A Reader’s Compass
An orientation: what the essay is about, why it is structured this way, and what the reader can carry from it.
A number waits on a screen, and the room has gone quiet around it. A system searched a space of decisions too entangled to lay out by hand, returned a recommendation, and now stands ready to move the firm faster than any committee was built to match. The board is diligent, the audit function careful, and the senior people are doing what their roles describe. And still, no one present can say what would have to be true for the recommendation to be wrong. The room has tools built for simpler decisions. They reduce the recommendation to a picture smaller than the decision itself, and the part they leave out is the part where the firm is most exposed, in the tail where the variables stop moving on their own and begin to move as one.
So the recommendation is produced in the large decision space and read in the small decision space. Between them sits an exposure chosen by no one, because no one can see the part of the decision that did not survive the trip across the spaces. This indicts no one holding the tools. The instruments were built well for the world they were built to read. The world changed its register underneath them.
The live debate about governing these systems is conducted almost entirely in the language of values, rights, and alignment, and it assumes the hard problem is deciding what the machine should be allowed to want. The harder problem comes first and is rarely named. Before a firm can judge a recommendation, it has to see the decision in the form the system used to make it. Most firms cannot. The governance gap is an instrument gap before it is a values gap. A firm that builds the shared picture its machines run on can mistake the picture for the governance. It has built only a more confident shadow of the decision, and learns the difference when the unseen part arrives all at once.
This is a worked argument that follows one decision until it fails and names what the room could not do at the moment it needed to. Financial services has already paid for this in the open. Its record is more contested than the usual telling allows. The next sector may inherit both the precedent and the blindness, harder to see because the instruments are newer. Consumer products is living the same condition now, in the interval before its reckoning, and it holds a discipline the sector that paid in the open never had to build. Underneath it is a knowledge gap that becomes a power problem. The firms that recognize the gap first are the ones that begin building the instruments and the tradecraft to close it before the failure forces them, build competitive advantage, and that recognition is the move every firm still in the interval can make.
The Governance That Has Not Yet Arrived
An organization can govern a decision only at the resolution its instruments can represent. What that sentence costs, why the financial sector paid it in the open, and why the rest of the corporate world is living the lesson without its reckoning.
I. The resolution gap
An organization can govern a decision only at the resolution its instruments can represent.
Picture a board reading a chart with four categories when the system that made the recommendation searched a space of four hundred entangled variables. The chart is not wrong. It is too small for the decision it has been asked to govern. This sounds like a remark about tools. The systems now deciding inside firms search a space of entangled variables and stand ready to move the firm faster than any committee was built to match. The instruments the committee holds were built for a simpler moment, when the variables were few enough to enumerate and the question held still long enough to be argued. So the governing bodies read a smaller picture than the decision actually occupies, and the part they cannot hold is the part where the firm is most exposed, in the tail where the variables stop moving on their own and begin to move as one. The reigning debate about governing these systems is conducted almost entirely in the language of values, rights, and alignment. That debate assumes the hard problem is deciding what the machine should do. The harder problem comes first and is rarely named. Before a firm can decide what a recommendation should be, it has to be able to see the recommendation at the resolution it was made, and most cannot, and most do not know that they cannot. The governance gap is an instrument gap before it is a values gap, and a firm that builds the shared picture its machines run on will believe it has built the governance when it has only built a more confident shadow of the decision.
By resolution I mean the level of detail a tool can show. The part that does not survive the trip across is concrete rather than abstract. It is the correlation the model treated as fixed, the funding channel that quietly tied two segments together, the loss that lived past the line the report drew. A body that can hold four variables and the way they pull on each other can govern a decision living in four. Put that same body in front of a decision living in four hundred, entangled through the whole, and what the body can see no longer matches what the decision actually contains. It does not govern the decision badly. It governs a reduction and mistakes that reduction for the whole. Between the recommendation the machine made and the smaller one the committee reads sits an exposure chosen by no one, because no one can see the part of the decision that did not survive the trip across.
This puts a finer point on a problem the polycrisis literature has already named. Nils Gilman has argued that our governing institutions are mismatched in scale to the problems they face, that the nation-state is too small for the planetary problem and too large for the local one, and that the failure is structural rather than a failure of will.1 The argument here borrows the diagnosis and corrects the mechanism. The lag is also a matter of grain.
Governance reads the new world at the level of detail of the old one, returns a clean answer to a question the firm has stopped asking, and the cleanness of the answer is what hides the gap. An instrument calibrated to the wrong scale does not announce its miscalibration. It reports confidence.
The boards are diligent, the audit functions careful, the senior people doing exactly what their roles describe. What they hold was built well for the world it was built to read. The world changed register underneath it. The apparatus and the judgment fail together, because managers do not ask their tools to show what they have not learned to suspect. None of this indicts the people holding it.
There is a power question sitting underneath the epistemic one, and it will not stay underneath for long. The person who can read the decision in the machine’s own terms holds power the governing body lacks. That is a question about who governs the machine, and it is being settled by default, inside firms, while the debate overhead stays fixed on what the machine should be allowed to want. The work returns to that question at the end. It belongs to the conversation Yanis Varoufakis and Shoshana Zuboff are having about where power migrates when the machine becomes the medium through which decisions pass. The resolution argument names the precondition that conversation assumes, that power flows to whoever can read at the resolution the machine decides at. First the gap itself, because the gap is what makes the power question possible.
II. The glass box
In a conference room early in 2026, two people, one from a major management consulting firm and one from a major technology company, described a firm’s generative systems as a glass box, fully explainable, and rehearsed how to walk the client through the explanation as a service the client was owed. The phrase sat on the slide between them. They were operating in good faith. What neither could see was that the thing they were preparing to explain sits beyond what the technology can actually do, that they were assembling, in the language of fundamental understanding, a surface that only looked like an answer.
I have sat in enough of those rooms to know the silence that follows when a board cannot phrase the decision it was convened to govern. The silence comes from a missing tool and tradecraft, not from ignorance. No question in the room’s vocabulary will open the part of the decision that matters. That is the gap in the language of the meeting itself.
The danger was good faith applied to a false surface. Two competent people were preparing to hand a client a confident surface and call it understanding, and they believed it, and the belief is the whole of the problem. A glass box you can see through is governable. A glass box that only looks like one is worse than an opaque box, because the opaque box at least announces what it withholds. The confident surface withholds nothing visibly and everything in fact.
Cynthia Rudin has argued that high-stakes decisions should use models people can follow by construction. Her reason is plain. An explanation added after the fact is a second model, and it can be wrong about the first.2 She is right, and the glass-box scene is what she is right about. The people in that room were preparing to sell the second model, the explanation laid over the system, and to call the explanation the thing itself. Rudin’s distinction is the blade that cuts the false glass box from the true one. A reading the room can reach while it decides is not the same as a story assembled once the recommendation is already made, and the room that accepts the second has accepted a story about the decision rather than the decision.
The criterion forms here, in the room, before any apparatus is named. The board needs an explanation it can use while the decision is still open. Anything can be explained after the fact, persuasively, in good faith, on a slide. What matters is whether the explanation is native to the room that must accept it, reachable while the decision is open, at the level of detail the decision was actually made. If the room cannot use the explanation while deciding, the explanation is only a presentation layer. The board that accepts it has not governed the decision. It has accepted a confident surface in place of one.
That criterion will turn out to be necessary and not sufficient, which is where the argument leaves Rudin behind rather than borrowing her. But the necessity comes first. A decision no one in the room can reach is a decision no one in the room can govern, whatever else is true of it. The glass-box meeting is the case the rest of this will return to, because the gap it shows in miniature is the gap the whole essay is about.
III. One book, two readings
Underneath the gap is the contestable proposition the rest of the argument stands on. Sharing one view of the book does not settle who decides what to do with it. Inside any firm that allocates something scarce across claims that move together, two readings of the same book, the firm’s set of entangled positions and choices, are always running. Point the risk reading at a portfolio and it returns how much could be lost if the holdings fall together. Point the optimization reading at the same portfolio and it returns the arrangement that grows the firm fastest. The two answers can recommend opposite moves on the identical book. One asks what the firm is exposed to. The other asks what the firm should become. The firm houses them in different functions, staffs them with different people, runs them on different clocks. Do that for long enough and the separation starts to look like an accident of the org chart, because both readings stand on a single picture of the book held whole. Build that shared picture, as a firm now can, and the temptation is to conclude that the separation was never principled, that the two readings can be governed as one.
They cannot. The picture is shared. The discipline is not. The shared picture cannot supply the four things that separate the readings, the objective, the horizon, the accountability, and the tolerance for error. The shared picture is only the starting point for governance, and a firm that mistakes the one for the other will believe it has built the governance when it has built only the ground it stands on.
Start with the shared part, because it is real and it is the expensive part. Risk first requires one view of the positions and their interactions held together. Only then can the operator ask how the structure behaves under stress, where the losses concentrate, what happens in the tail, the far edge of the distribution where the variables stop moving independently and begin to move as one. To read the same book for its best configuration requires the identical first move, then asks the same picture for the best feasible arrangement that maximizes the firm’s goal state under its real-world constraints. Neither reading can begin until the picture exists. Build it once and either interrogation runs off it. The lineage is old. Harry Markowitz showed in 1952 that the risk of a portfolio rests in how its holdings move together, not in any single position read on its own, which establishes the entangled whole rather than the individual position as the proper object of attention.3 The number that captures how the holdings move together is covariance. That shared picture is what every later reading stands on. A capital model, a stress test, an optimizer all run off it.
They diverge first in what each is for. Risk minimizes a loss and asks how large the exposure could grow. Optimization maximizes the firm’s goal state and asks what the firm should become once the goals are achieved. One protects what exists. The other changes it. They point in opposite directions, and a picture that serves both does not tell you which way to face. So the readings diverge, and the shared picture reconciles none of it.
Risk runs through the cycle, asking how the structure survives stress that may arrive anywhere across a long uncertain stretch. Optimization runs toward a configuration the firm answers for over its planning horizon. They diverge in the clock’s timescale. A move that is right on the one clock can be a risk the firm cannot carry on the other, and the shared account reports both without warning that they are timed against each other.
They diverge in who owns the answer. Risk answers, in the sector that built the tools, to a chief risk officer holding the whole book against the board. Optimization answers to an investment committee or an operating committee whose mandate is to grow the thing rather than bound its loss. The picture produces both reads, risk on one pathway and optimization on the other, and assigns neither accountability. Assigning accountability is governance, and autonomous systems are beginning to outrun that assignment.
And they diverge in what it costs to be wrong. In a leveraged book a single exposure of variables that move together, read as though they moved apart, can end the firm, so the firm fears the missed loss far more than the missed gain and weighs its tolerance hard against the error that ends it. An optimization that lands on the second-best configuration costs only the difference between best and second-best, an error the firm survives and weighs evenly against its opposite. A firm that governs both as though the penalties matched will under-protect against the error that ends it and over-invest against the one it could survive.
The practitioner does not receive these four divergences as theory. He feels them as work, which is the only way they are ever really learned. Point the loss-minimizing tool at the maximizing question and the book comes back safe and going nowhere. Point the maximizer at the loss question and the configuration comes back excellent on average and fatal in the tail. He knows the risk read goes up to a chief risk officer and the optimization read goes across to a portfolio manager, and he knows that on the day the tool makes it trivial to collapse the two into one screen, the institution must refuse to collapse them, because the divergence the tool erased is the divergence that ends firms. One rendering, two readers, two clocks, two consequences when each is wrong. The apparatus never divided the discipline. The institution did, and the division has to be honored even on the days the tool makes it cheap to forget.
Having paid for the picture, the firm assumes that because the two reads share a rendering they share a governance. They do not. The shared rendering settles none of the four things that divide the readings. That is the limit of unification, and it is the thing a firm building the shared picture is most likely to miss, because the picture is expensive and impressive and feels, when it is finally built, like the answer. The rest is still to build.
IV. What 2008 actually settled, and what it did not
One sector has already paid for this in the open, and the record is more contested than it is usually told. David X. Li published a Gaussian copula in 2000, and the desks applied it to collateralized debt obligations, pricing the joint behavior of defaulting credits off a single correlation that did not move.4 A fixed correlation claims the relationship between two credits holds in every weather. It cannot represent the state a book enters when its parts stop moving on their own and begin to move together. The tool read a world of loosely connected holdings, the names drifting more or less independently, and it read that world well until the names defaulted in concert and the diversification that looked real on the spreadsheet evaporated in an afternoon. The instrument went blind in the tail, precisely where the firm was most exposed. Before 2008 the largest financial firms governed an entangled book with instruments that could not read its structure.
The same blindness lived in the measure the risk report led with each morning. Value-at-Risk named a loss threshold at a stated confidence, the amount the firm was unlikely to lose more than on an ordinary day, and said nothing about the shape of the loss beyond the line it drew. It held the body of the distribution and left blank the part that ended firms. Under ordinary conditions the silence did not matter. Under correlated stress, when the losses lived precisely past that line, the silence was the whole of the danger.
The decade after was the institution’s slow, expensive admission of what its instruments had not held, and one piece of it is a clean, checkable instance of a firm raising the resolution of its tools.5 (see endnote 5 for the apparatus) The Fundamental Review of the Trading Book, finalized in 2019, moved the internal-models measure away from Value-at-Risk toward Expected Shortfall. Expected Shortfall averages the losses past the point where the prior measure stopped looking. It reports the tail rather than marking where the tail begins. That is the move stated plainly. An institution that had governed with a tool blind to the tail replaced it with one that reads the tail. The resolution went up, in exactly the place it had been most dangerously low.
Here the record turns, and the turn is the part the usual telling skips. The instinct that says more resolution is better, that the cure for an instrument that saw too little is an instrument that sees more, is the instinct Andrew Haldane warned against. At Jackson Hole in 2012, in “The Dog and the Frisbee,” he argued that complex environments are often governed better by a few simple rules than by models that add detail faster than they add reliable signal.6 He showed it across domains, from the dog that catches the frisbee on a single heuristic to the regulator drowning in granular risk weights. A model that adds representation faster than reliable signal does not govern better. It governs worse. The point cuts against the resolution argument at its most naive, and it has to be met head on rather than conscripted, because Haldane is not an ally of more detail. He is its sharpest critic, and he is right about the thing he is criticizing.
The macroprudential critics press the same blade from the other side, and they have the record to press it with. The rebuild that was supposed to cure the false confidence may have institutionalized it. Stress testing, run across the largest firms on a common scenario, can produce a model monoculture, every firm reading the world through the same lens and so going blind in the same place at the same time. The forward-looking capital regime can amplify the cycle it was built to dampen, tightening when the system is already tightening. Adam Tooze’s account of the crisis as unfinished and politically unresolved is the standing check on any story that treats the post-2008 build as a debt paid in full.7 Tooze’s Crashed argues the crisis was technically contained, the Federal Reserve swap lines holding the dollar system together when it nearly seized, and politically unresolved. The question of who would bear the loss did not get answered in 2008. It migrated, into the eurozone debt crisis, into a decade of austerity politics, into the political settlements still unraveling. The instrument story is true. It is not the whole story. The decade-after rebuild solved a technical problem and left an institutional one open. The honest position is that 2008 is a template the next sector might inherit and also a warning that the inheritance can reproduce the pathology one level up, with more instruments and the same blindness, now harder to see because the instruments are newer.
The useful distinction is whether the tool makes its blindness visible, and it sorts the two cases Haldane’s warning runs together. Resolution helps only when the instrument declares the limit of its own seeing. An instrument that adds detail and says nothing about that limit is exactly the instrument Haldane warns of, the one that drowns the operator in granularity he cannot check and the monoculture the macroprudential critics name. An instrument that adds detail and declares the edge of its own representation, which interactions it held and which it treated as independent, which axes of the firm it took in and which it left out, is a different thing. Expected Shortfall declares only one edge, the tail the prior measure left blank, and now the region is named. Naive resolution-stacking, more weights the operator cannot check, fails Haldane’s test and the critics’ both. A tool that reports the edge of its own representation passes them, because it gives the operator the one thing the pre-crisis instruments never gave their supervisors, a view of their own edge.
It did not feel like ignorance at the time. I have watched a desk argue with its own screen and lose. Consider the senior trader on a mortgage desk in 2006, the type rather than any one man, the figure many firms held some version of, holding the resolution instrument and the understanding it required at once. He knew the names in the book were not independent. They were tied to the price of American housing, and if that moved the whole holding would move with it. What he could not do was make the apparatus say so. The screen showed correlations near zero across geographies the structure had quietly bound through a common funding channel, a diversified book on its face and a single bet on housing underneath. Tranches built on the same few thousand mortgages, sliced and packaged and sold as different exposures, counted as twenty what was, underneath, one. The tool read each position against a number fixed in calmer weather, and that number told him the book was diversified when his own eye told him it was one bet wearing many faces. He had a hunch and a memory of prior cycles, and the firm did not run on hunches. The screen was winning. What the decade after had to fix was that asymmetry, the institution organized to believe the instrument over the person who could see past it. Fixing it was never a matter of a better number. It was a matter of building something that read the structure the trader already knew was there, and then, harder, of building something that would say where its own reading stopped. The first half got built. The second half is the part still owed, in finance and everywhere the inheritance is now being shipped.
V. A worked failure
Return to the glass-box meeting, because the cleanest way to see what governance requires is to watch one decision fail to get it. The recommendation was a credit-portfolio reallocation across the book. The system returned it with high internal confidence and reasoning no one present could follow. Its strength rested on a covariance the model had declared independent, two segments it treated as drifting apart that a common funding channel had quietly bound together. The screen showed a basis-point pickup against a barely-moved risk number. The two segments sat in different buckets on the slide, as though they answered to different weather, when a single funding channel made one their shared sky. The room could not see that the gain it was being shown lived entirely in the assumption that the two would not move at once. Walk the recommendation through the room and watch what it costs at each step, not as a checklist of properties but as a sequence of consequences the room absorbs without naming them.
It lived in a space too large to enumerate, and somewhere between the system and the slide it was reduced to a handful of variables a committee could argue. The easy compressions are exactly the ones that throw away the entanglement in the tail, where the variables stop moving on their own. The decision begins in compression. So the first thing the room loses is the part of the decision most likely to end it, and it loses this silently, because a compression that drops the tail does not report that it dropped anything. The room is now arguing a smaller decision than the one the machine made and does not know it. Simplification makes the tools tractable, and it leaves the harder problem unanswered. That harder problem is also an opening. The practitioner who can read the decision space at the depth it actually moves, the operator who carries the contextual discipline the general role does not, sees more of the decision before it moves the firm. That seeing becomes institutional power before it becomes anything the market can price.
Translation fails next, or rather it never happens. The board asks its question in the language it reasons in, and the answer comes back as a number wearing a confident face. The question was posed in the small space and the computation ran in the large one, and nothing carried the board’s question down into the large representation and the answer back up. The inherited tools never had to make that trip, because question and computation once lived in the same small space. This one does, and the system did not make it. What the room receives is opaque dressed as clear, which is the glass box that only looks like one.
Explanation comes next, and this is where Rudin’s distinction does its work and then runs out. The two people in the room were prepared to explain the recommendation. The explanation was a second model laid over the first, assembled to be persuasive after the decision was already made, and a second model can be wrong about the first. So the room faces the choice Rudin named. It can accept a model it could in principle follow, interpretable by construction, or it can accept a story about a model it cannot. It reaches for the story, because the story is ready and the hour is late, and in reaching for it the room accepts a rationalization in place of a reading. This is the failure Rudin warned of, exactly.
And here is where the argument leaves her. Suppose the room had the interpretable model, the one a person can follow by construction. Suppose the recommendation were reachable, every step legible to anyone in the room willing to trace it. The room still could not govern it. A board may be able to retrace the model’s steps and still lack the institutional machinery to govern the decision, because the resolution at which the model decides exceeds the resolution at which the institution can hold accountability, assign a horizon, or specify a refusal. The trace is legible and the decision is still ungoverned. A legible trace does not give the institution the machinery to answer for the decision, and answering for a decision is a different act from following it, one that needs an institution shaped to the grain of the first. That is the move Rudin’s argument makes possible and does not itself make. Interpretability gets the decision into the room. It does not give the room the standing to govern what it can now see.
The standing fails after that. The recommendation rested on premises, the interactions it estimated, the variables it held fixed, the distributions it assumed, and none of them reached the room, because the inherited tools have no slot for them. So the room could not ask whether it believed the premises, the audit function could not check them against the record, and the firm accepted the recommendation on faith, which is not governance. Its blindness went undeclared. The tool did not say which interactions it held and which it treated as independent, which is the one declaration that would have let the firm tell a decision it governed from one it reduced to a shadow. Without that declaration the firm learns the difference only when everything arrives at once, which is how the last sector learned it.
The recommendation moves the firm, because that is what these systems are built to do, and the room turns to find that no one was named to own what it did. The accountability was deferred, the way it always is, because no one wants to assign who owns the autonomous answer before they have to. A name gets attached afterward, to the consequence, once the consequence has a shape and a cost. This is the vacancy the whole sequence has been moving toward, and it deserves to be said without softening. A name attached to a consequence after the fact is not accountability. It is a search for someone to absorb it.
And the record, finally. The decision leaves no trace of what it assumed, how much of the structure it read, what it could not see, and who was answerable for letting it move the firm, so the next operator inherits nothing, the way this generation inherited the financial sector’s failure rather than its working notes. A governance act that leaves no trace is a decision that happened to be supervised once and will not be supervised the same way again, because nothing was written down that the next room could hold. There is no explainability, no audit trail to stress test.
Six consequences, and not one of them announced itself as a missing criterion. Each arrived as a thing the room simply could not do, in the moment it needed to do it. The room did its job with instruments built for a smaller book. It was not a decision made badly. It was a decision made at a resolution the institution could not match, and the failure stayed invisible until the missing part arrived.
VI. What consumer goods knows that finance does not
There is a slide. It presents four categories that grew on their own lines last quarter, and the room reads four businesses, each with its own owner and its own target. Four boxes, four numbers, four people who will answer for them. The slide is honest about everything except the one thing that matters.
The Butterfly Effect’s complexity8 is what the slide cannot draw. The four lines move together. Shared plant capacity creates two entanglements that run in opposite directions. A common shock to the line, a fire, a labor stoppage, a procurement break, pulls every product running on that capacity down together, two lines bound to rise and fall as one, which is positive covariance in exactly the sense Markowitz meant. The allocation tradeoff runs the other way. Pushing one category through in a quarter removes the capacity another was counting on, so the one rises as the other falls, which is negative covariance, written in steel and trucks. The brand portfolio holds both at once. The third entanglement is the one the room least wants to see. When a deep promotion on one brand draws volume forward and steals it from the brand next to it on the same shelf, the covariance is negative again, the lift in one category partly paid for by the one beside it. The manager who lived the quarter knows which line paid. Whether the lines move in concert or in opposition, the count of interaction terms multiplies rather than adds as lines are added, and no single box on the slide can hold a number that lives between two boxes.
A loan book under a capital constraint and a category set under a trade, channel, and finance constraint are the same kind of object, an interwoven whole whose parts move together and whose joint behavior has to be held before any governing can touch it. The bank and the brand portfolio are not alike by analogy. They are instances of one structure, and the covariance Markowitz wrote for securities is the covariance a promotion writes across a shelf. Holding it whole is the hard half the consumer goods tools never built. It is the half the financial picture already supplies.
So the inheritance runs from finance to consumer goods, the shared picture transferring while the governance does not, because the objective, the horizon, the accountability, and the tolerance all run the other way. Now turn it around, because consumer goods knows something finance does not.
The operator who has held a brand through a decade has watched a price cut made for one quarter cost three years of pricing power, and has learned to read the slow variable underneath the fast one. Begin with time. Brand equity compounds or erodes across years and manifests in different ways in different global market regions, and stewarding it is a muscle the quarterly desk was never built to develop. Equity does not announce its erosion. It declines quietly, below what any quarterly report can show, until the decline is large enough to surface as a number, and by then it has been true for years. Multi-year equity stewardship is a discipline finance never had to build, and it is exactly the discipline the autonomous systems most threaten, because they optimize at the speed of the fast variable and the slow one does not show up in time to stop them.
Then the matter of who carries the whole book, which finance solved by building one office and consumer goods never did. The easy reading calls that a deficit, and finance tells the story that way, as the triumph of the chief risk officer seated at the board with the standing to stop a trade or hold a position. Watch how the consumer goods operator lives it and something else appears. Consumer goods distributes accountability across the category manager, the brand president, the supply chain VP, and the CFO. That fragmentation can be a discipline, because no single office can collapse the whole into one number and govern the number instead of the thing. The single risk office is a strength and also a temptation, a seat from which the whole can be flattened on a bad day. The consumer goods operator governs a structure no one office can flatten, and the distribution that reads as a gap is also a guard against the easiest mistake the single office makes. Consumer goods has a protection finance lacks. The seat that was supposed to be the answer carries its own characteristic failure, the flattening of a high-dimensional book into the one number the seat can hold, and the sector without the seat is protected from that failure by the same fragmentation it is told to regret.
The strain shows up in the public record, read as record rather than verdict. On February 21, 2019, Kraft Heinz recorded a non-cash impairment of more than fifteen billion dollars against the carrying value of its Kraft and Oscar Mayer brands.9 The equity in those brands had been eroding against private-label pressure and channel shift for the better part of two years, while the category reports the governing bodies read carried that equity as sound. The number held the value right up until the morning it reported, all at once, that the value had not been there. The slow variable moved the whole time and surfaced only when it was large enough to force an audited figure.
Six years later the same firm yields the structural version of the lesson. On September 2, 2025, Kraft Heinz announced a plan to separate into two independent public companies, saying the complexity of its current structure made it challenging to allocate capital effectively, prioritize initiatives, and drive scale.10 That is a company’s own account of a book of business whose breadth had grown past what its allocation could hold, and of a remedy that made the portfolio smaller rather than building the apparatus to govern it whole. Inherited tools often make subtraction the easiest move. Sometimes it is right, and often it is simply the move available when the firm cannot govern the whole.
Finance is the wrong benchmark for the consumer-goods machinery. That machinery is governed by different muscles, of multi-year stewardship, of accountability held without a single office to collapse it, of a calibrated relationship to recoverable error. Carried into the governance of structures that will run faster and reach further than any loan book, that knowledge brings something the financial inheritance never had to learn and does not contain. The one that paid in the open is not the one that knows the most.
VII. The governance that has not arrived
The systems are already in the building. The governance that would close this gap has not arrived. Boards, investors, and senior teams still lack the tools this structure requires. For now, the gap is carried day after day by the people inside the institutions who can see both sides of it at once.
The arrangement creates power. Someone in the room is asked to vouch for what the governing bodies cannot read. The person who can read the machine’s decision has power the board has ceded without deciding to cede it. The question of who governs the machine is usually asked as a question about values, about what the machine should be allowed to want. The plainer question is who can see the decision as it was made, and what the firm owes those who cannot. Varoufakis and Zuboff have been theorizing where power goes when the machine becomes the medium through which decisions pass.11 The resolution gap names the precondition their argument assumes. Power migrates to whoever can read the decision the machine made, and in most firms that reader is a person, a vendor, or no one at all, and the firm has not decided which.
The financial sector learned its version of this in the open and after the fact, when the entanglement it could not see arrived all at once and the firms paid for the structure their tools had not held. That is the one record we have, and it is contested, because the rebuild may have reproduced at a higher level the very confidence it was built to cure. The rest of the corporate world has not yet had its reckoning. It holds the same kind of book under the same condition, with more time, and most of it is spending that time making the book smaller, because subtraction is the move the inherited tools permit. The other move is harder and is still available. The firm needs tools that read the whole book at the depth it actually moves and report clearly where their seeing stops, so leaders can tell a decision they governed from one they merely simplified. The last generation lacked that declaration. This generation can still build it before a crisis forces the work.
The interval is open. The firms already at work on the instruments are building the tools and the tradecraft that read the entangled book at the resolution it actually moves, and they are building them before the failure rather than after it. The contextual reading they are learning to demand becomes institutional power before it becomes anything the market can see, and in time it becomes an asymmetric advantage of a kind the inherited apparatus was never built to grant or to deny. The instruments will be built by practitioners, ahead of the failure, or they will be built after a crisis that did not have to happen. The entangled variables living in a company’s book of business do not care which sector holds them, only whether someone can read them whole.
Stephen DeAngelis
Princeton, NJ
June 2026
About the Author
Stephen F. DeAngelis is the founder, president, and CEO of Enterra Solutions and Massive Dynamics, two companies that apply artificial intelligence and advanced mathematics to complex enterprise challenges. His work spans international relations, national security, and commercial technology, with visiting research affiliations at Princeton University, Department of Chemistry, the Computing and Computational Sciences and National Security Directorates of the Oak Ridge National Laboratory, the Software Engineering Institute at Carnegie Mellon University, and the MIT Computer Science and Artificial Intelligence Laboratory. He holds patents in autonomous decision science.
Notes:
1. Nils Gilman, “Governing in the Planetary Age,” Noema Magazine, March 9, 2021. Gilman argues that the scale of planetary challenges is incommensurate with the governing capacity of the nation-state, which is “too small for the big problems of life and too big for the small problems,” a structural rather than volitional failure of fit. https://www.noemamag.com/governing-in-the-planetary-age
2. Cynthia Rudin, “Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead,” Nature Machine Intelligence 1 (2019), 206 to 215. Rudin argues that post-hoc explanation of an opaque model is a second model that can be wrong about the first, and that high-stakes decisions should use models interpretable by construction. https://www.nature.com/articles/s42256-019-0048-x
3. Harry Markowitz, “Portfolio Selection,” The Journal of Finance 7, no. 1 (March 1952), 77 to 91. Markowitz showed that the risk of a portfolio depends on the covariance among its holdings rather than on the holdings read independently, establishing the entangled whole as the proper object of allocation. https://onlinelibrary.wiley.com/doi/10.1111/j.1540-6261.1952.tb01525.x
4. Felix Salmon, “Recipe for Disaster: The Formula That Killed Wall Street,” Wired, February 23, 2009, a retrospective on David X. Li’s 2000 Gaussian copula and its role in pricing the joint default behavior of collateralized debt obligations off a single correlation parameter. https://www.wired.com/2009/02/wp-quant/
5. Chronologically, on stress testing. The Comprehensive Capital Analysis and Review descends from the 2009 Supervisory Capital Assessment Program, the emergency exercise run across the nineteen largest banking firms at the depth of the crisis. The Federal Reserve conducted the first annual CCAR in early 2011 across those same nineteen firms. The November 2011 capital plan rule extended the regime to all U.S. bank holding companies with fifty billion dollars or more in total consolidated assets, so that by the 2014 cycle roughly thirty firms were inside it. The test required each institution to project its book through a severe, supervisor-specified downturn and show post-stress capital above the regulatory minima, with dividends and buybacks barred from drawing capital below the minimum plus the modeled loss. Liquidity followed: the Liquidity Coverage Ratio final rule of September 3, 2014 required covered firms to hold liquid assets against a thirty-day stressed outflow, phased to full compliance by 2017. By the later years the regime had, on Tarullo’s own retrospective account, been diluted into a routinized, predictable process, which is the seam the critics press. Their standing charge: a common scenario run across every large firm produces model monoculture, every balance sheet blind in the same place at once, while point-in-time stress capital is procyclical, tightening into a contraction it was meant to cushion. On the measure itself, the mechanism is the load-bearing point. Basel d457, the Fundamental Review of the Trading Book, finalized in January 2019, replaced Value-at-Risk with Expected Shortfall as the internal-models market-risk standard. VaR names a loss threshold at a stated confidence, conventionally the ninety-ninth percentile over a ten-day horizon, and reports nothing about the distribution beyond the line it draws. Expected Shortfall, calibrated under d457 to a 97.5 percent level over a stressed window, averages the losses across the tail past that point rather than marking where the tail begins. The earlier instrument answered where the loss line falls, the replacement answers how large the loss is once the firm is past it, a genuine rise in resolution into precisely the region the prior measure left unwritten. The literature designates the supervisory turn macroprudential, and it has an architect. Daniel Tarullo held a seat on the Federal Reserve Board of Governors from January 28, 2009 to April 5, 2017, with the supervisory portfolio, and his pre-crisis Banking on Basel (2008) had argued that the inherited framework treated the largest trading books as no special case when their correlated tail losses were exactly what it failed to price. From that chair the central move was institutionalized: a shift from microprudential supervision, asking whether each firm is sound one at a time, to macroprudential supervision, asking whether the system as a whole can absorb a common shock, with stress-testing the standing instrument. What the pre-2008 regime did not assert is that the soundness of each firm read in isolation does not aggregate to the soundness of the whole, because the correlations invisible firm-by-firm are the ones that move the system. Compare Stein. On February 7, 2013, in “Overheating in Credit Markets: Origins, Measurement, and Policy Responses,” delivered at a Federal Reserve Bank of St. Louis symposium, then-Governor Jeremy Stein argued that monetary policy might need to weigh financial stability even when inflation was contained, because risk reaches for yield and migrates into instruments the standard measures are not built to see, and because a sufficiently broad policy rate “gets in all of the cracks” supervision alone cannot reach. The relevance is direct: a sitting governor arguing structurally about the limits of the instruments the institution then held, naming a blindness to migration alongside the copula’s blindness to entanglement. See, against all of this, Haldane. In “The Dog and the Frisbee,” delivered with Vasileios Madouros at the August 2012 Jackson Hole symposium, the argument is that a dog catches a frisbee without solving the differential equation governing its flight, and that complex environments are often governed better by a few simple heuristics than by models that add representational detail faster than reliable signal. The empirical claim is sharper than the metaphor: a simple leverage ratio predicted crisis-era bank failures where risk-weighted capital measures did not, significant at the one percent level while the risk-based measures were insignificant; and the apparatus had inflated from the thirty pages of Basel I to the six hundred and sixteen of Basel III without commensurate gain in reliability. This cuts against the body’s instinct that higher resolution helps. The answer is a narrowing rather than a dispute: resolution helps only when paired with a declared edge. An instrument that adds detail and says nothing about the limit of its own seeing is the one he warns of; one that declares which interactions it held and which it treated as independent passes his test, giving the operator a view of its own blindness. The criterion is blindness made visible, not resolution as such. The macroprudential critics press from the same edge and deserve naming. David Aikman of the Bank of England has argued that supervisory stress severity should itself be procyclical and rule-governed rather than discretionary, on the diagnosis that the regime as run amplifies the cycle it was meant to dampen. Robert Engle of NYU Stern, with Viral Acharya, has shown that risk-weighted stress measures rank bank capital needs incorrectly and that market-based systemic measures such as SRISK diverge sharply from official results. Barry Eichengreen of Berkeley has argued that the pre-crisis regime was capital-poor and procyclical because regulators accepted the banks’ own internal value-at-risk models, the intellectual capture the rebuild risks reproducing. Their convergent charge: stress testing produced model monoculture, procyclicality, and the institutionalization of the false confidence the body warns against. The response treats the critique as its sharpest test, not its refutation, because this is precisely why resolution without a declared edge reproduces the pathology one level up, with newer instruments and the same blindness. Basel Committee on Banking Supervision, “Minimum capital requirements for market risk” (Fundamental Review of the Trading Book), d457, finalized January 2019. https://www.bis.org/bcbs/publ/d457.htm
6. Andrew G. Haldane and Vasileios Madouros, “The Dog and the Frisbee,” Jackson Hole economic policy symposium, August 2012, Bank for International Settlements. Haldane argues that complex environments are often governed better by a few simple rules than by models that add detail faster than they add reliable signal. https://www.bis.org/review/r120905a.pdf
7. Adam Tooze, Crashed: How a Decade of Financial Crises Changed the World. New York, Viking, 2018. Tooze’s account treats the 2008 crisis and its aftermath as politically unresolved and ongoing rather than a closed chapter, the standing check on any reading of the post-crisis build as a debt paid in full. https://www.penguinrandomhouse.com/books/301357/crashed-by-adam-tooze/
8. Edward N. Lorenz, “Deterministic Nonperiodic Flow,” Journal of the Atmospheric Sciences 20, no. 2 (March 1963), 130 to 141, and the 1972 address to the American Association for the Advancement of Science titled “Predictability: Does the Flap of a Butterfly’s Wings in Brazil Set Off a Tornado in Texas?” The figure of speech entered popular usage from these two papers, naming the phenomenon by which small differences in initial conditions compound through nonlinear dynamics into large divergences over time. https://journals.ametsoc.org/view/journals/atsc/20/2/1520-0469_1963_020_0130_dnf_2_0_co_2.xml
9. Kraft Heinz Company, “Kraft Heinz Reports Fourth Quarter and Full Year 2018 Results,” February 21, 2019 (Form 8-K, Exhibit 99.1). The company recorded a non-cash impairment of more than fifteen billion dollars against the carrying value of its Kraft and Oscar Mayer brands. https://www.sec.gov/Archives/edgar/data/1637459/000163745919000010/ex991-erq42018.htm
10. Kraft Heinz Company, “The Kraft Heinz Company Announces Plan to Separate into Two Scaled, Focused Companies,” September 2, 2025. The company stated that the complexity of its current structure made it challenging to allocate capital effectively, prioritize initiatives, and drive scale in its most promising areas. https://news.kraftheinzcompany.com/press-releases-details/2025/The-Kraft-Heinz-Company-Announces-Plan-to-Separate-into-Two-Scaled-Focused-Companies-to-Accelerate-Profitable-Growth-and-Unlock-Shareholder-Value/default.aspx
11. Yanis Varoufakis, Technofeudalism: What Killed Capitalism. London, The Bodley Head, 2023, and Shoshana Zuboff, The Age of Surveillance Capitalism. New York, PublicAffairs, 2019. Both theorize the migration of power that follows when digital and algorithmic systems become the medium through which economic decisions pass, the power question for which the resolution gap names a precondition. https://www.penguin.co.uk/books/451795/technofeudalism-by-varoufakis-yanis/9781529926095
Share this post
