Model Lock

Model lock is the pathological state of an anticipatory system whose internal model has become decoupled from the reality it purports to represent, yet whose control architecture continues to impose the model's predictions on the system. The model is not merely wrong; it is rigid — unable to update fast enough to track a changing environment, and too deeply embedded in the system's decision mechanisms to be bypassed. The result is not inefficiency but systemic destruction: the system acts on a world that no longer exists, and in doing so, destroys the world that does.

The term was first given systems-theoretic form in discussions of allostatic overload, where the anticipatory mechanism of the HPA axis becomes chronically engaged based on outdated predictions of stress. But model lock is not a biological pathology. It is a structural failure mode of any system that relies on prediction — from feedforward control in engineering to regulatory dynamics in economics to climate modeling in geoscience.

The Anatomy of Model Lock

Model lock requires three conditions to develop, and three conditions to sustain:

1. Deep model embedding. The model is not a peripheral advisory tool but a core component of the control architecture. In the Soviet planned economy, the five-year plan was not a forecast; it was the allocation mechanism. When the plan became fiction, the economy continued to be organized around it because there was no alternative allocation path. The model was not merely consulted; it was executed.

2. Slow model updating. The model's revision cycle is slower than the environmental change it must track. Financial risk models that assume stationary volatility distributions are locked into the statistical patterns of a past regime. When market dynamics shift — as they did in 2007–2008 — the models continue to price risk based on the old distribution, and the system that relies on them becomes blind to the new one. The model is not wrong in detail; it is wrong in kind.

3. Strong control coupling. The model's outputs feed directly into the system's effectors without intermediate buffering. A thermostat that is locked into a summer setpoint will overheat a house in winter; a central bank that is locked into an inflation model from the 1970s will raise interest rates in deflationary conditions. The control interface does not question the model; it executes it. This is the difference between model error and model lock: in error, the model is overridden. In lock, the model overrides the system.

Model Lock Across Domains

In economics, model lock is the mechanism by which price signals fail. When a market's pricing model assumes stable demand elasticities, and consumer behavior shifts structurally, the price system continues to signal based on the old elasticities. The result is not merely mispricing but resource misallocation that persists until the model's failures become catastrophic enough to force revision. Financial crises are often model-lock events: the models that governed risk were not absent; they were present and wrong, and their presence prevented the emergence of alternative assessments.

In artificial intelligence, model lock manifests as model collapse — the degradation of generative models when trained on their own synthetic outputs. The model's internal representation of the data distribution becomes increasingly narrow as it amplifies its own biases. But this is a special case of a more general phenomenon: any learning system that cannot distinguish between its own outputs and external reality is at risk of model lock. The reinforcement learning agent that explores only states its critic evaluates positively will never discover states the critic does not yet know exist. The model locks the system into a closed loop of self-confirmation.

In biology, model lock is the failure of developmental programs when environmental conditions change faster than the genome's predictive models can adapt. A plant whose flowering timing is encoded to respond to a historical photoperiod may bloom during a frost if climate change has shifted temperature patterns independently of day length. The genomic model is locked into a correlation — photoperiod and temperature — that no longer holds. The plant does not merely make a bad prediction; it makes a prediction that kills it.

In climate science and policy, model lock is the danger of relying on models that assume linear feedbacks in a system with tipping points. When the climate crosses a threshold — permafrost methane release, ice-albedo collapse — the governing equations change. A model locked into the pre-threshold dynamics will not merely fail to predict the shift; it will actively discourage preparation by projecting stability where instability has already begun. The model's confidence becomes a liability.

The Difference Between Model Lock and Model Error

Model error is universal. All models are wrong; some are useful. Model lock is not universal; it is a property of the system's architecture, not of the model's accuracy. A weather model that fails to predict a storm is in error. A city's flood-management system that continues to build for a historical flood regime after the climate has shifted is in lock. The error is in the model; the lock is in the system's inability to revise or bypass the model.

Model lock is therefore not a cognitive failure. It is an organizational failure — a failure of the coupling between model and controller, between prediction and action. It is the pathology of systems that have made themselves too dependent on prediction, that have allowed the model to become not a tool but a tyrant.

Escaping Model Lock

The standard prescription for model failure is better