The loop, run once
Take a single position: a spread between two nodal prices either side of a transmission constraint, entered because a line is derated for maintenance and the desk quant expects congestion to widen. Follow that position from entry to exit and the whole argument about intake becomes visible as a sequence of concrete events, not a thesis.
The position is not a belief held once and then defended. It is a belief that must stay answerable to a grid that keeps moving under it. Situated and embodied cognition, applied to a trading desk rather than a body in a room, says the same thing Gibson said about a walking animal: the relevant structure is not stored once and consulted, it is picked up continuously from an environment built to be probed. The desk is the body. The feeds are the coupling.
What arrives
Four streams run into the desk simultaneously, at different rates and with different half-lives.
Grid telemetry arrives fastest: SCADA sweeps, four-second real power flows, frequency, the transmission operator's contingency table showing which lines are currently derated and by how much. Outage notices arrive as discrete events: a planned maintenance window, a forced outage after a fault, a rescission of either. Weather reanalysis arrives on an hourly cadence, feeding wind and solar output forecasts that determine how much power needs to cross the constrained corridor in the first place. Regulatory filings arrive rarely but decisively: a system operator's emergency order, a tariff change, a curtailment instruction that overrides commercial dispatch outright.
None of these streams is optional and none substitutes for another. A forecast without the current outage table is a forecast of the wrong grid. An outage table without weather is a static fact about a corridor whose load nobody can predict. The position's meaning — what "the constraint" refers to — is fixed jointly by all four, continuously, not by any one of them read once at the start of the trade.
What is held, and how it decays
The desk holds a belief, not a fact: "the corridor between node A and node B is constrained until Thursday 06:00, with 340MW of derated capacity, because Line 7 is under planned maintenance." That belief has provenance — which outage notice, filed by which operator, timestamped when — and it has a decay profile. A planned maintenance window is reliable for days. A forced-outage return-to-service estimate is reliable for hours. A weather-driven wind forecast feeding into flow expectations is reliable, at usable resolution, for perhaps six.
The quant's model does not store "the constraint is active" as a fact. It stores a claim with a clock attached. That is the discipline situated cognition demands and unstructured belief-holding does not: every claim about a moving world needs a marked expiry, not because the desk is being cautious, but because the claim was never a fact about the world, only a fact about the world as of the last observation.
The trigger: a constraint lifted overnight
Here is the failure the domain actually produces. Line 7's maintenance was scheduled to run through Thursday. At 23:40, the transmission operator completes the work early, tests the line, and returns it to service — an unplanned rescission of a planned outage, filed as a routine notice, not flagged as urgent because nothing failed. The notice sits in a queue alongside forty others. The quant's overnight batch process, built to re-price positions against the day-ahead weather update, does not re-read the outage table, because the outage table "doesn't usually change" between the evening close and the morning open.
By the time the desk opens, the corridor is unconstrained. The congestion the position was built on has nowhere to come from. The spread the quant is holding is priced against a topology that stopped existing seven hours earlier. Nothing about the model failed. The belief was correct when it was formed and correct for as long as the coupling that produced it stayed live. The coupling was cut overnight by a batch schedule that treated one stream as slower-changing than it is.
This is the exact shape of the Held and Hein result, restated in market infrastructure. Passive intake — a nightly pull that assumes yesterday's topology holds — produces the same visual field as active, continuous intake, right up until the moment it does not. The deficiency is invisible until the position is marked and the mark is wrong.
What the quant sees, and what it costs
What the quant sees at 08:15 is a position that has moved against the book with no headline to explain it. The screen shows a spread collapsing, not a reason. Reconstructing the reason means walking backward through four streams to find the one notice, filed eight hours earlier, that changed the premise. That reconstruction is itself a coupling failure: the desk is now doing after the fact what should have been running continuously — checking whether the constraint that licenses the position still holds.
The cost is not abstract. It is the difference between the entry spread and the exit spread on a position sized to a corridor's derated capacity, times however long it takes the desk to notice. On a corridor with meaningful nodal price separation, a 340MW derating disappearing overnight can move the spread most of the way to zero before the local market absorbs the news. The desk pays for the gap between when the world changed and when its belief was told the world had changed. That gap is intake latency, priced in real currency, on a single stale premise.
Why this is a coupling problem, not a data problem
The natural objection is that the desk simply needs a faster pipeline: poll the outage table more often, alert on any status change, done. That is real and worth doing, but it treats the failure as a plumbing defect rather than what it is. The deeper issue is that the position's meaning was never fixed by the outage notice alone. It was fixed by the standing relation between the position and four moving streams, and a Large Language Model style architecture — reason well over what you were given, once, at a cutoff — cannot hold that relation at all. A model retrained nightly on yesterday's filings is a Large World Model at best: it senses, prices, and acts, but only for the scene it was given, and the scene goes stale the moment the transmission operator files something new.
What the domain actually needs is closer to the third position on this ladder: every stream still running, the constraint belief carrying its own provenance and its own decay clock, revision triggered by the arrival of the rescission notice itself rather than by the next scheduled batch. That is not a claim that such a system trades better in some general sense. It is the narrower claim that a position referencing a physical grid state is only as good as the desk's live access to that state, and no amount of skill in pricing spreads compensates for referencing a corridor that no longer exists.
Two objections worth taking seriously
Every real desk samples, filters and compresses. Nobody actually watches every SCADA sweep on every corridor. Calling continuous multi-stream intake "terminal" just describes a bigger polling loop, not a different kind of cognition.
This is fair and should not be waved away. No trading system, human or otherwise, achieves total coverage; every desk decides which corridors matter enough to watch closely. What makes the third position categorically different is not coverage, it is posture: a standing commitment that any stream may revise any belief, with no scheduled stopping point and no stream presumptively exempt from re-checking. A nightly batch job is a different category from a continuously polled feed, even if both eventually sample the same outage table, because the batch job has a built-in blind window and the continuous feed does not. Adding a fifth stream to either does not change which category it is in.
Regulatory filings and outage notices are text. Distributional structure in filings language already tells you a lot about what tends to follow a "planned maintenance" notice versus a "forced outage" notice, without needing a live feed at all.
Granted, and it is a genuine efficiency: pattern in filing language does carry real signal about typical duration and typical rescission risk. But that signal describes the population of past notices, not the state of Line 7 at 23:40 on a particular Tuesday. The rescission that sank this position was a one-off, filed once, about a single named asset. No amount of distributional structure recovered from historical filings substitutes for reading that one notice when it arrives. Text about the past is not access to the present, however well it is modelled.