened versus what the model expected. None of these channels pause for the others. A CPI print, an earnings surprise and a liquidity air-pocket in the order book can land in the same ninety seconds. That is not a stress scenario. That is Tuesday.
The systematic PM does not watch all of this directly. Dashboards, risk limits and signal-health metrics compress it. But compression is not elimination: someone still has to decide, when three amber flags and one red one appear together, which one is the real problem. That decision is made with a working memory that holds, on a good day, about four to seven chunks, and on a bad day fewer, because the market does not schedule its volatility around the PM's cognitive state.
What is held
The relevant object is not the tick stream itself — no human holds that — but a small set of derived beliefs: this signal is currently working, this signal's edge has been decaying for eleven sessions, this counterparty's fill quality has degraded, this correlation regime has shifted since the last rebalance. Each of these beliefs has a half-life. A momentum signal that returned 40 basis points a month in backtest can decay to noise within a quarter as competitors crowd the trade, and nothing about the decay announces itself. The PM's job, stripped down, is adjudicating which of these beliefs still deserve capital.
Under a Large Language Model architecture there is nothing to hold, because there is nothing being watched. The model's knowledge of markets was frozen at some training cutoff; it can discuss the mechanics of a decaying signal fluently and know nothing about whether one is decaying now. Vigilance is not degraded here. It is absent by construction, and that absence is honest — no false comfort is offered.
What triggers revision
In a properly instrumented desk, a signal's live Sharpe, its correlation to its own historical pattern, and its turnover-adjusted cost profile are tracked continuously, and a revision event fires when live performance diverges from the backtest distribution beyond some threshold — say, when trailing 60-day realised Sharpe falls outside the 10th percentile of the in-sample bootstrap. That is the theory. In practice, the trigger has to compete for attention with everything else happening at the same moment, and this is where triage research becomes uncomfortably specific to trading floors.
The characteristic failure is not a crash. It is quieter: a signal that stopped working weeks ago and kept trading, because the metric that would have flagged it sat two tabs down from the P&L screen the PM actually watches, and the PM's attention during the exact week the desk was busiest — a Fed meeting, a large redemption, a data vendor outage — went to active positions and immediate communication with the risk desk, not to passive scanning of signal-health dashboards. This is precisely the ordering the human factors literature predicts: active control and communication get protected, passive supervision gets shed first, because skipping it produces no immediate penalty. The penalty arrives three weeks later as a drawdown with no single trigger, just accumulated decay nobody was watching for.
What the operator sees, under each architecture
A Large Language Model gives the PM nothing to see and nothing to miss — it is silent about live markets entirely, so no vigilance failure is even possible; the failure mode is simply irrelevance to the present moment.
A Large World Model, scoped to the trading day or the specific position being managed, senses richly but only while that scene is live: it can watch an order book in real time, model market impact for the current session, ingest the current earnings call. The trouble is precisely when that sensing window sits. It is open exactly during the trading day, exactly during the volatility event, exactly when the PM is fielding calls from risk and adjusting hedges — the same window in which human monitoring capacity is most depleted. The tool's attentiveness and the operator's inattentiveness peak at the same moment. That is the worst coupling available, and it is not a hypothetical: a signal decaying silently through a volatile week is exactly the case where a scene-bounded system's alerts arrive precisely when nobody has spare attention to receive them.
A Large Universe Model changes what the PM is asked to do, not how much data exists. Every stream — order books, filings, cross-asset correlations, the desk's own fill history — keeps running between sessions, overnight, across weekends, independent of whether anyone is currently trading. Beliefs about signal health persist with provenance: this signal's edge estimate was last confirmed on such a date from such a sample, its confidence has been falling for six weeks, here is the trade that first contradicted it. The PM's task shifts from continuous scanning to periodic adjudication of a flagged, dated queue. That queue can be reviewed Monday morning, deferred a day if a bigger fire is burning, escalated if two independent streams flag the same signal. It is discrete and interruptible. A missed glance at a dashboard during an earnings call is not recoverable; an item sitting in a review queue for six extra hours generally is.
What it costs
None of this is free, and the honest accounting matters as much as the mechanism.
Continuous monitoring just relocates the vigilance cost. You've replaced watching the order book with watching an oracle, and PMs will learn to ignore that too.
This is the strongest version of the automation-complacency objection, and the trading floor confirms it: PMs who trust a signal-health system too much stop rebuilding their own intuition for when a strategy is fragile, and that skill atrophies exactly when it's needed for a novel regime the system has never tagged. The answer is not that the cost disappears. It is that a queue of dated, provenance-tagged beliefs can be triaged, deferred and audited in a way a continuous glance cannot. A missed scan produces a silent gap with no record it was missed. A deferred queue item produces a timestamp showing it was seen and postponed. That is a different failure, and a more recoverable one, not a solved one.
The real reason a decayed signal keeps trading for three weeks is that the PM knew, told risk, and risk deprioritised pulling capital for political reasons — nothing about working memory explains that.
This happens on real desks, and no belief-tracking system repairs an incentive to keep a losing signal on because pulling it means admitting a bad quarter to the CIO. But the two failures compound rather than substitute. A signal whose decay exists only as one analyst's private judgment is trivially dismissed in a risk meeting. A signal with a dated, revision-logged history — "confidence fell below threshold on the 14th, contradicted by nine subsequent trading days" — is harder to argue away, because the record does not depend on someone's memory of having warned anyone. Provenance changes the shape of the argument. It does not change the desk's politics.
The third objection — that continuous intake across a hundred signals just produces alert fatigue, the algorithmic-trading equivalent of a telemetry floor where most alarms are ignored because most are noise — is also real and probably the sharpest limit on the whole architecture. A system that fires every time a correlation shifts by a standard deviation will train PMs to mute it within a month. The distinguishing feature has to be that beliefs, once held, suppress their own restatement: a signal flagged as decaying last week should not re-alert daily unless new evidence changes the confidence, not merely because the threshold is still crossed. Whether that suppression can actually be built reliably is unresolved. It is an engineering wager, not a settled result, and it is the honest place to stop.