What arrives
The desk's screens carry four separate feeds, and none of them is optional. Grid telemetry comes in from SCADA systems at sub-minute resolution: line flows, transformer temperatures, breaker states. Outage notices arrive from transmission operators and generators, some scheduled weeks ahead, some filed at 2am when a relay trips. Weather reanalysis updates conductor thermal ratings, because a transmission line's safe carrying capacity is not a fixed number — it rises when the wind picks up and cools the conductor, falls on a still, hot afternoon. Regulatory filings, mostly through the relevant ISO or RTO's market notices and FERC tariff postings, tell the desk what the rules governing all of the above currently are, and those rules change more often than intuition suggests.
None of these four streams, on its own, tells you the thing that actually prices a spread trade: whether the transmission network, as a graph, currently has a path of adequate capacity between two nodes. That is a property of the whole structure, assembled from all four streams at once, and it can flip without any single feed showing an unusual reading.
What is held
The position under discussion is a locational spread: long at one pricing node, short at another, sized against the belief that a transmission constraint between them stays binding. When a constraint binds, power cannot flow freely from the cheap side to the expensive side, and the price difference — the basis — persists and can be captured. The trade is, structurally, a bet on a cut in the network graph: a set of lines whose combined outage or de-rating separates the grid into two components that cannot fully exchange power.
That belief is not stored as a scalar. It is stored with provenance — which outage filings, which SCADA readings, which rating updates support it — and with a decay function, because a constraint confirmed live forty minutes ago is a stronger belief than one confirmed six hours ago. The position is sized against the current strength of that belief, not against a number frozen at the moment the trade was put on.
The threshold hiding in the topology
This is where percolation stops being an analogy and becomes the actual mechanism. Treat the transmission network as a graph: substations as nodes, lines as edges, each edge carrying a capacity. An outage removes an edge, or a de-rating shrinks it. Below some combination of outages, the graph still has a path of sufficient capacity from the cheap node to the expensive one — the constraint holds, the spread survives. Above that combination — one more line restored, one rating uprated by a cold front — the graph reconnects with enough capacity to arbitrage the spread away, and the price difference collapses, often within the hour, not proportionally to how much capacity came back but categorically, because a new path opened.
This is the mechanism Broadbent and Hammersley described in 1957, working on gas diffusing through carbon granules in a respirator filter, and that Kesten later pinned down exactly — bond percolation on the square lattice crosses at density one half, sharply, not gradually. Real transmission networks are not square lattices; they have hub substations and long radial branches, and the honest objection is that classical lattice thresholds do not transfer cleanly to them. That is correct, and it does not help the case for ignoring the mechanism — it sharpens the case for watching it, because in a heterogeneous, hub-heavy network the crossing depends on which edges return, not on aggregate megawatts restored. A single 345kV line coming back into service at the right substation can reconnect two components that ten smaller lines returning elsewhere would not touch.
What triggers revision
Revision is triggered by any of the four streams contradicting the current belief, and the desk's system is built to treat these asymmetrically. A scheduled outage notice extending its return date strengthens the constraint-binding belief and lets the decay clock reset. A SCADA reading showing flow on a nominally out-of-service line — a crew finishing early, a breaker reclosing ahead of the filed schedule — triggers immediate revision downward in confidence, because it means the cut may already be closing. A weather reanalysis update showing a cold front arriving overnight, raising thermal ratings across the constrained corridor by 15–20%, is itself sufficient to trigger a revision even with no outage notice at all, because the effective capacity of the existing lines has changed.
The constraint was filed as binding through Thursday. Nothing in the filing changed. Why would the position need revisiting before then?
Because the filing describes intent, not current capacity, and capacity is a function of temperature, wind and whichever crew finished ahead of schedule. The filing is one edge weight among many; the belief that matters is the state of the whole cut, and that is exactly the kind of global, connectivity-dependent quantity that a single document cannot certify.
What the operator sees
The desk quant does not see a static constraint table. The interface shows the position, the current belief about the constraint — expressed as a probability the cut remains below the effective capacity needed to hold the spread, with a timestamp on last confirmation and a visible decay curve — and a log of which feed last touched that belief. When an outage notice posts a line's early return, the system does not wait for the next scheduled review; it recomputes the graph's cut capacity immediately, flags the position, and shows the estimated new basis alongside the old one. The quant's job at that point is judgement about sizing and exit, not detection. Detection has already happened, continuously, without anyone needing to have been watching a screen at the right minute.
Two objections worth answering
The stronger version of the "wrong model" objection is that transmission networks have hubs and correlated failures, so there is often no single sharp p_c to miss — connectivity can be fragile at almost any density once a hub line is involved. That is true, and it cuts the other way from how it is usually meant. A vanishing or hub-dependent threshold does not mean the system is safe from sudden reconnection; it means the reconnection can be triggered by very few specific edges, which makes it harder to infer from anything short of live, edge-level data. A corpus that recorded last month's outage pattern cannot tell you which single line, out of forty on outage, is the one whose return dissolves the cut. Only current, provenanced tracking of each edge can.
The other serious objection is that grids are engineered against exactly this: N-1 contingency standards, reserve margins, redundancy built in so operation stays far from any critical density. That is real, and where margins are genuinely current it does most of the work. The trouble is that "far from criticality" is itself a claim about present topology, and margins computed against last quarter's network topology are not automatically true of tonight's. The 2003 Northeast blackout and the cascading line trips that followed a series of individually survivable outages are the standard illustration: the margin existed on paper against a network state that had already moved. An engineered buffer is a belief with an expiry date, not a permanent exemption.
What it costs
Running the four streams continuously — SCADA integration, an outage-notice parser watching every relevant utility and ISO filing, a weather reanalysis feed re-run against thermal rating models, and a filings tracker against FERC and market notices — is not free, and for a desk trading a handful of small, non-constrained hubs it may not be worth building. But price a specific failure against it. A 500 MW spread position, mispriced by $6 per MWh once a constraint that was assumed binding through Thursday actually cleared overnight, produces a swing of roughly $72,000 across an eight-hour settlement window before the position can be unwound at the market open. Set against that, a monitoring system that catches the outage notice at 11:40pm rather than at the next morning's review is buying back most of that loss for the cost of a feed subscription and some parsing logic. The asymmetry is the entire argument: the crossing costs tens of thousands of dollars per position when missed, and the intake needed to catch it costs a fraction of one such loss, repeated across every position exposed to that interface.