The Unstable Target
Modeling Entities That Refuse To Remain The Same.
Prediction does not fail because information is absent. It fails because the thing being modeled stops holding still.
The Hidden Assumption Beneath Prediction
Every prediction system rests on an assumption it rarely examines. Not that the future is knowable. Not that the forecast is accurate. The deeper assumption is structural: that the entity being predicted possesses sufficient continuity to be modeled at all. That it will remain recognizably itself between observation and outcome.
Markets. Negotiations. Warfare. Intelligence assessments. Competitive strategy. The underlying machinery differs. The assumption does not. A framework can account for uncertainty. It can account for deception, incomplete information, nonlinear behavior, adversarial adaptation, even recursive reasoning. What it cannot easily survive is a target that is changing faster than the framework can stabilize its understanding of it.
The failure is rarely identified because it feels like insufficient data. We assume the forecast failed because the model lacked data, because the variables were misunderstood, because the signal was hidden inside noise. The possibility that the target itself may have ceased to possess a stable form rarely enters the analysis. The question most prediction systems ask is straightforward: given what is known about this entity, what is it likely to do next?
The prior question is almost never examined: Is this entity sufficiently stable for that question to have a meaningful answer at all? Most of the time, the assumption holds. Institutions drift slowly. Incentives persist. Markets exhibit enough continuity for historical information to remain useful. Strategic actors generally preserve coherent objectives long enough for forecasts to converge on something actionable. Which is precisely why the assumption remains invisible. A premise repeatedly validated by reality eventually ceases to appear as a premise. It becomes part of the environment itself. Only when the assumption breaks does it become visible. And when it does, the failure rarely appears as a failure of prediction. It appears as confusion. Contradictory signals. Sudden irrationality. A growing sense that the environment is becoming harder to read despite no obvious reduction in available information.
The model has not necessarily failed. The target has stopped holding still.
What Prediction Actually Assumes
The natural response to prediction failure is to improve the model. More data. More variables. More computation. Deeper reasoning. If the forecast was wrong, the assumption is that the methodology was insufficient. Given enough sophistication, the gap between prediction and outcome can eventually be reduced. This creates a familiar progression. Prediction evolves from forecasting actions, to forecasting reasoning, to forecasting how an opponent forecasts you. Each layer appears to solve a limitation in the previous one. Each promises greater precision through greater depth.
What rarely changes is the assumption underneath them. Every layer inherits the same premise: that the entity being modeled possesses a sufficiently stable decision-generating process for additional sophistication to matter. This is what most strategic modeling quietly assumes. Not that opponents are simple. Not that they are transparent. The assumption is narrower and far more important. That a policy exists. It may be rational. It may be boundedly rational. It may be adaptive, adversarial, or stochastic. But beneath the observable behavior, there remains some underlying process stable enough that sufficiently persistent analysis can converge toward it.
Recursive reasoning inherits this assumption entirely. The recursion problem is usually framed as a synchronization problem. You expect your opponent to move left, so you move right. Your opponent anticipates this and moves left. You anticipate their anticipation and move right again. The loop appears infinite. But the structure of the problem remains unchanged. The question is assumed to be: which recursive layer is currently governing the environment? Not whether the environment is operating inside a stable recursive structure at all.
This distinction matters. A recursive model assumes an opponent progressing through recursive layers. It does not account for an opponent who abandons the process entirely. An opponent who becomes bored. Distracted. Emotional. Random. An opponent who changes objectives halfway through the game. An opponent whose decision-generating process is no longer the same process the model was attempting to synchronize against. The deeper the recursion becomes, the more fragile the assumption becomes with it.
Additional layers of inference move the model further from direct observation and deeper into assumptions about how reality is being generated. The failure is often misidentified as insufficient recursion. The deeper failure is that recursion itself presumes there is something stable to recurse against. The question was never which loop is correct. The question is whether the thing being modeled remains stable long enough for any loop to matter.
The Unstable Target
The problem is not that the opponent is hiding information. The problem is that the opponent may not currently possess a stable form to hide. Strategic modeling assumes continuity. Beneath every forecast sits an entity expected to remain recognizably itself across time. The objectives may evolve. The tactics may adapt. The behavior may become deceptive. But the system assumes some underlying process persists long enough for observation to accumulate into understanding.
Even adversarial models inherit this premise. A deceptive opponent is still assumed to be pursuing a stable objective. An irrational opponent is still assumed to express that irrationality through a sufficiently consistent pattern. Something is generating the behavior. Something that remains intact long enough to be mapped. The assumption holds until the thing being modeled stops holding. Human opponents violate this assumption constantly, often without intention. Not because they are strategically sophisticated. Because the process generating decisions is rarely as stable as the models built to describe it.
The individual making a decision at one moment may not be operating under the same cognitive conditions a short time later. Objectives shift. Attention drifts. Boredom emerges. Emotional states override strategic reasoning. Novelty becomes more attractive than optimization. Pattern-breaking becomes rewarding for reasons unrelated to the original objective. This is not hidden information. It is ontological instability. The target is not concealing its decision process. The target is transitioning between decision processes faster than observation can stabilize an understanding of any single one.
What the model encounters is not a stable policy obscured by noise. It encounters a policy generator moving through multiple operational states - strategic, emotional, indifferent, random, reactive, anti-pattern - without fixed sequence, without clear boundaries, and often without the awareness of the entity itself. The question changes completely. Not: What is the opponent thinking? But: What is currently generating the opponent’s thinking? Most prediction frameworks can model deception. They can model bounded rationality. They can model incomplete information, uncertainty, and noise. What they struggle to model is a target whose decision-generating process is itself unstable. A target that changes faster than the observer can converge upon it.
This is why certain opponents appear uniquely difficult to predict. Not because they are exceptionally intelligent. Not because they recurse more deeply than everyone else, but because the object being modeled refuses to remain the same object long enough for modeling itself to converge. The prediction system continues refining its understanding. The target continues changing the thing being understood. By the time convergence appears possible, convergence is no longer occurring against the same entity.
Projection Leakage
When the target becomes unstable, prediction does not stop. It continues. What changes is the object being predicted. Under stable conditions, a forecasting system remains anchored to an external entity. Observations accumulate. Patterns persist long enough to be compared against one another. The gap between model and target gradually narrows as new information corrects old assumptions. This process depends on continuity. When the target begins transitioning between states faster than observation can stabilize an understanding of any single one, that continuity disappears. The model loses the ability to update against a consistent external signal.
Prediction does not fail at this point. It substitutes. The model continues generating forecasts, but the source material begins to change. Unable to converge on a stable target, it increasingly relies on assumptions about how targets are supposed to behave. The forecast remains active. The anchor has moved. This transition is difficult to detect from inside the process itself. The reasoning still appears rigorous. The outputs remain coherent. Nothing announces that the model has crossed a boundary between observation and projection. Yet the boundary has already been crossed. The system is no longer primarily modeling the opponent. It is modeling an opponent-shaped construct assembled from its own ontology: assumptions about rationality, strategic consistency, incentive structures, plausible recursion depths, meaningful signals, and legitimate forms of behavior.
Every prediction engine reveals, often unintentionally, the structure of the reality it expects to encounter. An observer who values strategic consistency tends to perceive strategic consistency even after the evidence supporting it begins to weaken. An observer who reasons recursively tends to infer recursive reasoning in others. An observer who finds randomness implausible tends to reinterpret randomness as hidden structure. The model does not reject contradictory information. It translates the information into categories it already knows how to process. The forecast remains operational. Its connection to the target becomes progressively weaker.
As instability increases, prediction gradually transforms into self-reference. The model becomes less a description of the opponent and more a description of the assumptions through which the observer understands opponents in general. The collapse does not announce itself. It arrives as increasingly coherent reasoning about an increasingly fictional target. The forecasts remain internally coherent. The reasoning remains persuasive. The conclusions continue fitting together elegantly.
What disappears is not consistency. What disappears is correspondence. This is the recursive trap in its deepest form. Not an infinite chain of anticipated anticipations. But the gradual replacement of an external target with an internal one. A prediction system begins by attempting to understand reality. Given enough instability, it can end by understanding only itself.
Incentives Survive Longer Than States
If actions are unstable and states are transient, prediction requires a different point of contact with reality. The behavioral surface of an opponent changes continuously. Strategies adapt. Attention drifts. Emotional states emerge and disappear. The same individual may behave strategically in one moment and impulsively in the next. Forecasting at this layer means forecasting the output of a process that is itself unstable. The problem is not insufficient information. The problem is that the object being observed is changing faster than the observation can converge. Not everything changes at the same rate.
Beneath the behavioral surface exist structures that are slower to move, not because they are fixed, but because other parts of the system depend on them. An opponent’s actions can change abruptly. The constraints governing those actions often cannot. Status. Security. Autonomy. Recognition. Survival. The avoidance of specific costs. The preservation of particular identities or commitments. These are not behaviors. They are constraints. The distinction matters.
Prediction frequently fails because it attempts to forecast what an opponent will do. Yet actions are often the most unstable component of the system. The same incentive structure can generate radically different behaviors under different conditions. What remains relatively persistent is not the action itself, but the set of outcomes the system is attempting to preserve or avoid.
This does not make incentive mapping easy. Incentives can be hidden, conflicting, ambiguous, or poorly understood even by the actor pursuing them. But incentives generally degrade more slowly than behavioral patterns. They operate on a timescale that allows observation to accumulate before the underlying structure has already changed. The objective therefore shifts.
Not: What will the opponent do next? But: What outcomes is the opponent structurally constrained from accepting?
Not: Which strategy are they currently running? But: Which realities are they attempting to preserve regardless of which strategy they choose?
This is a different form of prediction. It does not seek to forecast specific actions. It seeks to identify the boundaries within which actions can occur. The forecast becomes less precise. It becomes more durable. The target continues moving. The observer no longer attempts to freeze the movement into a stable image. Instead, attention shifts toward the forces shaping the movement itself. Actions change. States change. Strategies change. The constraints that govern what remains acceptable often change more slowly. And in environments where stable behaviors no longer exist, constraints frequently become more informative than behavior.
Blackwood Verdict
Prediction is usually treated as an information problem. When forecasts fail, the instinctive response is to search for missing variables, hidden signals, better models, deeper reasoning. The assumption is that reality remained stable while the observer failed to understand it. Some failures emerge from a different source entirely.
Some failures emerge from a different source entirely. The observer may possess sufficient information. The methodology may be internally sound. The reasoning may be rigorous. What has changed is the object being reasoned about. The target itself has ceased to provide the continuity the prediction process requires in order to converge.
The distinction is subtle but consequential. A forecasting system can survive incomplete information. It can survive deception. It can survive uncertainty, noise, and adversarial behavior. These are all conditions that preserve the existence of a stable object being modeled. The more difficult failure occurs when the object itself becomes unstable. At that point, prediction begins operating against a moving ontology rather than a moving target.
The problem is no longer that reality is hidden. The problem is that reality is changing shape faster than understanding can accumulate around it. This is why deeper sophistication does not always produce better forecasts. Additional complexity often assumes the target remains sufficiently stable for the added complexity to converge upon. When that assumption fails, more modeling does not necessarily reduce uncertainty. It can amplify it. The prediction system continues refining itself while the object of prediction continues transforming. The distance between them grows even as the analysis becomes more elaborate.
What appears from the outside as prediction failure is often a deeper mismatch. The observer is attempting to converge. The target is no longer remaining the same thing long enough to be converged upon. In such environments, the objective shifts. Not toward perfect prediction. Toward identifying the structures that survive state transitions: the constraints, incentives, dependencies, and boundaries that persist even when behavior does not.
The question is no longer: What will happen next? The question becomes: What remains stable enough to matter regardless of what happens next?
Because prediction rarely collapses at the limits of information. More often, it collapses at the limits of identity.
Blackwood Analysis 006 — Published May 2026