It’s therefore pretty amazing that in the data world, those roles collapse routinely.

The same people who specify what’s to be measured are usually the ones who collect it, distribute it, and control what it means. This isn’t negligence. It’s structural. Data represents the basis of the narrative that people want to permeate the organization. Of course the people who own that narrative own the measurement system that validates it.

This came up this week in the comments on a sharp piece by my colleague Bill Schmarzo arguing that AI doesn’t have a trust problem — your metrics do. His core claim: AI optimizes exactly what you tell it to optimize, so if your metrics track activity rather than value, you’ve handed an extraordinarily capable engine a bad set of instructions. He’s right, and the piece is worth reading.

But the thread surfaced something the piece doesn’t fully address. Two problems, actually — related but still distinct. Getting them confused leads to solutions that fix one and leave the other intact.

The conflict of interest is structural, not accidental.

The food analogy explains why the metric problem is so persistent. It’s not that data leaders are unaware that metrics should connect to outcomes. Most of them know this. It’s that the people responsible for that connection have an inherent stake in what the connection looks like. The data represents the evidentiary basis for the story leadership wants told. Asking that system to objectively audit itself is like asking the food producer to determine whether the distribution criteria are fair. The role structure doesn’t allow for the objectivity the task requires.

Bill’s prescription — audit your KPIs before you automate them, trace each metric to the business outcome it’s supposed to represent — is the right prescription. But it implicitly assumes the audit can be conducted with genuine independence. In most organizations, it can’t, for exactly this reason.

The fix this problem implies is organizational: separate the roles. Create genuine independence between those who specify metrics, those who collect them, and those who interpret them. That’s necessary.

But it’s not sufficient.

Even independent auditors may be using the wrong method.

Here’s where it gets harder. The audit Bill calls for is typically conducted by data scientists — the profession theoretically positioned to serve as the check-and-balance on organizational narrative. And that profession, by training and by tool, approaches reality from a correlative perspective.

This isn’t a criticism of data scientists. Correlation is powerful. It’s the right tool for many problems. But it is not the right tool for establishing whether a metric actually causes the outcome you care about, or merely travels with it under current conditions.

Goodhart’s Law — the principle Bill correctly invokes — is precisely a causation problem wearing correlation’s clothes. A metric correlates with business value in a stable environment. You optimize against it. The optimization severs the causal relationship the correlation was tracking. The metric keeps climbing. The value stops following. You can’t diagnose that risk using correlative methods, because the correlation looks fine right up until it doesn’t.

Which means: you can solve the conflict of interest problem completely — perfect role separation, zero stake in the outcome, fully independent audit team — and still arrive at metrics that won’t survive contact with an AI optimization engine. The auditors are independent. But they’re using the wrong map.

This is a reasoning architecture problem.

The gap between what organizations measure and what they value is real and persistent. But the reason it persists isn’t primarily that data teams lack good data, or that the roles aren’t separated, or even that leadership doesn’t care. It’s that the reasoning method most commonly used to evaluate metrics can’t see the gap from where it’s standing.

What’s needed isn’t better correlation or cleaner governance structures alone. It’s a fundamentally different interrogation posture — one that asks not “does this metric move with the outcome?” but “would this metric still point toward the outcome if we optimized hard against it?” That’s a causal stress-testing question. And it requires a reference standard that doesn’t share the reasoning architecture of the system being interrogated.

This is the class of question that documented human failure across centuries and disciplines — military strategy, medicine, law, engineering, organizational behavior — has consistently shown we are structurally bad at asking ourselves. Not because we’re not smart. Because the method most of us default to wasn’t built to interrogate its own load-bearing assumptions. The auditors and the system they’re auditing learned to think the same way. That’s the problem.

The implication for AI governance:

Bill frames the ultimate problem as a leadership gap hiding inside your KPIs. I’d propose to go two levels further.

First, it’s a structural conflict of interest hiding inside your org design.

Second, and harder, it’s an epistemic (how we know what we know) gap hiding inside your methodology — the map problem. Leadership can’t close either one by just caring more. The first closes with genuine role separation. The second closes only when the interrogation method is structurally independent of the assumptions being interrogated — when the map and the mapmaker don’t share the same blind spots.

This is the core reason why audit firms can be so very necessary.

That’s the work I’ve been focused on with #ProvingGround — building an instrument that asks the questions authors of a plan or decision are structurally least likely to ask themselves, against a bulletproof reference standard with no stake in their reasoning.

Bill’s piece is pointing at the right problem. The solution space is wider than better KPI governance and deeper than better org design. It starts with a question most audit processes never ask: does the tool we’re using to evaluate our thinking have the same blind spots as the thinking we’re evaluating?

If it does, we haven’t solved the problem.

We’ve just made everything look more official.

This belief is running, simultaneously, through two of the most consequential resource decisions in modern business: the race to scale AI compute, and the push to scale GTM investment. And in both cases, the belief rests on a formal error — one with serious consequences for organizations that haven’t yet named it.

The Scaling Premise and Its Hidden Assumption

The scaling premise, stated plainly, is that input and output have a stable, predictable, roughly linear relationship. More data plus more compute produces better AI. More sales reps plus more marketing spend produces more revenue. The logic seems self-evident.

But embedded inside that logic is an assumption that rarely gets examined: that the system you are scaling behaves like a closed system — one where the relationship between inputs and outputs doesn’t change in ways you can’t control.¹

Closed systems are tractable. You can map their parameters, hold them stable across experiments, and extrapolate from historical patterns with reasonable confidence. The problem is that GTM environments and AI’s operating reality are both open systems — systems where new variables, new actors, new causal pathways, and new environmental conditions continuously enter from outside the model. The relationships between inputs and outputs don’t stay fixed. They shift in ways that past data cannot fully anticipate.²

Scaling a closed-system strategy in an open-system environment doesn’t just fail to solve the problem. It actively makes things worse, because it generates more evidence confirming a model that is increasingly misaligned with how reality actually works.

What the AI Scaling Debate Reveals

The most rigorous version of this critique has emerged, somewhat awkwardly, from inside the AI industry itself. After years of confident assertions that scaling data and compute would produce progressively better AI — eventually leading to artificial general intelligence — cracks are widening in the consensus.³

The fundamental issue isn’t computational. It’s epistemic, meaning “how we know what we know.” Large language models and their predecessors are trained on large amounts of historical data. They learn the patterns that existed in that data. But as researchers have documented, when the environment changes structurally — when what economists call a “regime shift” occurs — the historical patterns become not just less useful but actively misleading.⁴

A 2025 ACL paper demonstrated the mechanism directly: obsolete information significantly degraded model accuracy even when correct current information was also available in the retrieval context.⁵ The model’s statistical confidence in older patterns overwhelmed the signal from newer ones. This is not a bug that more compute will fix. It is a structural property of systems trained on accumulated history and deployed into a continuously changing present.

The deeper point is what might be called the inertia of human knowledge — this means that the larger the historical corpus, the more statistical mass backs the old regime. Immediately after a structural break, you face the worst possible combination: maximum historical evidence, minimum current-regime evidence, and maximum divergence between the two. Scaling data collection in that environment makes the problem worse before it makes it better, if it ever does.⁶

This has a formal analog. Any system attempting to model an open environment by treating it as closed will eventually encounter what might be called a horizon problem: the open system keeps generating genuinely novel variables and causal structures that have no prior representation in the training data. You cannot collect observations from a reality that hasn’t happened yet. Scale cannot purchase future adequacy from present data.⁷

The Same Architecture in GTM

The GTM scaling bet is built on identical intellectual foundations — and in many cases, by the same people.

The doctrine of blitzscaling, formalized by Reid Hoffman and widely propagated through Y Combinator and Silicon Valley’s venture networks, argued explicitly that speed of scale justifies significant inefficiency. As the assertion goes, we should accept massive operational waste in exchange for the competitive advantage of moving faster than the market.⁸ This philosophy was generative and occasionally correct — when network effects are real, when winner-take-all dynamics apply, when the window is genuinely narrow.

The problem is what happened next. The logic got generalized far beyond the conditions that made it defensible. “Spend your way to market dominance” became a template applied across categories regardless of whether their competitive dynamics actually resembled the cases where blitzscaling worked. The doctrine that produced some exceptional outcomes in consumer social networks was applied to B2B SaaS, to enterprise sales, to markets with long buying cycles and complex stakeholder structures — environments where the causal relationship between spend and revenue is anything but linear or predictable.

The GTM function’s version of this epistemic inertia is the MQL and the Funnel. Most organizations have accumulated years of pipeline data demonstrating apparent correlations between marketing activities and revenue outcomes. That data feels like evidence. It has the weight of organizational history behind it. But in volatile markets — where ICP definitions have shifted, where buyer journeys have fragmented, where economic conditions have reset willingness-to-pay — those historical correlations are presently describing a world that no longer exists.⁹

The GTM function cannot easily see this, because the failure mode of scaling a broken motion is diffuse and lagged. Revenue underperformance gets attributed to macro conditions, to comp plan design, to product gaps, to competitive pressure — to anything that distributes accountability away from the motion itself. The causal signal is obscured precisely when you most need to see it.

The Accountability Asymmetry

There is a structural reason why the scaling bet persists even as its evidence base erodes. Changing the motion is visible, immediate, and owned. Whoever calls for a strategic pivot owns the outcome of that pivot. Scaling the existing motion, by contrast, diffuses accountability — underperformance in a scaled-up version of the current approach is harder to trace to specific decisions.

This creates a systematic bias toward incumbency of thought, if not necessarily the thought leaders. Leaders who have built careers on particular GTM architectures — large field sales organizations, SDR-led outbound, demand-generation funnels, category-creation spend — face an existential threat in a diagnosis that says the architecture itself is wrong. The resistance to structural change is not simply organizational inertia. For many people, it is identity.¹⁰

The AI industry faces the same dynamic at institutional scale. Acknowledging that the scaling paradigm has fundamental limits would require writing down enormous capital commitments, dismantling narratives that have driven valuations, and admitting that the path to capable AI runs through architectural changes rather than raw compute. The financial and reputational costs of that admission are so large that the industry has strong structural incentives to keep betting on scale even as the evidence accumulates.¹¹

The Causal Question That Changes Everything

Both of these situations resolve when you shift the question.

The scaling question is: are we doing enough of this? The causal question is: is this the right investment or action, given how the system actually works today?

In AI, that shift requires moving from pattern-matching architectures toward what researchers working on causal AI frameworks call genuine causal modeling. These are systems that reason about interventions and their effects rather than correlations in historical data.¹² Judea Pearl’s causal hierarchy formalizes the distinction: association (what co-varies?), intervention (what happens if I change this?), and counterfactual reasoning (what would have happened differently?). Most current AI, and most current GTM analytics, operates almost exclusively at the first level.

In GTM, the causal question forces an examination of the actual mechanism by which buyers make decisions — not what correlated with closed revenue in the historical dataset, but what causally drove it, and whether those drivers still hold in the current environment. This is a fundamentally different kind of analysis than waterfall funnel reporting, and it often produces fundamentally different conclusions about where to invest.

The organizations that will navigate this transition successfully are not necessarily those with the largest budgets. They are those willing to ask whether their confidence in current methods reflects the actual causal structure of their markets — or the accumulated statistical weight of a regime that may already be behind them.

That distinction will determine a great deal about who is still standing in five years.

Footnotes

¹ The closed-system assumption in economics and strategic planning has a long lineage, but its application to scaling strategy was made most explicit in the blitzscaling literature. See Reid Hoffman and Chris Yeh, Blitzscaling: The Lightning-Fast Path to Building Massively Valuable Companies (Currency, 2018), particularly the discussion of prioritizing speed over efficiency in winner-take-all markets.

² On open vs. closed systems in organizational and economic contexts, see W. Ross Ashby’s foundational work on regulatory systems and, more recently, complexity economics literature from the Santa Fe Institute, including W. Brian Arthur, The Nature of Technology (Free Press, 2009).

³ The public fracture in the scaling consensus began surfacing notably in 2024–2025, with reports from multiple AI labs indicating that benchmark gains from pure compute scaling were flattening, and that new architectural approaches would be required for continued progress. See reporting in MIT Technology Review, The Information, and academic preprints on scaling law limitations circa 2024–2025.

⁴ On the concept of regime shifts and distributional shift in machine learning, see the literature on covariate shift and dataset drift, including Joaquin Quiñonero-Candela et al., Dataset Shift in Machine Learning (MIT Press, 2009), and subsequent work on temporal distribution shift in deployed systems.

⁵ Findings on the degradation of retrieval-augmented generation by obsolete information: research presented at ACL 2025 demonstrating that stale information in retrieval corpora could mislead models even when current correct information was simultaneously available. The benchmark created specifically to test outdated information handling in RAG systems showed significant accuracy degradation attributable to temporal staleness.

⁶ The concept of epistemic inertia as applied to AI systems extends the standard machine learning notion of catastrophic forgetting in a different direction: not that models forget old information, but that old information retains disproportionate statistical weight relative to new information in structurally changed environments.

⁷ The formal argument about open systems and the impossibility of complete closure draws on Gödel’s incompleteness theorems as applied to formal modeling systems — any sufficiently complex formal system contains truths it cannot prove from within its own axiomatic boundary. Applied to data systems, closure itself is the binding constraint, not computational scale.

⁸ Hoffman and Yeh, Blitzscaling, op. cit. The explicit framing of accepting inefficiency for speed is the book’s central strategic argument, distinguishing blitzscaling from conventional “smart scaling” approaches.

⁹ On MQL waterfall metrics and the measurement of marketing contribution, see work from the Marketing Accountability Standards Board (MASB) on marketing mix modeling and attribution, and critiques of last-touch and linear attribution models in volatile buying environments.

¹⁰ The organizational behavior literature on identity threat and resistance to change is extensive. For a strategic management lens, see Clayton Christensen’s work on the innovator’s dilemma, which documents how incumbent expertise and identity create systematic blindness to architectural disruption — The Innovator’s Dilemma (Harvard Business School Press, 1997).

¹¹ On the financial incentive structures reinforcing AI scaling commitments, including circular investment relationships between infrastructure vendors and AI labs, see analysis of Nvidia’s venture investment portfolio and the structural dependencies created by hardware-contingent funding arrangements, widely reported in technology financial press through 2024–2025.

¹² On causal AI and the distinction between correlational and causal reasoning systems, see Judea Pearl and Dana Mackenzie, The Book of Why: The New Science of Cause and Effect (Basic Books, 2018), and Pearl’s earlier technical treatment in Causality: Models, Reasoning, and Inference (Cambridge University Press, 2000, 2nd ed. 2009).

There is an extraordinary idea buried inside that premise. For more than a century, statistics has warned us that correlation does not imply causation. But a significant part of the AI scaling thesis effectively proposes an operational workaround: perhaps we do not need to understand causality explicitly if we can observe enough correlations, across enough variables, over enough time, at sufficient resolution. A sufficiently complete statistical representation of Reality might behave as though it understood causality.

There is ample evidence that this can work remarkably well — right up until Reality changes. And Reality is changing, not merely faster, but more discontinuously. That distinction may prove to be one of the most important unresolved problems in AI.

Data does not have to become false to become wrong

We normally think about bad data in familiar ways: inaccurate, incomplete, corrupted, duplicated, biased, poorly labeled, missing provenance. But there is another category that is much harder to see. The data can remain perfectly accurate historically while becoming progressively less representative of present Reality.

Customers really did behave that way in 2022.

That sales strategy really did work in 2019.

Customers really did see risk differently in 2005.

None of the records have become false. What changed is the world to which we are trying to apply them.

Researchers studying human capital increasingly describe knowledge itself as possessing a half-life. Knudsen and Lien, writing in Human Resource Management Review, argue explicitly that knowledge stocks depreciate and that their rate of depreciation varies by domain and across time. That is intuitively obvious once you think about it.

Here’s another example: Newtonian mechanics has an extraordinarily long useful life, while knowledge of a particular software architecture may not. The geology beneath a bridge changes very slowly, while customer buying behavior can change dramatically in 12 months. The human body retains many fundamental characteristics across centuries, while our understanding of how to treat a particular disease can leave physicians wondering what to do. The useful age of knowledge is therefore not primarily chronological — it depends on how far Reality has traveled since the knowledge was created. That same principle applies to data.

The problem is not forgetting. It is perpetuating.

We often imagine knowledge depreciation as forgetting, but in modern organizations forgetting may be the lesser problem. Digital systems are extraordinarily good at remembering. We preserve presentations, transactions, emails, customer records, operating metrics, research, forecasts, plans, policies, analyses and decisions indefinitely. Storage became cheap enough that organizations largely stopped asking whether information should remain active knowledge. We built data lakes, and then we built AI systems capable of swimming through them.

The resulting problem is almost the opposite of forgetting: we preserve knowledge after the conditions that made it valid have disappeared. An organization discovers something that works. It becomes a best practice, then a process, then software, then metrics, then management doctrine. Consultants teach it. Universities codify it. Executives build careers around it. Eventually AI trains on an enormous historical record documenting that the idea works — and then Reality changes, the knowledge survives, and what we are left with is not memory loss but persistence beyond validity.

External changes drag data into obsolescence

The mechanism is fairly straightforward. Organizations operate inside open systems where customers change, competitors react, technology advances, interest rates move, governments regulate, supply chains reorganize, demographics shift, and new behaviors continuously emerge. Those exogenous changes alter the system that generated the data in the first place, producing a sequence: exogenous change → regime divergence → data depreciation → decision distortion.

The crucial issue is that this relationship is not necessarily smooth, because there are really two distinct dimensions of change. Velocity describes how rapidly the environment is moving; volatility describes how unevenly and discontinuously it is moving. Smooth change allows data to depreciate gradually, but volatile change creates structural breaks — and a structural break can make information that was representative yesterday materially nonrepresentative tomorrow. This suggests that data behaves economically more like other assets than we have wanted to admit: some of it experiences gradual depreciation, some of it experiences sudden impairment, and we generally do not account for either.

Three industries show the continuum

Consider three very different industries, arranged by how quickly the Reality underlying their data tends to change.

Structural engineering sits at one end. The core relationships are anchored in unusually persistent Reality — gravity has not changed, steel does not wake up one morning with different political preferences, and concrete compression is not disrupted by a startup. Much of the historical knowledge surrounding fundamental physics and material behavior retains enormous usefulness across decades. But move outward from the physics and depreciation appears: construction techniques change, building codes change, labor costs change, climate assumptions change. Even here, data does not have a single expiration rate; its half-life depends upon what part of Reality it represents.

Property and casualty insurance sits toward the middle. Property insurance depends enormously on historical observation — claims frequency, severity, weather, geography, property values, construction characteristics, replacement cost — but the surrounding system is continuously moving. Climate patterns change, people migrate, development expands into previously less-populated risk zones, replacement costs rise, litigation changes, reinsurance markets shift. A wildfire claim from 1995 remains a true record of what happened in 1995, but it does not follow that it deserves equal evidentiary weight in estimating a wildfire loss 30+ years later. The historical event remains true; its causal portability has changed.

Enterprise software and AI sits at the opposite end — an almost pure laboratory for rapid knowledge depreciation. Products change, platforms change, distribution changes, capital economics change, customer expectations change, procurement changes, and AI itself is now changing many of these variables simultaneously. A company can possess fifteen years of immaculate customer data and still discover that much of it describes several extinct civilizations of customer behavior. That creates a deeply counterintuitive result: historical abundance can become a biasing prior. Ten years of old-regime data may statistically overwhelm one year of new-regime data even when that single year represents the Reality that now matters. The past has more votes; the present has more relevance; and our systems frequently confuse the two.

The problem here is that human perspectives inside the insurance and software industries are not keeping up with the external changes or the rate of those changes.

There is a point where the data lake stops representing Reality

A data team should therefore be asking a question that is still surprisingly uncommon: at what point has Reality moved far enough that historical and current observations should no longer be treated as samples from the same system? Call this the representativeness boundary. Below it, pooling historical and recent observations may be perfectly reasonable. Near it, older evidence should be discounted and important relationships revalidated. Beyond it, the data should be partitioned by regime.

Historical information should not be deleted — history remains extremely valuable. But it should stop automatically qualifying as current evidence. The right question is no longer how old is this dataset? It is has the regime that made these observations informative survived? That is a profoundly different approach to data governance.

Why correlation makes this especially dangerous

Most business decision systems are fundamentally correlative. They discover that when X happens, Y tends to follow, and given sufficient stability, that can work extraordinarily well. But consider what happens when the underlying system changes.

A dataset can contain millions of observations supporting a relationship with extraordinary statistical significance. The model becomes more precise, confidence intervals narrow, sampling error declines — and yet the relationship itself may no longer describe current Reality. So we can have precision rising while validity falls, which is one of the most dangerous conditions in analytics because the system becomes more confident at exactly the same time it is becoming less relevant. Larger samples reduce uncertainty inside an assumed distribution, but they do not prove that the distribution remains the right one.

This is why the familiar data-science instinct that more data is always better becomes dangerous in nonstationary environments. You absolutely can have too much data — specifically, when observations from obsolete regimes dominate observations from the regime you are actually trying to understand.

This is a really big challenge to the “scale data, scale compute” premise

Modern AI pushes this problem much further. The scaling premise says, roughly, that more data plus more compute plus larger models produces better representations of Reality. That works exceptionally well when Reality remains sufficiently stable. But in a volatile open system, while the AI accumulates information about current Reality, it simultaneously retains enormous quantities of information about previous Realities. The gross information stock keeps increasing, but its relevance density does not necessarily increase with it. The model can become vastly better at understanding the historical record without becoming proportionately better at distinguishing which parts of that record still govern the present. At that point the problem is no longer computational scale — it is epistemic inertia, meaning that we lose awareness of how we know what we know.

The larger the historical regime in the corpus, the greater the statistical mass behind it. When Reality undergoes a structural break, the new regime initially possesses very little data precisely because it is new, which means that immediately after major change, the system faces the worst possible combination: maximum historical evidence, minimum current-regime evidence, and the historical evidence is least portable at exactly that moment. Scale cannot solve this simply by scaling harder, because a closed historical representation cannot fully envelop an open system that continues creating variables, interactions and causal pathways that have never existed before. More memory cannot manufacture observations from a Reality that has not yet happened.

Which brings us to hallucination

We generally describe hallucination as AI making things up — a definition that is technically useful but too narrow for the user, who experiences it much more simply as the AI confidently telling them something that was not true. Under that broader definition, data obsolescence opens several additional failure modes that go unrecognized, including their compounding effects. A model can fabricate something outright. But it can also give an answer that used to be true, apply a relationship from an old regime to a new one, or combine individually valid facts from incompatible periods into a conclusion that is false now.

Research already demonstrates part of the mechanism. A 2025 ACL paper created a benchmark specifically to test outdated information in retrieval-augmented generation and found that obsolete information significantly reduced answer accuracy — and could mislead models even when the correct current information was also available. The system does not merely suffer when current information is missing; it can suffer because old truth competes with new truth, and the old truth may possess far greater representational mass.

This creates what we might call temporal majority bias: the dominant historical representation overwhelms the smaller amount of evidence describing current Reality, operating like a standing vote where the past holds more seats simply because it has been around longer. A new-regime observation has to overcome not just uncertainty but sheer statistical mass. AI may therefore not merely hallucinate facts — it may regress toward extinct realities. NIST’s latest work on deployed AI systems makes the necessity of continuing real-world monitoring explicit: controlled pre-deployment evaluation is not enough when systems operate under dynamic real-world conditions, because without some ongoing mechanism for determining whether the Reality represented by their evidence still exists, AI systems cannot reliably know when their own epistemic foundation has been impaired.

Go-to-Market execution makes this phenomenon very visible

B2B go-to-market offers a useful illustration of how temporal majority bias operates in practice, though the verdict on exactly where the regime boundary sits remains genuinely contested and reasonable practitioners disagree. The accumulated GTM corpus contains enormous amounts of information supporting a familiar worldview: funnels, lead generation, seller-driven progression, outbound prospecting, attribution, pipeline conversion, CRM activity as a proxy for buyer movement. These ideas are not imaginary — they describe practices that produced results in particular historical environments.

Meanwhile, a growing body of evidence suggests buyer behavior has changed substantially. Buyers conduct far more research independently, buying groups have become larger and more internally complex, supplier interaction frequently occurs later in the process, and unwanted outreach frequently creates resistance rather than acceleration. Whether that constitutes a full regime change or a significant evolution within the same regime is a live argument. What is not in dispute is the structural phenomenon: the old worldview possesses enormous representational mass regardless of which interpretation is correct. Books describe it, software operationalizes it, consultancies sell it, organizations measure it, executives defend it, and now AI generates even more content repeating it.

This produces a particularly interesting dynamic: obsolete epistemology acquires new timestamps. A blog post published yesterday may contain a 2015 model of Reality, which means recency alone cannot solve the problem. Ask a generic GenAI system how to improve B2B growth and it will likely reproduce the dominant historical consensus. Push it hard enough on contemporary buying evidence and it may articulate an entirely different picture. The knowledge is often inside the model; the problem is determining which Reality should govern the answer. A sophisticated user can force that distinction. Most users will not know they need to.

Data science itself is not immune

There is an uncomfortable irony here. Data science has developed sophisticated techniques for dealing with nonstationarity, concept drift, change-point detection, regime switching and temporal validation — the discipline knows these problems exist. Yet practical data culture still frequently rewards larger datasets, longer histories, greater sample sizes, more variables, more compute. Those are valuable things, but only after a more fundamental question has been answered: does this dataset still represent the system we are trying to understand?

The correct ordering is Reality validity → regime validity → data relevance → model specification → optimization, but too often organizations begin near the end. No amount of model optimization can repair a dataset whose governing relationships no longer transport into the current system. You simply create a better model of an obsolete Reality.

So what kind of historical knowledge maintains its relevance across time?

This was originally a CFO question in 2025 that led somewhere unexpected. If successful strategies are highly dependent on their environments, then the historical record of what worked may depreciate rapidly. But what about the historical record of how humans fail? That turns out to be very different.

Human beings repeatedly exhibit remarkably persistent reasoning failures: confirmation bias, overconfidence, base-rate neglect, motivated reasoning, the planning fallacy, exceptionalism, sunk-cost thinking, false causal inference, ignored externalities, untested assumptions, confusing a persuasive story with strong evidence, and mistaking historical success for proof of future portability. Technology changes enormously, markets change enormously, political systems change, wars change, institutions change — and human beings continue making many of the same mistakes.

This creates an important asymmetry: what worked yesterday may not work tomorrow. But how people fool themselves barely changes at all.

In sum, what works changes with the environment. How humans make mistakes has changed very, very little since the Bronze Age.

This asymmetry is the foundation underneath a very different way of using generative AI. Most GenAI applications are generative in the obvious sense — we ask what we should do, what is likely to happen, what strategy will work. Those questions require the model to infer forward from accumulated historical knowledge, which means they inherit all of the regime-sensitivity problems described above.

But there is another side of these systems. A reductive use of GenAI asks something different: where might this reasoning fail? What assumptions are unsupported? Where has correlation been mistaken for causation? Which externalities have been ignored? What must remain true for this conclusion to survive? Where is the author assuming that their situation is exceptional? This reductive mode does not primarily ask historical knowledge to predict the future — it uses the historical record of recurring human failure to stress-test assertions made in the present, which dramatically changes the epistemic burden.

The simplest analogy may be a crash test. A crash test does not predict the exact accident your car will experience; it asks whether the structure survives known classes of failure. Reasoning can be subjected to the same discipline. And because the failure modes being tested are regime-resistant — because humans keep arriving at new situations carrying the same cognitive vulnerabilities — the historical record of human failure becomes one of the most durable datasets available to AI. We cannot safely assume that yesterday’s recipe for success will work tomorrow. But we have much stronger reason to believe that tomorrow’s humans will remain susceptible to many of yesterday’s reasoning failures.

The larger lesson

The AI industry has spent enormous energy asking how much human knowledge we can preserve and place inside machines, and we may eventually discover that this was only half of the problem. The harder question is determining which knowledge should continue governing decisions after the world that produced it has changed. That changes how we should think about data lakes, the economic value of proprietary data, and hallucination itself. It challenges the assumption that longer historical series are intrinsically better, and it places an important boundary around the idea that enough data and compute can eventually make correlation functionally equivalent to causal understanding.

Scale can reduce uncertainty inside a stable regime. It cannot prevent Reality from changing regimes. And when volatile change outruns the ability of a model to recognize that transition, something particularly dangerous becomes possible: the AI can stop representing Reality accurately without stopping looking intelligent. Its prose remains fluent, its logic remains coherent, its historical evidence remains enormous, its confidence may remain intact — and it can simply become wrong about the world.

Which brings us back to the strange asymmetry at the center of all this. We have spent decades believing that the great informational asset of humanity is the accumulated record of what we have learned. Maybe. But in an era when knowledge depreciates faster and Reality changes more violently, another part of the record may prove even more durable: the record of how we repeatedly got things wrong.

The world keeps changing, up and down and all around. With lots of volatility and velocity across most domains.

And we keep bringing ourselves along for the ride.

Time for a different approach. #ProvingGround

The first is the desire to always be right. It sounds like intellectual confidence. It presents itself as expertise, certainty, decisive leadership. But underneath, it is defensive. Its primary function is to protect the current model — the worldview, the framework, the prior conclusion — at the expense of reality. It is the engine behind confirmation bias, behind the 95% confidence score quietly treated as 100%, behind every organization that mistook its map for the territory until the territory hit back.

This is not a character flaw unique to bad people. It is the default human setting. The brain is a prediction machine that rewards the sensation of being right and punishes the sensation of being wrong. It’s also the default GenAI setting. That’s not an accident, given that AI is authored by people and leverages human-generated data. Under pressure, under time constraint, under the accumulated weight of a career’s worth of positions publicly taken, the desire to protect the model becomes almost irresistible. And the model closes.

The second relationship is the desire to not be wrong when you could have been correct. That sounds like a small distinction. It is not.

The desire to not be wrong when you could have been correct is forward-facing. It does not pretend that wrongness is always avoidable. Fog of war exists. Unknown unknowns exist. The future is genuinely uncertain in ways that no amount of rigor can fully resolve. Sometimes you will be wrong despite doing everything right, and the honest response to that is equanimity, not self-flagellation.

But avoidable wrongness is different. Avoidable wrongness — the kind that results from not looking when you could have looked, from not testing when you could have tested, from treating a pattern as a cause because it was convenient — is in many ways a moral failure, not just an analytical one. The measure this imposes is precise and uncomfortable: the gap between what you knew and what you did with what you knew.

There is a third element, and it is the one that makes the other two dangerous at scale.

Lying is the attempt to create an alternative reality without telling anyone that you’re doing that. And this definition matters because it includes the lies that don’t feel like lying. The guidance that extrapolates optimistically beyond what the data supports. The confidence score presented as certainty. The strategic narrative or the financial guidance that emphasizes what confirms the case and quietly omits what doesn’t. The dashboard that shows what was measured rather than what matters. None of these require a conscious decision to act in bad faith. But they are all, by this definition, attempts to create an alternative reality without disclosure.

The private version is no less damaging. The self is not actually contained. Your model of reality — however privately held — is the instrument through which you act in the world. Every decision flows from that model. If the model is built on a lie you’ve told yourself, the lie propagates outward through everything you touch. And from inside a self-constructed alternative reality, you cannot see what you’re doing to the people downstream of your decisions, because from inside it, you’re not doing anything wrong. You’re just acting on what you know.

This is why the epiphany is sometimes genuinely traumatic rather than just intellectually exciting. It doesn’t just add new information. It retroactively reveals the shape of the lie you’d been living inside.

All three failure modes — wanting to always be right, failing to avoid preventable wrongness, and lying — share the same structure. A model is substituted for reality without the substitution being declared. The correlative analyst isn’t usually lying in the conscious sense. But the architecture of what they produce — the closed system, the false certainty, the map presented as territory — is structurally identical to a lie. It creates an alternative reality. It doesn’t tell anyone that’s what it’s doing.

The most dangerous lies are the ones the liar has told themselves first.

There is one structure that prevents all three simultaneously. It is the confrontation — sustained, compressive, uncomfortable, metabolically expensive — with Known Reality on one side and Your Own Ignorance on the other. You cannot lie your way through that. The compression finds it. Every assumption gets tested. Every load-bearing assertion gets asked whether it is actually bearing the load. Every gap between what you knew and what you did with what you knew becomes visible.

That confrontation is what Proving Ground was built to produce.

Proving Ground is a fiduciary oversight and decision safety tool that’s launching pretty soon. It takes any proposal, plan, strategy, or assertion and then subjects it to structured stress-testing using the subtractive side of Generative AI. Total elapsed time is about 30-40 seconds. It leverages the power of Claude models and the most dependable performance patterns in the world: the patterns of human failure. Failure patterns don’t hallucinate.

Working with Anthropic, PG surfaces the load-bearing assumptions, identifying what would have to be true and remain true in order for the claims to be investable. It doesn’t tell you what to do, but it makes sure that when the claims fall apart, you and your colleagues are fully aware of what not to do.

Proving Ground helps you to not be preventably wrong from the beginning, which is where most teams go wrong. And it extends that standard past the comfortable, private boundary of the self because decisions are not abstract objects. They live in the world. They have consequences for people who had no voice in making them, in a future Reality the decision-maker may never inhabit. So Proving Ground is ultimately transparent about its analysis.

Being wrong when you could have been right has consequences for many people.

That is not just embarrassing. Today, it is an ethical and fiduciary breach.

This trend has been slowly forming for 2-3 years and is now accelerating.

By the end of 2026, Finance will be the lead economic buyer for all such systems in this company. And they are placing much greater weight on data reliability and fiduciary oversight than “ease of use.”

I asked about specifics re “systems of record”. His answer was all ERPs and functional technology stacks. He drew a sharp contrast between cost-side tools and systems which “Finance has controlled for a long time” and what he called “value for money” performance tools and tech stacks like CRM, martech, HR tech, supply chain tech, etc.

I asked him how he saw this affecting the knowledge workers in these areas. “We believe they will become part of a significantly expanded Finance organization.” I was particularly interested to hear that they’d already pulled the CDAO organization under Finance, and that those teams will be used to operate the systems of record according to Finance mandates.

My last question was why they were doing this. Clearly, I have a view here, but I wanted to understand their impetus.

“We have learned the hard way that most functions operate tech stacks in a very self-interested way that is typically not a very accurate representation of reality. We can’t afford to continue down this path given the regulations being promulgated by a whole range of global jurisdictions. Finance, by its very nature, is the function who most represents reality day-in-and-day-out.”

This is not the first company to have done something like this, but it is an exemplar of a slow-moving trend that has started to accelerate. What do you think about this?

Features that attract suspicion

Both humans and automated systems tend to associate these features with AI:

• Uniformly medium-length sentences

• Paragraphs of nearly identical size

• Excessively smooth transitions

• Repeated thesis restatement

• Balanced constructions: “not merely X, but Y”

• Enumerated triads

• Comprehensive coverage without strong selection

• Abstract claims unsupported by observed particulars

• Constantly competent grammar

• Qualification without genuine uncertainty

• An emotionally frictionless voice

• Conclusions that summarize rather than transform the argument

• Familiar metaphors used correctly but unsurprisingly

• A tendency to explain what the reader already understands

None of those characteristics proves AI involvement. Several are also hallmarks of competent academic, professional, technical, and second-language writing. Research has found that detectors can disproportionately flag non-native English writers precisely because controlled, less idiomatic English is more statistically predictable. In one Stanford-linked study, detectors classified a large majority of sampled TOEFL essays by non-native writers as AI-generated. Stanford HAI

What reads as more fully human

Human-authored prose tends to become persuasive as human work when it exhibits situated specificity:

• Details that matter to the author but are not needed to complete the assignment

• Uneven emphasis—spending 500 words on one point and 40 on another

• Changes in rhythm that correspond to changes in thought

• Genuine intellectual commitment rather than synthetic balance

• Admissions of confusion, error, embarrassment, or changed belief

• Observations that could only come from being somewhere, knowing someone, or doing something

• Metaphors that organize the author’s thinking rather than merely decorate it

• Arguments that develop through discovery instead of announcing their entire structure in advance

• Deliberate omissions and unresolved tensions

• A recognizable relationship to the author’s earlier work

That last point is important. “Human” is increasingly judged longitudinally, not textually. Draft history, notes, citations, correspondence, previous writing, and the ability to discuss one’s choices are better evidence of authorship than stylistic detection.

A crucial distinction: AI-generated versus AI-aided

Style alone is especially incapable of distinguishing these categories:

1. Fully AI-generated

2. AI-generated and heavily revised

3. Human-drafted and AI-polished

4. Human-written with AI research or structural assistance

5. Fully human-written

They can converge on virtually the same final surface. Current detection research consistently finds substantial degradation when detectors encounter different domains, unfamiliar models, or edited text; there is no robust cross-domain method that can reliably infer the production history of a finished passage. 2026 cross-domain benchmark OpenAI itself withdrew its early text classifier because of its low accuracy and warned that it should not be used as a primary decision-making tool. OpenAI classifier notice

So the practical hierarchy is:

Most vulnerable to being labeled AI: polished, generic, comprehensive, symmetrical, convention-following prose.

Least vulnerable: selective, situated, rhythmically irregular, intellectually committed prose with a documented relationship to a particular person’s experience and prior thinking.

The irony is that traditional advice to make writing “clearer, smoother, more objective, and better organized” often pushes human prose directly toward the statistical and rhetorical profile now associated with AI.

“I once heard a very wise saying.

Your hard work is the dream of all unemployed people.

Your restless child is the dream of everyone who doesn’t have children.

Your house, no matter how big or small, is every homeless person’s dream.

Your lack of debt is every debtor’s dream.

Your health deteriorating is the dream of every patient with an incurable disease.

Your peace, restful sleep, and your food are the dreams of everyone in a country at war.

Be grateful for everything you have.

After all, nobody knows what tomorrow will bring.”

That’s where we are right now with correlative analytics. Nowhere is that failure more consequential than in two domains that literally price and govern civilizational risk: macroeconomic policy and property insurance.

Both fields were built on the same foundational bet: that past relationships between variables contain enough signal to predict future outcomes with acceptable confidence. For most of the 19th and 20th century, that bet paid off. The world was stable enough, slow-changing enough, and well-understood enough that correlation-based models could function as reasonable proxies for causal understanding.

That world is gone. The models haven’t gotten the memo.

What Correlative Analytics Actually Is (and Isn’t)

Before we talk about collapse, we need to be precise about what we’re critiquing. Correlative analytics — the broad family of statistical methods that includes regression, econometric modeling, actuarial tables, and most machine learning — is not wrong in any simple sense. It is a legitimate and powerful tool for describing relationships between variables in historical data.

The problem is ontological, not mathematical. Correlation describes what has co-occurred. Causation describes what produces what. These are not the same question, and in a stable, slowly-changing world, conflating them is an affordable shortcut. In an unstable, rapidly-changing world, it is a source of systematic error.

Judea Pearl’s causal hierarchy is useful here. At the first rung, we observe and describe — what correlates with what. At the second rung, we intervene — what happens when we change something. At the third rung, we counterfactualize — what would have happened under different conditions. Virtually all of econometrics and actuarial science operates at the first rung while making claims that require the second and third. That gap — between the epistemic foundation of the model and the epistemic demands of the decision — is where the collapse is happening.

Econometrics: The Policy Model Is Eating Itself

Central banks, finance ministries, and intergovernmental bodies like the IMF and World Bank run on econometric models of staggering complexity. These models are not naive — they incorporate sophisticated statistical techniques, massive datasets, and decades of theoretical refinement. The people who build and use them are genuinely brilliant.

And yet the models keep failing in ways that aren’t random. They fail systematically at structural breaks — moments when the underlying causal relationships between variables change. The 2008 financial crisis. The inflation surge of 2021-2023. The supply chain dislocations of the pandemic period. The energy price shocks from the Ukraine war. In each case, econometric models either failed to anticipate the event or dramatically mispriced its trajectory.

This isn’t a calibration problem. It’s an architecture problem.

Econometric models are trained on historical data that encodes the causal structure of a prior world. When that causal structure changes — when globalization reverses, when energy systems reorganize, when labor markets shift their behavioral foundations — the historical correlations become actively misleading. The model confidently extrapolates from a world that no longer exists.

Consider the Federal Reserve’s inflation models in 2021. The Fed models saw low inflation expectations, anchored by decades of price stability, and projected that pandemic-era monetary expansion would produce modest, transitory inflation. What the models couldn’t see was that the causal mechanism keeping inflation suppressed — globalized supply chains absorbing demand shocks through cheap overseas production — was in the process of breaking down. The correlation between money supply growth and inflation was structurally different than it had been. The models were looking at a rearview mirror and calling it a windshield.

There is also a subtler problem: the policy interventions themselves change the causal landscape. This is the econometric equivalent of the Heisenberg problem. When the Fed raises rates to cool inflation, it changes the behavior of mortgage markets, corporate borrowing, bank lending, and consumer confidence in ways that alter the very relationships the model depends on. The model cannot adequately account for the feedback between the policy instrument and the system it’s measuring, because doing so would require a level of causal understanding the model doesn’t possess.

The confidence intervals get wider. The error bars grow. Policymakers receive model outputs with diminishing predictive validity while the models’ institutional authority remains intact.

Actuarials: The Great Risk Mispricing

Property insurance is in a different kind of crisis, but the mechanism is the same.

Actuarial models price risk by analyzing historical loss data — how often certain types of events have occurred, what they’ve cost, and how those patterns vary by geography, property type, and other observable characteristics. The statistical machinery is sophisticated. The data sets are large. The profession is rigorous by its own standards.

But actuarial models are fundamentally backward-looking instruments. They price the risk that has existed, not the risk that is emerging. In a stationary risk environment — where the frequency and severity of loss events is relatively stable over time — this is a workable methodology. In a non-stationary risk environment — where the underlying causes of loss are changing faster than the historical record can capture — it becomes a systematic mispricing engine.

We are now firmly in the non-stationary era.

Wildfire risk in the American West has been repriced catastrophically as climate patterns shifted, vegetation dried, and the wildland-urban interface expanded. The historical loss data, which informed actuarial models for decades, encoded a world of different fire behavior. Insurers who stayed loyal to their models lost enormous amounts of money. Many of them — including major carriers — have now exited entire state markets: California, Florida, Louisiana. This is not a market correction. It is a model failure.

The deeper problem is compound causality. Wildfire risk isn’t just about weather. It’s about the interaction of temperature, humidity, wind patterns, vegetation density, urban development patterns, building codes, fire suppression capacity, and the nonlinear feedback between these variables. Actuarial models can incorporate some of these factors as correlates. They cannot adequately model the causal interactions between them, because doing so would require moving up Pearl’s hierarchy — into interventional and counterfactual reasoning — and actuarial practice is not built for that.

The same dynamic is playing out in hurricane risk, flood risk, and increasingly in liability lines that are exposed to climate-adjacent litigation. The historical correlations are breaking down because the causal mechanisms that generated those correlations are changing. The models trained on that history are not just imprecise — they are systematically wrong in predictable directions.

And here’s the dangerous part: the confidence with which these models report their outputs doesn’t diminish as the underlying signal degrades. The model doesn’t know what it doesn’t know. It produces probability estimates and loss projections with the same apparent precision regardless of whether the historical data it’s trained on reflects a causal environment that still exists.

The Half-Life of Knowing

What’s happening in both domains can be understood through a single framework: the half-life of a model’s causal validity.

Every predictive model encodes a set of causal assumptions about how the world works — which variables drive which outcomes, through which mechanisms, at what magnitudes. Those assumptions have a half-life. In a stable world with slow structural change, the half-life is long — decades, potentially. Historical data remains a good guide to present reality. In an unstable world with rapid structural change, the half-life shortens dramatically.

We are now in a world where the half-life of causal validity for both macroeconomic and property risk models has shortened to years, or in some cases months. But our institutional use of these models has not adjusted. We are still treating them as if they have the epistemic authority of long-half-life instruments. We are borrowing confidence from a past that no longer applies.

This is what Correlative Collapse looks like in practice. It’s not that the models produce nonsense — they produce plausible-looking outputs with genuine historical grounding. The collapse is in the gap between apparent precision and actual validity. The models become numerically sophisticated instruments for measuring a world that has already changed.

What Actually Needs to Happen

The answer is not to abandon statistical modeling. It’s to change what we ask these models to do and how we govern their use.

At the methodological level, both fields need to invest seriously in causal model architectures — tools that explicitly represent the mechanisms by which outcomes are produced, not just the historical patterns of co-occurrence. This means structural causal models, directed acyclic graphs, and interventional reasoning frameworks. These are not exotic research tools — they exist and are increasingly tractable. They are simply not yet integrated into the institutional workflows of policy modeling or actuarial practice.

At the governance level, both fields need epistemic audit functions — systematic processes for tracking how fast the causal assumptions embedded in a model are degrading relative to real-world evidence. Model accuracy is not enough. You need model validity — ongoing verification that the causal structure encoded in the model still corresponds to the causal structure of the world it claims to represent.

At the institutional level, both fields need to develop what might be called fiduciary epistemic standards — enforceable obligations to disclose not just model uncertainty, but model vintage and causal coherence. A model built on pre-2020 data and applied to post-2020 risk environments should carry something like a structural warning label: this instrument’s causal validity has not been verified against current conditions.

None of this is technically impossible. What it requires is an institutional willingness to acknowledge that the floor is falling out, and that the methodological inheritance of the 20th century — however sophisticated, however refined — is no longer adequate to the epistemic demands of the 21st.

The Stakes

Why does this matter beyond the professional communities directly involved?

Because econometric models are the epistemic foundation of macroeconomic governance. When they fail, policy fails — and the costs are borne by people who never heard of a DSGE model in their lives. Mispriced inflation. Mismanaged recessions. Fiscal policy calibrated to a causal world that no longer exists. These aren’t abstract errors. They are transfers of harm from institutions to populations.

And because actuarial models are the mechanism by which private risk is socialized or not socialized. When they fail systematically, insurers exit markets, homeowners lose coverage, and governments face pressure to become the insurer of last resort — at costs that haven’t been properly modeled either. The people most exposed are almost always the people with the fewest alternatives.

Correlative Collapse isn’t a technical problem dressed up in policy language. It’s a civilizational infrastructure failure dressed up in technical language.

The models we built to understand the world have become obstacles to understanding it. That’s the diagnosis. The prescription requires something harder than better statistics.

It requires an honest reckoning with the limits of knowing — and the discipline to build institutions worthy of that honesty.

The first is accuracy — let’s call it Scylla. This the obligation to get it right, to honor the actual structure of the argument, to resist the seductive simplification that makes a thing easier to say but subtly wrong.

The second is understandability — let’s call it Charybdis. This is the obligation to reach the reader where they are, to build a bridge between what you know and what they can receive, to accept that even the most important truth, unheard, is functionally inert. Writers who have thought seriously about their craft know these two as permanent adversaries. Compress too hard and you mislead. Refuse to compress and you lose the room.

But there is a third monster, the Sirens. It is and the most dangerous of the three precisely because it is the hardest to see because it is the most human.

You've navigated the rocks of accuracy, survived the whirlpool of understandability, arrived at a piece that is both true and clear — and then nothing happens. You’ve finished, feeling the pleasant weight of having understood something. And then you set it down. No decision gets made. No behavior changes. No mental model shifts into action.

The piece was beautiful and convincing. But it did nothing. The Sirens didn’t attack you. They seduced you into inaction by reminding you of the cost of taking action.

The difference between this failure and the other two is that it often goes undetected — by the writer and the reader alike.

A piece that fails on accuracy eventually embarrasses someone.

A piece that fails on understandability loses its audience visibly.

A piece that fails on practicality gets praised for being insightful and then forgotten.

In some domains, that outcome is merely disappointing. In domains where decisions carry real fiduciary weight, where the gap between knowing and doing has real consequences — it is a serious failure dressed as a success.

The question is not how to find a midpoint between the three. There is no midpoint. Odysseus didn't split the difference between Scylla and Charybdis — he chose the lesser danger deliberately and accepted the cost. The craft is in knowing which cost to accept in which piece for which reader, and being honest with yourself that you are making a choice, not achieving a synthesis.

Most writing about difficult subjects operates on a single plane — as though accuracy, understandability, and practicality are competing for the same real estate. They are not. They operate at different layers of the same piece, and the writer's job is to architect those layers rather than collapse them.

The surface of a piece — what a skimming reader takes away — should optimize for understandability and practicality together. A sharp claim, a concrete implication, an image that lodges. This is not a concession to shallow reading. It is a recognition that the surface claim is what determines whether anyone reads further, and that the reader who finishes understanding only the surface claim should still leave with something true and actionable. The surface is not decoration. It is the promise.

The middle carries the argument — enough logical structure that an engaged reader can follow the reasoning without a specialized vocabulary. This is where analogy earns its keep, where the concrete example does the work that formal language cannot, where the reader who wants more than the surface claim finds the architecture that supports it.

The deep layer is where accuracy lives in its full form. It may be a footnote, a linked appendix, a reference to underlying work that the specialist reader knows how to find. Its job is not to reach everyone — it is to ensure that the piece cannot be legitimately accused of being wrong, that there is somewhere the precision lives even when the surface has made its necessary compressions.

This is how serious legal documents work. The operative clause must be clear and actionable — a reader who only reads that clause should be able to act correctly. The whereas clauses carry the precision that protects against the edge cases. Both are necessary. Neither substitutes for the other.

But there is a specific problem that doesn't appear in most writing about difficult subjects, and it matters to be clear about it.

Most writers navigate these tradeoffs while writing *about* a domain. The trickier situation is writing from inside a domain where the distinction between correlation and causation is not rhetorical — it is load-bearing. A wrong simplification in that context doesn't just confuse a reader. It can actively reinforce the epistemic errors the work exists to correct. It can send someone away more confident in a flawed mental model than they arrived.

This sharpens the role played by the Sirens considerably. When you offer a reader an actionable conclusion, you are implicitly telling them the causal model is settled enough to act on. That is a strong claim. The honest version of practicality is not false certainty. It is a tripartite structure: here is what you can act on with confidence, here is what you should treat as provisional, here is what you should not yet decide. That structure is more actionable than false certainty, not less, once readers are trained to receive it. It also happens to be true.

There is a related and underappreciated problem in where we locate the standard for truth itself.

Current discourse around AI-generated content treats provenance — who wrote it —as a proxy for validity. It’s as though the origin of a claim tells you something reliable about its accuracy. It does not.

Origin and Provenance are a heuristic, and in this case an increasingly unreliable one. The correct standard has always been: what does the content represent, and does that representation hold under interrogation? This is a harder standard than checking a source. It is also the only standard that actually works.

The human outputs that serve as the implicit baseline are themselves the product of motivated reasoning, agenda, preference masquerading as analysis, and the full distribution of human cognitive error at scale. The corpus is not clean. The baseline is not reliable. And style markers — the length of a dash, the construction of a sentence — tell you nothing about correspondence to reality. They tell you something about the training data that shaped a model's stylistic defaults, which is a different question entirely.

Detecting what wrote X and then naming it authentic or inauthentic is not a quality standard. Indeed, and I say this with care, it is a stupid, amateur-hour distinction.

The deeper question underneath all of this is one that has no technical answer.

Reality is the state of things as they are at a moment — contingent, dynamic, subject to causal forces that are themselves always changing. Truth is the set of relationships and principles that govern how Reality behaves — the deep structure underneath the flux. I tend to stick with Reality as it is largely unarguable. Truth is very often about what you or I believe.

The confusion between these two concepts is not a modern problem. Pontius Pilate asked his question in real time, to someone who was making precisely that distinction, and Pilate either couldn't hear it or couldn't afford to. *What is truth?* from a Roman prefect who had to manage a volatile political situation that morning was not a philosophical inquiry. His question of Jesus was a practical man's rhetorical impatience with a claim that transcended his all-too-present circumstances.

Many religious traditions characterize people as either “lovers of Truth” or lovers of their own view. These characterizations are exact. The lover of Truth has already made peace with the possibility that Reality will disappoint them. Their preference is ordered correctly — they want to know what is actually true more than they want confirmation of what they hoped was true. To love Truth is a moral achievement before it is an intellectual one. And it is counterfactually bulletproof: the person who genuinely loves Truth cannot be destabilized by evidence, because evidence is not a threat to them. It is the point. The same applies to Reality, which is neither our enemy nor our friend. Reality is the most valuable teacher, should we decide to learn from it.

The reason this matters for writing is that it matters for reading. An audience of Truth-lovers is the only audience that will follow you all the way through the argument — past the comfortable stopping points, through the conclusions that complicate their prior positions, into the territory where the practical implication is genuinely demanding. Writing for that audience is different work than writing for an audience that wants to feel informed. The piece has to earn the right to make hard demands, and then it has to make them.

The three monsters are real. The path through requires knowing which cost you're accepting, being honest that you're accepting it, and trusting that the reader who loves Truth will meet you where the argument actually goes — not where it would have been more comfortable to stop.

The model still runs. The dashboard still populates. The confidence intervals still look fine. But somewhere between the training data and the present moment, something fundamental shifted — and the system has no idea. It is still navigating by a map drawn in a world that no longer exists.

That moment is the Wall of Reality. And for a significant number of data science and AI teams, it is coming faster than they think.

The Wall Has Another Name

I’ve been calling this dynamic Correlative Collapse for some time now, and the Wall of Reality is its most concrete expression.

Correlative Collapse describes what happens when systems optimized on historical co-occurrence patterns lose their predictive grip as the conditions that generated those patterns change. It’s not a gradual, measurable degradation. It’s a cliff. The system functions, then it doesn’t — and because the failure is epistemically invisible until it’s operationally catastrophic, organizations rarely see it coming.

The Wall of Reality is what Correlative Collapse looks like when you’re standing in front of it.

Business leaders experience the Wall as a governance problem: why didn’t our models predict this? Why did we invest so heavily in tools that failed us?

Data scientists experience it as a technical problem: the model’s performance metrics looked fine right up until they didn’t.

Both framings are incomplete. The Wall is an epistemic problem — a failure not of computation, but of understanding. And that distinction matters enormously for what comes next.

What Makes the Wall Invisible

Here’s the diagnostic core of it: correlation-based systems are structurally incapable of detecting their own obsolescence.

A model trained on three years of pre-pandemic consumer behavior doesn’t know that a pandemic happened. A revenue prediction model trained on a decade of stable interest rates has no representation of the rate environment we’ve lived in since 2022. A hiring algorithm trained on the profiles of past successful employees encodes the selection biases of that era as signal.

These systems are not broken. They are doing exactly what they were built to do — find patterns in historical data and project them forward. The problem is that “forward” keeps changing, and the models cannot follow.

The silence is the tell. High R². Low RMSE. Clean backtests. The metrics that signal model health were all generated before the conditions changed. By the time the performance data reveals the problem, the decisions have already been made.

This is what I call epistemic debt — the accumulating liability of decisions made on correlative confidence without causal grounding. Like financial debt, it accrues quietly and comes due suddenly.

Why GenAI Makes This Worse, Not Better

The most dangerous misconception in the current AI moment is that large language models solve this problem.

They don’t. They accelerate it.

LLMs are extraordinarily sophisticated pattern-matching engines — the apotheosis of correlative reasoning. They can synthesize, translate, explain, and generate with remarkable fluency. What they cannot do is reason causally. They do not know why things happen; they know what tends to follow what, in the text they were trained on.

When organizations layer LLM-assisted analytics on top of existing correlative infrastructure, they produce outputs that are more articulate, more confident-sounding, and just as epistemically fragile. The fluency creates an illusion of understanding. The Wall doesn’t move — you just approach it faster, with better prose.

The Three Signs You’re Closer Than You Think

Most teams don’t know they’re approaching the Wall until impact. But the approach has signatures:

1. You can predict well but not explain why. If your team can tell you what will happen next quarter but cannot construct a defensible causal account of why, you are running on correlative momentum. That momentum has a half-life — and you don’t know how long it is.

2. Your model’s last major recalibration was triggered by failure, not by design. Healthy epistemic systems have explicit protocols for stress-testing assumptions against changing conditions. If your last big model update came because something went wrong rather than because you were proactively auditing causal validity, you’re reactive by default.

3. Your confidence doesn’t degrade as your data ages. Confidence in correlative models should decline as the temporal distance between training data and present conditions increases. If your dashboards don’t reflect that degradation — if a model trained in 2022 is still generating the same confidence scores in 2025 — you are not measuring uncertainty. You are performing it.

The Exit Is Causal, Not Correlative

Here’s the solution-forward part, and it requires honesty about what it demands.

The exit from Correlative Collapse is not better data or bigger models. It is a shift from correlative reasoning to causal reasoning — asking not just what tends to follow what, but why, and what we could do to change the outcome.

Judea Pearl’s causal hierarchy gives us the structure: prediction (rung one), intervention (rung two), counterfactual reasoning (rung three). Most enterprise AI operates at rung one. The Wall is what happens when rung one reasoning meets a world that has moved on. More recently, Fiduciari has added a “Fourth Rung” we call Causal Engineering that uses high fidelity synthetic data to be able to use Causal AI to probabilistically reverse-engineer what it would take to reach a business objective in the future, net of time lag and externalities. The future is the anchor point, not the past.

Moving up the hierarchy is not primarily a technical challenge — though the technical challenges are real. It is an organizational and epistemic challenge. It requires building cultures that reward causal humility over predictive confidence, that treat “we don’t know why this works” as a risk flag rather than an irrelevant philosophical question, and that distinguish between fit and understanding.

It requires, in other words, governing AI the way you would govern any other source of fiduciary risk — with explicit accountability for the assumptions baked into the system, not just the outputs it produces.

What This Means for Leaders

For business leaders, the Wall of Reality reframes AI governance. The question is not “is our model accurate?” It is: “Do we understand why our model works well enough to know when it will stop working — and what we’ll do when it does?”

If you cannot answer that question, you have epistemic debt on your balance sheet. It doesn’t show up in any traditional audit. But it will show up eventually — in decisions that fail, in reputational damage, in regulatory exposure, in the corrosive realization that your AI investment was built on a foundation that couldn’t hold.

For data science and AI teams, the Wall of Reality is a professional challenge. The skills that got you here — feature engineering, model selection, hyperparameter tuning, MLOps — are real and valuable. But they are not sufficient to navigate the epistemic terrain ahead. Causal inference, structural causal models, and the capacity to reason about interventions rather than just predictions: these are the competencies that determine what’s on the other side of the Wall.

This has profound implications for the intensifying standards of fiduciary duty of oversight and the SEC’s new strictures on decision governance / decision safety, both for AI and HI.

The Wall Is Not the End

I want to be precise about something: Correlative Collapse is not an argument against machine learning or predictive modeling. It is an argument for understanding what those tools actually do, and what they cannot do on their own.

Correlation finds patterns. Causation explains them. You need both — but in the right relationship, with the right epistemic humility about which is doing which.

The Wall of Reality is not the end of AI-driven decision-making. It is the end of the era when we could pretend that prediction was the same as understanding.

The organizations that get to the other side will be the ones that figured that out before impact.

Mark Stouse is Chairman and CEO of Fiduciari.ai, Inc., a technology-enabled consultancy operating at the intersection of AI, capital allocation, and fiduciary duty of oversight. He writes here on Substack.

And somewhere in that silence, the gap between what the model believes and what is Reality continues to widen — until it doesn’t anymore, because something breaks.

This is Correlative Collapse. It is not a bug. It is a structural feature of every system built on the assumption that the statistical relationships observed in historical data will persist into the future.

Correlative Collapse is the problem behind the collapse of econometric models discussed by the last Fed chairman in the spring of 2026, and echoed by a governor of the Bank of England. It is the issue forcing property insurers into binary business decisions they’d rather not make, such as “we are no longer writing policies in California.” Specifically, the actuarials that have both assessed their risk levels and enabled them to price risk to their customers are failing because the basis — correlation — is losing fidelity with Reality because Reality is changing fast.

Understanding it precisely — not just intuitively — is one of the most important things a technical practitioner can do right now.

The Core Mechanism

A correlative model learns a function f that maps inputs X to outputs Y based on patterns observed in training data D, drawn from a distribution P(X, Y | θ_t) where θ_t represents the underlying data-generating process at time t.

The model then operates in production at time t + Δt, where Δt may be months, years, or — in the case of many enterprise systems — a decade.

The implicit assumption is that P(X, Y | θ_{t+Δt}) ≈ P(X, Y | θ_t). That is: the joint distribution of inputs and outputs has not changed meaningfully since training. This assumption is never formally stated. It is never tested on a schedule. It is simply inherited by every prediction the model produces.

When that assumption holds, the model performs. When it breaks — due to market shifts, behavioral change, regulatory change, technological disruption, demographic drift, or any of a hundred other mechanisms — the model does not know. It continues producing predictions at the same confidence level, because its confidence is calibrated against D, not against reality at t + Δt.

This is the central pathology: a correlative system’s confidence is a function of its internal coherence, not its current validity.

The Numerator/Denominator Problem

Think of model confidence as a ratio. The numerator is the evidence supporting the model’s learned associations. The denominator is the total evidence relevant to whether those associations still hold — including evidence the model has never seen.

A model optimized on historical data maximizes the numerator. But it has no mechanism for growing the denominator as the world changes. As Δt increases and conditions diverge from the training distribution, the denominator grows in reality while remaining fixed in the model’s accounting. Confidence becomes increasingly overstated — not because the model is miscalibrated against its training data, but because the training data itself is becoming a less valid sample of the current world.

Standard calibration techniques don’t fix this. Platt scaling, isotonic regression, temperature scaling — these adjust the model’s confidence relative to held-out data from the same distribution. They don’t address temporal distribution shift. A perfectly calibrated model can still be catastrophically wrong if the distribution has moved.

The Half-Life of Knowing

Every dataset has a half-life — a period after which the predictive signal in that data has degraded by half relative to its value at collection.

Half-life is not uniform across domains. It is a function of the rate of change in the underlying system being modeled:

• Financial market microstructure: half-life measured in seconds to minutes

• Consumer behavioral patterns: months to low single-digit years

• Demographic and cultural trends: years to decades

• Physical constants: effectively infinite

Most enterprise ML systems are built without explicit half-life estimates for their training data. This is equivalent to running a chemistry experiment without knowing whether your reagents have expired.

The practical implication: model refresh cycles should be derived from data half-life estimates, not from performance metric degradation observed in production. By the time production metrics degrade visibly, you have already been operating in the red zone for some time. The decay is a leading indicator; the failure is a lagging one.

A rough obsolescence index for a model’s training data can be expressed as:

O(t) = 1 - e^(-λΔt)

Where λ is the domain-specific decay rate and Δt is the time since data collection. When O(t) exceeds a domain-specific threshold, model outputs should be treated as epistemically untrustworthy even if performance metrics appear stable.

The Oil Refinery Analogy

A useful structural analogy: a correlative model is like an oil refinery designed for a specific crude feedstock. The refinery is highly optimized — yields are maximized, waste is minimized, throughput is efficient. But change the feedstock — heavier crude, different sulfur content, different viscosity — and the refinery doesn’t adapt. It continues processing, but the outputs degrade in quality, and in extreme cases the process itself breaks down.

The refinery has no sensor that measures “feedstock similarity to design spec.” It has sensors for temperature, pressure, and flow rate — proxies for performance. But those proxies can remain nominally acceptable even as the feedstock diverges significantly from the design envelope.

Data is the feedstock. The model is the refinery. Standard monitoring metrics are the temperature and pressure gauges. What’s missing is a sensor for distributional distance between current input data and training data — and even more fundamentally, a sensor for whether the causal structure that generated the training data still holds.

The ATS Case Study

Applicant Tracking Systems provide a clean, well-documented business case of Correlative Collapse in practice, driven not by a desire for more effectiveness in recruiting but by a desire for a less costly talent acquisition process. The problem is, it merely reallocates the cost and the risk to later in the employment life cycle.

ATS algorithms trained on profiles of past successful employees learn associations between input features (education, keywords, prior employer prestige, career trajectory) and the target variable (success as defined by tenure, performance rating, or promotion). These associations are real — they existed in the training data.

The causal structure they encode, however, reflects the hiring context of the training period: labor market conditions, industry norms, internal culture, the profiles of the people doing the evaluating, and the definition of “success” itself. As each of these changes, the learned associations become progressively less valid as causal proxies — but the model continues to use them as filters.

The result is a system that becomes increasingly good at selecting candidates who resemble past hires in contexts that no longer exist, while systematically filtering out candidates who are well-suited to the current and future context. The model’s confidence in its rankings is high. Its validity is declining.

The failure is not detectable from within the correlative framework. It requires an external, causal audit: mapping what the model actually uses as signal against a current causal account of what actually drives success in this role, in this organization, at this moment.

Why Pearl’s Hierarchy Frames the Exit

Judea Pearl’s causal hierarchy provides the formal structure for understanding why correlative systems can’t self-correct:

Rung 1 — Association: P(Y | X). Observational. What tends to follow what. This is where all standard ML operates.

Rung 2 — Intervention: P(Y | do(X)). Interventional. What happens when we actively change X. Requires a causal model, not just observational data.

Rung 3 — Counterfactual: P(Y_x | X = x’, Y = y’). What would have happened under different conditions. Requires a structural causal model and assumptions about individual-level mechanisms.

Correlative Collapse is what happens when a Rung 1 system is trusted to answer Rung 2 and Rung 3 questions — which is, functionally, what every business decision made on the basis of a predictive model is asking it to do.

The exit from Correlative Collapse requires ascending the hierarchy deliberately: building interventional models grounded in structural causal assumptions, testing those assumptions explicitly, and maintaining ongoing audits of whether the assumed causal structure still holds as conditions change.

This is not a call to abandon correlative modeling. Rung 1 is fast, scalable, and often adequate for stable domains. It is a call to know which rung you’re on — and to stop expecting Rung 1 tools to answer Rung 2 and 3 questions without penalty.

What Monitoring Gets Wrong

Most MLOps monitoring frameworks are designed to detect Rung 1 failure: drift in input distributions, degradation in held-out performance metrics, data quality anomalies. These are valuable signals. They are not sufficient.

What they miss is silent distributional shift with intact surface statistics — the scenario where the joint distribution of inputs and outputs has changed, but in ways that don’t immediately surface in the metrics being tracked. This happens when the shift is in the causal structure rather than the marginal distributions. The inputs look similar. The outputs look similar. The relationships between them have changed.

The practical implication: monitoring for Correlative Collapse requires causal auditing, not just statistical monitoring. This means periodically re-estimating the causal relationships the model implicitly assumes, comparing them against current data, and explicitly testing whether the interventional logic underlying the model’s use case still holds.

Most organizations do not do this. Most MLOps tooling does not support it. That gap is where Correlative Collapse lives.

The Honest Summary

Correlative Collapse is not a failure mode that better engineering eliminates. It is a structural consequence of the epistemic gap between correlation and causation — a gap that exists in every model trained on historical data and deployed into a changing world.

The question is not whether your models will encounter it. They will. The question is whether your organization has built the epistemic infrastructure to detect it before it becomes a governance event — and to reason causally about what to do when it does.

The wall is real. The map that gets you through it is causal, not correlative.

Mark Stouse is Chairman and CEO of Fiduciari.ai, Inc., a technology enabled consultancy operating at the intersection of causal AI, risk, capital allocation, and fiduciary compliance. He writes here on Substack.

One of the most enduring ideas in modern management is that organizations improve by comparing themselves with other organizations. Entire industries have been built around benchmarking, best practices, peer comparisons, market averages, analyst expectations, and conventional wisdom. Consultants sell it. Boards request it. Investors expect it. Executives rarely question it.

The underlying logic appears almost self-evident. If successful companies behave in similar ways, then studying those behaviors should reveal the path to success. If most organizations have converged upon a particular practice, then that practice presumably survived because it works. If thousands of experienced executives believe something, their collective experience should carry more weight than the opinion of any single individual.

There is, of course, considerable truth in that reasoning. Human knowledge has always depended upon accumulated experience. Civilization itself is built upon the ability to learn from those who came before us rather than rediscovering every principle from first principles. The problem is not that benchmarks are useless. The problem is that we often misunderstand what they actually represent.

A benchmark is not Reality.

It is a historical description of Reality.

That distinction appears almost trivial until the world begins changing quickly.

For much of the twentieth century, markets evolved at a pace that made historical comparisons extraordinarily useful. Industries changed, but they often changed gradually enough that yesterday remained a reasonable guide to tomorrow. Production methods improved incrementally. Consumer preferences drifted rather than lurched. Competitive landscapes evolved over years instead of quarters. Under those conditions, benchmarking worked remarkably well because the causal system generating business outcomes remained relatively stable.

Today that assumption has become increasingly difficult to defend.

Artificial intelligence is restructuring knowledge work. Global supply chains are reorganizing in response to geopolitical fragmentation. Capital has become more expensive. Regulatory environments shift with unusual frequency. Customer expectations evolve continuously as technology changes what is possible. In many industries, the environment that produced last year’s benchmark has already begun disappearing before this year’s planning cycle has concluded.

Yet most organizations continue managing as though the benchmark itself possesses enduring authority.

This is where management quietly confuses representation with Reality.

Every benchmark is a summary of historical observations. It tells us what happened. Sometimes it even tells us what happened consistently. What it cannot tell us, by itself, is why those outcomes occurred or whether the same causal mechanisms remain in place today. A benchmark captures patterns. It does not capture causes.

That distinction matters far more than most executives realize. When environments remain stable, historical patterns often continue because the underlying causal relationships remain largely unchanged. During periods of rapid disruption, however, the causal architecture itself begins to move. Relationships that once appeared reliable weaken or disappear. New constraints emerge. Time lags lengthen or shorten. Externalities become more influential. Historical correlations become progressively less representative of the Reality they once described.

The benchmark has not become incorrect. It has become unrepresentative.

But that’s not thinking. That’s reacting with extra steps.

Real thinking, critical thinking — the kind that builds businesses, spots opportunities others miss, and holds up under pressure — is deliberate. It has a learned method. It is not “intuitive.” And at the heart of critical thinking are two questions so simple they look almost trivial.

They are not trivial.

The first question: “If this is true, what else must be true?”

This is the question of consequences.

Every belief, assumption, or observation you hold carries with it a hidden architecture — a web of implications that follow necessarily from it. Most people never explore that architecture. They accept a premise and stop. The idea sits there, isolated, never tested against the world it implies.

The crucial move is to follow the premise wherever it leads.

If it’s true that technology is dropping barriers to entry in every industry simultaneously — what else must be true?

QED, it must be true that competitive advantage can no longer live in technical skill alone. Which means it must be true that something else is now doing the work that skill used to do. Which means it must be true that the people who understand what that something is have an enormous, and growing, edge over those who don’t.

Notice what happened there. One observation, followed honestly, becomes an orientation towards Reality.

This is how rigorous thinkers operate. Not by accumulating more information, but by extracting more signal from the information they already have. The question “if this is true, what else must be true?” is a forcing function. It demands that you take your own premises seriously — seriously enough to follow them into uncomfortable territory.

Because here’s the thing: most people sense the implications of their beliefs. They just don’t go there. It’s easier to hold a convenient idea loosely than to follow it to conclusions that require you to change something.

The first question won’t let you off that hook.

The second question: “For this to be true, what else must be true — and remain true?”

This is the question of foundations.

Where the first question looks forward — at what a belief implies — the second looks underneath. It asks: what has to be holding for this to stand? What are the load-bearing assumptions? And crucially: are they actually holding?

This is where most strategies, plans, and confident predictions quietly fall apart.

A business plan is a set of beliefs and assumptions. Nothing else.

You may be recoiling from this last statement. “I have data, I have facts, I’ve done research!” Ok, but definitionally these are all about the Past, otherwise they wouldn’t be available to us. What if that view of the Past is not Prologue? Isn’t a business plan about the as-yet unexperienced Future?

That’s the uncomfortable thing about Reality that is also its opportunity: it changes very dynamically. Just because X was Reality for your business in 2020 does not mean it’s still Reality today. In fact, as we can all see quite plainly, it’s not. Reality hasn’t changed a little. It’s changed a lot. Has your thinking, your assumptions, your beliefs changed that much in response?

Most business plans fail not because the idea was wrong, but because one or two foundational assumptions — things the plan needed to be true in order to work — turned out not to be. They were never examined. They were simply inherited.

“For this to be true, what else must be true and remain true?” is the question that surfaces those assumptions before they surface themselves — which they will, eventually, always, and usually at the worst possible moment.

Consider how this works in practice.

A company believes: we can win by offering a lower price than the competition. Fine. But for that to be true, what else must be true? The market must be primarily price-sensitive. Customers must know your price is lower. Your margins must support it sustainably. Your competitors must not be able or willing to match you. Each of those is a load-bearing wall. Remove any one of them and the whole structure comes down.

Most companies discover this by living it. That’s the most expensive way to learn.

The second question lets you discover it by thinking it.

Together, these questions form a complete instrument.

The first question is expansive. It opens territory. It shows you what your beliefs commit you to, what world they describe, what actions they demand.

The second question is structural. It stress-tests. It shows you what your beliefs depend on, what has to hold for them to hold, what would break them.

Used together, they do something remarkable: they turn thinking into a discipline rather than an intuitive reflex.

Socrates understood that most people don’t actually know what they know. They hold positions without understanding their foundations. They draw conclusions without following their implications. They mistake familiarity for comprehension.

His method — relentless, uncomfortable, occasionally infuriating — was simply to ask. To keep asking. To follow the logic wherever it led and surface whatever it revealed.

The two questions above are that method, distilled.

This matters more now than it ever has.

We are living through a period where information is effectively infinite and attention is effectively zero. Where AI can generate analysis, strategy, and some level of insight on demand. Where the raw material of thinking — data, research, frameworks — has never been more abundant or more accessible to everyone.

Again, QED, that means that AI is not the moat. It cannot be the moat if it’s equally available to everyone.

And precisely because of that, the thinking itself has never mattered more. That means that you have never mattered more. And that I have never mattered more.

Because here’s what GenAI cannot do:

It cannot tell you which questions to ask.

It cannot decide what matters.

It cannot sit with an uncomfortable implication and choose to follow it anyway.

It cannot look at a load-bearing assumption that everyone has accepted and ask, quietly, but does this actually hold?

That’s yours. That’s the work only you can do. Technology can make thinking faster, easier, less expensive. But it cannot originate it.

And the people doing that work — building the habit of these two questions, applying them to everything, following them honestly even when they lead somewhere inconvenient — those are the people who are going to see further, decide better, and compound that edge over time in ways that no tool can replicate and no competitor can easily copy.

We cannot be poseurs on this. Indeed, the more you claim to think critically but actually don’t, the faster the truth will be known. This is not about your intelligence or mine. This is about a discipline that aligns us as closely as possible to Reality.

Start small. Start today.

Take the next belief you encounter — in a meeting, in a strategy document, in your own head — and run it through both questions.

If this is true, what else must be true?

For this to be true, what else must be true, and remain true?

Don’t rush the answers. Sit with them. Follow them. You will be surprised how quickly a seemingly solid idea begins to reveal its architecture — and how often that architecture turns out to be less solid than it looked.

That discomfort is not a problem.

In many cases, I’ve seen it be the thing that saves me from a big mistake.

AI belongs on that list. Short of those few exceptions, nothing has transformed human performance at this scale or speed.

But here is what that history also tells us: technology has never homogenized human contribution. It has always stratified it. Every transformative technology created a new hierarchy of capability — those who mastered it, those who used it adequately, and those who didn’t engage with it at all. The distances between those groups widened, not narrowed, after each transition.

AI is no different. In fact it may be the most dramatic stratification event in generations. Which means this: we have never been more important as individuals. And AI is driving a completely new ordering of human contribution and performance based on how well — how honestly, how rigorously, how self-awarely — we use it.

For decades, competitive advantage in technology was essentially an asset problem. You owned something your competitor didn’t — proprietary data, switching costs, network effects, patents, distribution. The company with the biggest pile of defensible assets won. Technology was the primary moat generator because technology was scarce and hard to build.

That theory is now obsolete, and its obsolescence is accelerating at a pace that most organizations haven’t fully registered.

The models are extraordinary and increasingly accessible to everyone. The infrastructure is commoditized. The data advantages are eroding as training sets converge. You can’t build a durable moat from technology alone anymore because your competitor can access substantially the same technology tomorrow morning. What was once a castle wall is now a floor.

So where does the moat go?

It migrates to the only thing that can’t be commoditized: how specific human beings think, see, and behave in relationship with the technology.

This is the central irony of the current AI moment. The entire discourse has been about humans being displaced by technology. What’s actually emerging is the opposite: the humans who know how to work with AI in ways that are rare, calibrated, and hard to replicate are becoming more valuable, not less. The technology amplifies the human asymmetry rather than erasing it. The floor rises, and the distance between people who know how to stand on it and people who don’t grows wider, not narrower.

The moat hasn’t disappeared. It has moved — categorically, permanently — from attributes of technology to attributes of people with, and without, technology.

Consider what this means in practice.

A competitor can study a methodology. They can read the same literature, access the same models, hire engineers from the same talent pool. What they cannot replicate is the human journey that produced genuine insight — the years of accumulated practitioner experience with consequential decisions, the unusual tolerance for being wrong, the willingness to have one’s own certainty interrogated rather than defended. That journey cannot be purchased or downloaded. It can only be lived.

This is why the idea of a personal moat — once a concept without much strategic content — is now load-bearing. The professional who has spent a year stress-testing their own reasoning doesn’t just have a better tool. They reason differently. The repeated experience of having assumptions interrogated changes how arguments are constructed in the first place. The instrument leaves a residue of better thinking that compounds with use and cannot be transferred to someone who hasn’t done the work.

The person who has been doing this for two years is not just better equipped. They are genuinely harder to beat in any domain where reasoning quality matters.

There is a second dimension that’s equally important and almost entirely overlooked: human behavior as a product moat.

The most durable products in the current environment are not those with the best technology. They are those that, over time, accumulate an irreplaceable map of how human beings actually fail to reason — in specific domains, under specific pressures, across specific professional categories. That map lives in the relationship between the instrument and the humans who use it. It cannot be extracted by a competitor who copies the technology alone, because the technology was never the asset. The human behavior the technology observed and recorded was the asset.

This is a new category of competitive advantage. It is self-reinforcing in ways that traditional network effects are not, because it gets smarter about human reasoning failure the more human reasoning it encounters. The competitor who starts today starts from zero. The gap widens with every evaluation, every session, every user who brings their hardest thinking to the instrument and leaves with something more honest than they arrived with.

The organizations that survive the current transition are the ones that were never actually selling technology. They were selling what they know that technology doesn’t. They were selling the accumulated, calibrated, compounding intelligence of people who understood something important before the market caught up.

The moat is all about how we know what we know, and what we do with it. The moat is human. And it is, for the first time in more than a generation, genuinely hard to buy your way to.

That changes everything about how advantage is built, defended, and transferred. The organizations that understand this now are positioning for a world the rest of the market hasn’t seen yet. The ones that don’t will keep trying to build castles on ground that is no longer high.

Because in the end, what the best AI in the world gives everyone equally is capability and capacity. What it cannot give anyone is the ability to use those capabilities and capacities well. That is the scarce resource. Now, for the first time since the dawn of AI, the market waking up to it. There may be smaller teams of more capable people going forward. But capability and capacity will be defined by the quality of our judgment and how well we can use technology to think, to solve problems, and exploit opportunities.

It’s something much more subtle.

It’s substituting the seller’s Reality for the buyer’s Reality.

Every sales organization lives inside its own causal system. Quotas must be met. Pipeline must grow. Forecasts must improve. Investors expect results. Management wants certainty. Marketing is measured on engagement, Sales on conversions, Customer Success on retention.

All of these pressures are real.

But they’re real for the seller.

The buyer lives in an entirely different Reality.

Their budget may be frozen. A new CFO may have changed investment priorities. Internal politics may have stalled consensus. A critical operational problem may have eclipsed the one you’re trying to solve. Or perhaps nothing has changed enough to justify organizational change at all.

The mistake isn’t that sales teams ignore these factors.

The mistake is that they unconsciously assume the buyer’s decision process mirrors their own urgency.

It doesn’t.

This realization has profoundly changed my own thinking over the past several years.

When we first built Proof, we naturally viewed the world through the lens of the seller. We had developed a fundamentally different approach to understanding business causality, and we assumed that explaining it well would be enough. Like many technology companies, we believed the challenge was communicating our superiority.

Over time, we discovered something more important.

Our technology wasn’t the primary barrier.

The customer’s Reality was.

Many organizations simply weren’t structured to adopt a sophisticated causal platform. They lacked the organizational maturity, cross-functional governance, internal expertise, or executive alignment necessary to take advantage of it. The limitation wasn’t technical. It was contextual.

That insight changed everything.

It led us to shift from a pure software model toward software-enabled advisory. Not because the software became less valuable, but because customers often needed help changing how they thought before they needed another tool.

That’s a very different starting point.

Today, when we work with executives, we don’t begin by explaining causal AI. We begin by understanding the Reality they’re trying to negotiate.

• What pressures is the CFO facing?

• What risks concern the board?

• Where are political incentives aligned—or misaligned?

• What externalities are shaping the business?

• What assumptions are quietly driving today’s decisions?

Only after we understand that Reality does the technology become relevant.

This distinction also explains why so much of modern GTM struggles with forecasting.

Organizations measure signals: website visits, content downloads, demo requests, intent data, meeting frequency, email engagement.

These observations are useful, but they aren’t causes. They are effects.

The real question isn’t whether buyer activity increased.

The real question is: What changed inside the buyer’s Reality that made those behaviors emerge?

That’s a fundamentally causal question.

It also reframes the role of sales.

Great salespeople don’t persuade buyers to leave Reality.

They help buyers navigate the Reality they already inhabit.

Sometimes that leads to a purchase.

Sometimes it leads to waiting.

Sometimes it leads to deciding that another priority matters more.

Ironically, understanding that Reality often produces better commercial outcomes than trying to overcome it.

There’s a broader lesson here that extends well beyond sales.

Organizations frequently become “customer-centric,” but still fail because they never become Reality-centric.

Reality doesn’t belong to the buyer or the seller. It simply exists.

The organizations that consistently make better decisions are the ones willing to replace hope with understanding, assumptions with evidence, and signals with causality.

In the end, every forecast, every strategy, and every major business decision is really a prediction about Reality. The question is whether we’re modeling the customer’s—or merely projecting our own.

Fable 5 is magnificent but costly. Sonnet 5 has some substantial issues.

Specifically, if you have agents running on Claude, or you’ve built a product offering on its Sonnet 4.6 model, you need to immediately test and re-evaluate their operational integrity if they are running now on Sonnet 5.

We have picked up extensive issues between Sonnet 4.6 and 5, ranging from adherence to instructions, refusal of instructions by Sonnet 5 that are accepted by Sonnet 4.6 and Fable 5, and the generation of different outcomes to varying degrees of severity.

The baseline of activity on Sonnet 4.6 is very strong and consistent. But as soon as the operating model is switched to Sonnet 5, you can get — and almost assuredly will get — compliance and operational compliance issues.

Fortunately, Fable 5, which represents the strategic direction for Anthropic, does not exhibit these transitional issues. But Fable 5 also is expensive enough to preclude it as an option.

We will share more information as we can

.

Markets moved. Commentators and politicians called it historic. But it’s fair to say that they meant very different things in using that word.

The MOU language has been published. What no one showed you was how well the agreement’s logic actually holds up — what is dependably load-bearing and what isn’t.

The analysis below was produced in less than 60 seconds by Proving Ground, a decision stress-testing tool built on one operating principle: before you commit to a consequential position, find out where your reasoning breaks. Not where it might break. Where it does.

Built on causal logic, Proving Ground makes no recommendations. It doesn’t tell you what to think about the Islamabad MOU. What it does is apply a structured interrogation to the reasoning the agreement — or any content whatsoever — asks you to accept — and return an honest account of which parts hold and which parts don’t.

One thing worth noting before you look at the output: this was produced using Claude Sonnet 4.6 — the model roughly 9 in 10 Claude users access by default. Not a premium configuration. Not a showcase build. What you’re about to see is what the tool most people already have produces when the methodology is sound. We’ve also run the Proving Ground alpha very successfully on Opus 4.8 and for a brief moment on Fable 5.

The analysis begins by identifying every assumption the agreement depends on. Not the ones it states explicitly — those are easy. The ones it requires in order for its promises to be true.

A Hormuz reopening by a stated date requires Iran to act in good faith under a clause that says “best efforts.” A nuclear freeze requires a precise definition of what “status quo” actually means. A $300 billion reconstruction fund requires someone to fund it. Each assumption gets interrogated on a single question: for this to be true, what else must be true — and must remain true?

When you follow those threads, the document starts to look different from the headlines.

Two of ten assumptions in the MOU held under interrogation. Eight didn’t. That produces a Survivability Score of .222.

That score means the MOU is Unreliable, not a Failure. Why? The ceasefire is real. The Hormuz reopening mechanics are sound. Both sides have clear economic incentive to follow through on the near-term provisions. These are genuine achievements, and the analysis doesn’t dislodge them.

What it does find is that the agreement’s three hardest problems — Iran’s nuclear program, Israel’s behavior as a non-signatory actively operating in Lebanon, and a $300 billion reconstruction fund with no named contributors — are not solved. They are deferred. And the sequencing means the United States provides economic relief before those problems are resolved, spending the leverage that produced this agreement before the agreement’s hardest work begins.

This is what stress-testing reasoning looks like. It isn’t pessimism. It isn’t contrarianism. It’s the discipline of separating what has been established from what has merely been asserted — and refusing to let that difference collapse under the pressure of a good headline. In business, it shows you how confident you can be in a plan, proposal, or other assertion. It shows you how much of it survives the stress test, how much of it is investable.

The ten questions at the end of the analysis are not rhetorical. They are the specific conditions under which the rating changes. When those questions have answers, run it again. The score will move.

Until then, you have what the document actually supports.

Proving Ground is a decision integrity platform built by Fiduciari.ai. It stress-tests reasoning before consequential commitments are made. It returns decision authorship to you. If you’re interested in trying it when it launches, please send a LinkedIn PM to Mark Stouse with the word “Trial PG” in the headline.

A recent MIT study made headlines by suggesting that heavy AI use weakens cognitive engagement — that people who lean on tools like ChatGPT think less, retain less, and over time may become less capable. The findings were real. The interpretation was incomplete.

Here’s what the study couldn’t measure: why people reach for AI in the first place.

There’s a strange irony running through our professional culture right now. The same people who pride themselves on data-driven thinking have decided, largely without evidence, that AI use is either cheating or capitulation. So people hide it. They use it heavily and present the results as unaided work. The shame has driven the practice underground — which means we’re having a loud public debate about something most people won’t publicly admit to doing.

That’s not a conversation. That’s a performance.

The critics performing the loudest are making exactly the same mistake they accuse AI users of making: accepting a narrative without stress-testing it.

Let me offer a different frame for your consideration. One that’s been sitting in plain sight for about 10,000 years.

The first genuinely human-enhancing technology was arguably the lever. Simple. Elegant. Transformative. A lever multiplies the amount of weight a person can move — dramatically, reliably, every time.

So here’s the question: does using a lever mean you’ve misrepresented your personal strength?

Is refusing to use a lever somehow more virtuous, more pure?

Does using a lever frequently mean you are in its thrall, that it is making you weaker?

The questions are almost embarrassing to ask out loud. Of course not. Of course not. Of course not.

But here’s what’s less obvious and more important: put a lever on a table and walk away. Come back fifty years later. Assuming that no one else exercised their agency and moved it, the lever will be exactly where you left it. The lever has no agenda. No initiative. No judgment. It is entirely and only what you make of it.

So is AI.

The tool has no agency. None. It doesn’t reach out and grab you. It doesn’t seduce you into passivity. It doesn’t quietly erode your judgment while you sleep. It lies exactly where you put it, doing exactly nothing, until a person picks it up and makes a decision about how to use it.

If you’re saying, “what about Agentic?” right now, the reality of Agentic right now and as far as we can see now if that it is sophisticated rules-based automation. No agency. It executed the agency you gave it.

Which means every outcome — cognitive growth, cognitive atrophy, genuine augmentation, quiet redundancy — is a human decision. Fully. Without exception.

Now here’s where it gets interesting.

Most people reading this have encountered Maslow’s Hierarchy at some point. The pyramid. Survival at the base. Safety above it. Belonging. Esteem. And at the top, Self-Actualization. The full expression of what a person is capable of becoming.

I want to suggest something: the way you use AI is a direct reflection of where you currently live on that pyramid and where you want to go.

We are in a period of extraordinary, relentless change. The rate of disruption across industries, institutions, and daily life has exceeded most people’s capacity to integrate it meaningfully. When the map keeps changing faster than you can read it, you stop consulting the map. You navigate by habit, by feel, by principles, by heuristics, by whatever reduces immediate uncertainty.

This isn’t laziness. It’s rational triage under conditions of chronic overload. But it has a cost. The upper registers of Maslow’s hierarchy — growth, mastery, the hunger to understand harder things — require a kind of metabolic surplus. Psychological slack. The felt sense that the ground beneath you is stable enough to support upward movement. Relentless change is a direct tax on that surplus.

Many people are running deficits just managing the present. And AI, in that context, fits the shape of the exhaustion perfectly. It offers closure. Relief from uncertainty. A way to stop the discomfort of not knowing.

That’s not AI’s fault. That’s a civilizational condition AI is being blamed for. Most of the stress created by the incessant change we’ve seen was created not by random circumstances but by the unintended consequences of human action.

Someone reaching for AI from the bottom of the pyramid uses it as shelter. Someone reaching for it from the top uses it as a ladder. Same tool. Completely different outcome. And the tool gets blamed for the outcome that was already determined by where the person was standing when they picked it up.

I’ve been using various categories of AI intensively for years. In that time my thinking has gotten sharper, not duller. My frameworks have deepened. My tolerance for complexity has increased. I find myself capable of sustaining problems that would have been out of reach before.

But I’m honest about why.

I don’t come to AI for answers. I come to it looking for where the argument breaks. I aim the conversation directly at points of constraint — the places where clean logic gets complicated, where the framework makes an assumption it hasn’t earned, where the position has a soft underbelly. That’s where the actual work is.

The result is a continuous series of what I’d call micro-epiphanies. Gaps filled. Assumptions surfaced and examined. Errors corrected. Each one restructures the lattice that holds prior knowledge — which is genuine learning, not retrieval.

This is the Socratic method, essentially. And AI is an extraordinarily capable sparring partner for it — available at any hour, without ego, without the social cost of admitting ignorance to a colleague.

One more thing: GenAI is vastly more accurate when it’s being used to poke holes in assertions than when it’s suggesting what we should do next or how to optimize it. But you have to know how to prompt it.

It also requires something most people find genuinely uncomfortable: the willingness to be wrong. The appetite for correction over confirmation. The decision to find your own blind spots more interesting than your own certainty.

That decision is entirely yours. The lever doesn’t make it for you.

There’s one more dimension that doesn’t get discussed, because it’s uncomfortable.

Some people are not just offloading the work to AI. They are offloading the accountability.

If the AI said it — generated it, recommended it, structured it — then when it fails, when it’s shallow, when it’s wrong, there’s a ready-made defendant that isn’t you. The tool becomes a buffer between the person and the consequences of their own judgment. Or rather, their abdication of it.

But follow that logic to its destination.

If you can fully offload the work — and simultaneously offload the accountability for that work — to something that has no agency whatsoever, that will lie undisturbed on a table for fifty years if left alone — then you haven’t used a tool. You’ve made yourself redundant. You’ve demonstrated, with perfect clarity, that the judgment and authorship that justified your seat at the table was never really yours. Or you’ve abandoned it so completely that the distinction no longer matters.

Every organization right now is quietly running this calculation. Who here is genuinely augmented — bigger, faster, more accountable, higher output? And who is merely insulated — producing acceptable-looking work while the actual judgment has quietly left the building?

The first group becomes more valuable. The second group is already redundant. They just don’t know it yet.

I teach from time to time at USC. Grad students. On Day 1, I tell my students to use AI. Use it and use it effectively. No restrictions. Then I tell them I’ve made the projects and exams significantly harder to account for the power of AI.

My line is this: “If you’re going to be the Six Million Dollar Man or Woman — with an intellect augmented by AI — then a lot more can be expected of you.”

They are usually so stunned they never even try to mount a counter-argument.

I find that disappointing, actually. I’d welcome the argument. But what I think stops them isn’t confusion. It’s the sudden recognition that I’ve closed the exit. The tool is available. The escape hatch isn’t. They own the augmentation. Which means they own the result.

University faculty often find this somewhat controversial. I’ve heard their arguments. But I’d ask them the same question I’d ask anyone resisting this conversation:

What are you protecting? Virtue? Integrity? A version of your own authority that assumed the tools would stay the same?

My view is that anyone who uses AI to make Life easier is only shortchanging themselves.

Some raise the question of AI developing its own agency — of the tool becoming something more than a tool. That’s a real question and it deserves its own serious treatment. But it is a different conversation from this one.

Right now, today, the lever lies exactly where you put it. What we do with it is entirely ours. What it reveals is entirely about you and me.

That’s the mirror test.

And it has nothing to do with AI.

One utility company in Ohio introduced a new billing rule. Nothing dramatic — just a requirement that data centers pay for power they reserve, whether they use it or not. A standard commercial practice. Pay for what you order.

Seventeen gigawatts of pipeline vanished overnight.

To put that in physical terms: seventeen gigawatts is roughly the output of seventeen large nuclear power plants. It was there on Monday. It was gone by Friday. Not delayed. Not restructured. Gone.

The question worth sitting with isn’t where it went. It’s where it came from.

The numbers behind America’s AI infrastructure buildout are genuinely staggering. Alphabet, Amazon, Meta, and Microsoft have collectively committed approximately $650 billion to AI infrastructure in 2025 and 2026. Across 140 construction projects, data centers representing 12 to 16 gigawatts of new capacity were announced for completion this year alone.

As of now, just about 5 gigawatts are actually under active construction.

JPMorgan analysis puts more than 60% of capacity scheduled for 2027 as not yet having broken ground.

OpenAI’s $500 billion Stargate project — announced with considerable fanfare — has reportedly stalled at its Texas site. Between a third and a half of all US data centers planned for 2026 are expected to be delayed or cancelled outright.

This is not a supply chain story, though supply chains are involved. It is not an energy story, though energy is the binding constraint. It is not a permitting story, though permitting is a real obstacle.

It is a reasoning story. Specifically, it is a story about three things that look identical from the outside — and are almost never distinguished from the inside.

The first is belief.

Belief is the starting position. Something feels true. It pattern-matches to prior experience. It aligns with what respected people are saying. Smart people in the room are nodding. It requires no evidence to exist and no defense to persist. It simply is — a proposition that has found its way into the reasoning process and settled there.

Belief is not a failure of intelligence. The people who held it here were not uninformed. They had access to more data, more sophisticated models, and more credentialed advisors than almost any institutions in human history. Belief at this level is not ignorance. It is the default starting condition of every consequential decision ever made by every organization that has ever existed.

The belief here was straightforward: the infrastructure required to support the AI buildout would be available, on the required timeline, at acceptable cost. Power. Equipment. Permitting. Supply chain. Demand. All of it would follow the capital.

That belief was reasonable. It was not unreasonable to hold it. The error came in what happened next.

The second is confidence.

Confidence is what belief becomes when it has been agreed with enough times. It is belief that has been reinforced — by models that confirmed it, by advisors who endorsed it, by momentum that made dissent seem contrarian, by the fact that no one in the room pushed back with sufficient force to require a defense.

Confidence feels like validation. It has the same behavioral signature. It moves capital the same way. It produces the same quality of announced commitment, the same tone in the earnings call, the same certainty in the investor presentation.

But its foundations are social and internal, not evidential. Confidence is not what a proposition looks like after it has survived pressure. It is what a proposition looks like after it has been surrounded by agreement.

The $650 billion moved at the confidence stage. Announcements were made. Commitments were declared. Pipelines were built — on paper, in press releases, in gigawatt projections that were treated as load-bearing before a single foundation had been poured.

The organizations involved were not reckless. They were confident. The distinction felt meaningful from the inside. From the outside, and in retrospect, it was not.

The third is validation.

This is where the cascade fails.

Validation requires a proposition to survive contact with the most uncomfortable version of the opposing case. Not a devil’s advocate exercise conducted by someone who knows their job is to eventually agree. Not a pre-mortem checklist assembled after the decision has already been made. Not a risk register that documents concerns without ever requiring them to be resolved.

Genuine validation is adversarial. It asks what has to be true for the central assertion to be true — and then attempts, seriously and without predetermined outcome, to establish whether those things are actually true.

For the AI infrastructure thesis, those questions were available. They were not hidden.

What has to be true about power availability? Is reserved grid capacity the same as committed grid capacity — and what happens to the pipeline if it isn’t?

What has to be true about supply chains already strained before this demand was layered on top? Memory costs are up five-fold. Transformer lead times have extended to years. What does the thesis look like if those constraints don’t resolve on the assumed timeline?

What has to be true about demand? Not projected demand. Actual committed demand, from actual customers, at actual price points, on actual timelines that align with the capacity being built to serve them?

These questions were not unanswerable. They were unasked. Or asked gently, in forums where the answer was unlikely to stop anything.

The Ohio utility didn’t introduce new costs into the system. It introduced accountability for assumptions that had been present all along. Data centers were reserving grid capacity — committing the utility to hold power in readiness — while treating that reservation as a costless option. When the option acquired a price, the reservation was abandoned.

What vanished wasn’t capacity. It was the appearance of capacity — seventeen gigawatts of announced pipeline resting on an assumption so foundational it had never been named: that reserving power and being willing to pay for power were the same thing.

They were not the same thing. The utility rule didn’t create that gap. It made the gap visible.

Adversarial pressure doesn’t introduce weakness into a structure. It reveals weakness that was already there.

This is the permanent layer of the story — the part that will remain true long after the current AI construction cycle has run its course and been replaced by whatever comes next.

The cascade from belief to confidence to the appearance of validation is not a failure mode unique to AI infrastructure. It is the default failure mode of sophisticated reasoning in institutional settings. And it operates most destructively precisely in the institutions with the most elaborate apparatus for generating confidence — because more analysis, more models, more advisors, and more consensus all accelerate the journey from belief to the appearance of validation without ever requiring passage through the real thing.

The organizations that do this most expensively are not the ones that think too little. They are the ones that think a great deal, in environments structured to produce agreement, measured by metrics that cannot distinguish between confidence and validation, led by people whose professional identity is bound up in the quality of their judgment.

Those people are not wrong to be confident in their reasoning. They are wrong to mistake that confidence for something it isn’t.

There is a difference between a conclusion that has been tested and a conclusion that has been constructed. Between confidence that has survived pressure and confidence that has simply never encountered it. Between knowing something and having successfully defended it against the most uncomfortable version of the opposing case.

The $650 billion buildout proceeded as though that difference didn’t exist.

The market is now conducting the interrogation that should have happened before the capital moved. It is doing so at considerably greater expense, and without any of the advantages that upstream interrogation would have provided — when the questions were still cheap, when the assumptions were still adjustable, when the cost of being wrong was still a fraction of what it became.

The subtractive question is deceptively simple.

Not: what do we believe?

Not: what does our analysis show?

Not: what do our advisors recommend?

But: what has to be true for our central assertion to be true — and have we actually established that those things are true, or have we merely assumed them and surrounded those assumptions with enough agreement to make them feel like facts?

The answer to that question, asked honestly and before the capital moves, doesn’t guarantee a good outcome. Markets are adversarial. Execution is hard. Physical infrastructure is unforgiving. Some bets that survive rigorous interrogation still fail.

But seventeen gigawatts don’t disappear overnight from a pipeline built on defended assumptions.

They disappear from a pipeline built on the distance between confidence and validation — a distance that no one measured, because no one thought to ask whether it existed.

If $650 billion can move on undefended assumptions, what are you moving on?

Each of those nine men will walk to a four-foot-wide rectangle of dirt, alone, and attempt to do something that the entire opposing organization — its coaches, its analysts, its pitchers, its fielders, and its data infrastructure — has spent months preparing to make impossible.

And if the best of them succeeds three times out of ten, he’ll be considered excellent.

That number, the batting average, looks simple. It’s a fraction: hits divided by at-bats. But what it actually measures is something far more interesting than hitting proficiency.

It measures survival.

Start With a 1.000

Here is a thought experiment.

Every batter steps into the box with a perfect 1.000 — a complete possibility. We colloquially call this “batting a thousand.” Everything is still in front of him. The hit exists, somewhere in the probability space of the next few seconds, waiting to either materialize or be destroyed.

What happens to it?

What follows is not metaphor. It is, as close as current research allows, the actual math of what the system does to that 1.000 before the at-bat is over.

The Time Problem — ×.72 — Running Total: .720

The pitcher releases the ball at roughly 94 miles per hour. That is close to the current MLB average — and it means the ball covers sixty feet, six inches in approximately 395 milliseconds.

The human brain requires between 150 and 200 milliseconds just to process what the eyes are seeing.

By the time a batter’s brain has registered what pitch is coming, the window to swing has already closed or is closing. He cannot react to the ball in any meaningful sense. He is making a prediction — based on the pitcher’s arm angle, release point, grip, and the first few feet of flight — and committing to that prediction before conscious confirmation is possible.

He is not hitting the ball. He is betting on it. Every single time.

The 1.000 is now .720.

For context: in 2008, when MLB first began systematically tracking pitch velocity, the average four-seam fastball was 91.1 mph. That sounds like a modest three-mile-per-hour difference. In terms of the reaction window available to the batter, it represents the difference between a nearly impossible task and an even more nearly impossible one. The average has gone up every single year since — without exception.

The Deception Problem — ×.78 — Running Total: .562

Velocity alone would be manageable. The human nervous system is adaptive. Given enough repetitions, a batter could theoretically calibrate to 94 mph.

But the ball is also lying to him.

Spin rate is the mechanism. A four-seam fastball thrown with elite backspin creates an optical illusion: the ball appears to the batter to be rising, defying gravity, arriving higher than physics allows. He adjusts upward. He swings over it.

A breaking ball with elite topspin does the opposite — it appears to be tracking on a catchable plane and then drops off the table as it crosses the zone. He holds his swing a fraction too long, or commits to a pitch that isn’t there.

The average spin rate on sliders and curveballs has risen from roughly 2,106 revolutions per minute in 2015 to 2,475 RPM in 2024. That’s not a minor refinement. That’s a systematic enhancement of the lie the ball tells the batter’s eyes — developed deliberately, tracked precisely, and improved continuously.

He’s not just being deceived. He’s being deceived with increasing scientific precision.

The 1.000 is now .562.

The Intelligence Problem — ×.72 — Running Total: .405

Here is where it gets genuinely uncomfortable.

Every at-bat this batter has ever had is in a database. The pitches he chased out of the zone. The counts where he expands his strike zone. The sequences that made him look foolish. The location in his stance where he is weakest. All of it has been analyzed. Tendencies extracted. A game plan built around that data, specific to him, updated continuously as new information arrives.

The pitcher on the mound is, in a meaningful sense, executing a machine learning output. He knows — because his coaches told him, because the scouting report told him, because Statcast told him — exactly which pitch in which location at which point in the count produces the highest probability of a favorable outcome against this particular human being.

Research confirms what any batter already knows in his bones: when a pitcher hits his intended spot, batter success rates fall by more than half. It is not just what is thrown. It is the precision of where it lands.

The batter has his own preparation, his own film study, his own instincts built over decades. But he is one person with finite time. The intelligence arrayed against him is institutional, continuous, and improving.

The 1.000 is now .405.

The Volume Problem — ×.85 — Running Total: .344

Not long ago, a batter would face the same starting pitcher multiple times in a single game — building pattern recognition, learning tendencies, adjusting at-bat by at-bat. That pitcher would stay in the game until he tired or faltered, giving the lineup time to solve him.

That era is largely over.

Modern pitching strategy is built around specialization and freshness. A batter might face four or five different arms in a single game — each optimized for a specific role, each carrying a different arm slot, velocity profile, and pitch arsenal, each entering fresh while the batter accumulates fatigue and cognitive load.

There is no pattern to solve. There is only a new problem, every time, with a new pitcher who has studied you and whom you have limited information on. The information asymmetry consistently and deliberately favors the mound.

The 1.000 is now .344.

The Space Problem — ×.88 — Running Total: .303

For much of the last decade, teams added another layer: the defensive shift. Armed with years of batted-ball data showing where each hitter tends to put the ball in play, teams repositioned fielders to cover those zones — sometimes dramatically, with three infielders shading to one side of the diamond.

The batter could execute perfectly. He could read the pitch correctly, commit at the right moment, make solid contact, drive the ball exactly where his mechanics naturally send it — and be thrown out by a fielder standing there in advance because the data predicted exactly that outcome.

He hit it well. He was still out. The system had pre-solved his success.

MLB has since restricted the most extreme shifts. The underlying logic — using historical data to collapse the space available to each individual hitter — remains.

The 1.000 is now .303.

The Psychological Problem — ×.79 — Running Total: .239

None of the above accounts for the fact that all of it happens in front of tens of thousands of people, on camera, with his name and his average on the scoreboard, in a profession where failure is not just common but expected and utterly public.

Every other player on the field operates within a system that distributes responsibility. A shortstop who misplays a grounder shares the moment with the situation, the coverage, the defensive alignment. There is context. There is diffusion. There is recovery.

The batter steps into the box alone. The outcome is entirely his. No help is coming. And because failure is the norm — even for the very best — he must walk back to the dugout after failing, sit with it for approximately three minutes, and then be ready to do it again with the same composure and belief as if nothing happened.

Mental resilience is not a soft asset in this job. It is a core technical competency, without which all the physical skill in the world degrades rapidly and visibly.

The 1.000 is now .239.

The Number That Remains

The current MLB batting average is approximately .240.

That is not a coincidence. That is the system producing its output — the residue left after velocity, deception, intelligence, volume, positioning, and psychological pressure have each taken their cut.

The batting average doesn’t measure how good a hitter is in some abstract, graded sense. It measures what the system couldn’t take from him. It is not a score. It is what survived.

Which is why .350 is not an A+ on a hitting exam. It is evidence that a human being went to that box hundreds of times — against everything described above, a system that improves its ability to defeat him every single season — and the system failed to stop him more than one out of every three times.

That is not hitting proficiency. That is something closer to a force of nature.

The league batting average was .271 in 1999. It has fallen nearly every year since. The last player to hit .400 over a full season was Ted Williams in 1941, when fastballs averaged in the low-to-mid 80s, pitchers threw two or three pitch types, and “analytics” meant a manager’s gut. Whether that era or this one was actually harder is a genuine debate. What is not debatable is that the system has never stopped getting better at producing outs.

Nine names on a scoreboard. Nine people who will step into that box alone tonight. The best of them will succeed roughly three times in ten. They will be considered excellent.

Now you know what that number actually costs.

Next Up: What baseball figured out about understanding performance that business still hasn’t.