Model Based Regulation

Embed a decision-relevant, continuously tested model of the system inside its regulator so interventions are state-aware, predictive, auditable, and revisable.

Overview

Model-Based Regulation is the intervention pattern implied by the Good Regulator Theorem: a regulator becomes more effective when its decisions are conditioned on a workable model of the system it is trying to influence. The model does not need to be a perfect replica. It needs to preserve the distinctions that change what the regulator should do.

A traffic controller needs to distinguish congestion caused by ordinary demand from congestion caused by an incident. A hospital flow team needs to distinguish high census with imminent discharges from high census with a downstream staffing constraint. A cyber defender needs to distinguish an isolated anomaly from a coordinated intrusion. In each case, the visible signal alone is insufficient. The regulator needs a representation of hidden state, relevant dynamics, available actions, and likely consequences.

The archetype therefore adds more than a forecast or diagram. It creates a governed operating loop:

1. define the outcome to regulate;
2. estimate the current state;
3. represent how the system changes and responds;
4. predict or discriminate among action consequences;
5. act within authority, validity, and safety bounds;
6. compare prediction with outcome;
7. revise, replace, or suspend the model when correspondence fails.

The governing principle is decision-relevant correspondence. A model is good enough when it reliably preserves the distinctions needed for action, exposes where it is uncertain, and can lose authority when evidence contradicts it.

The problem it solves

Many regulators are model-based in practice but model-free in governance. Their implicit model is hidden inside habits, thresholds, dashboards, legacy rules, institutional stories, or a vendor system. This creates three recurring failures.

First, the regulator mistakes an observation for a state. A temperature, queue length, loss rate, or risk score can arise from several different configurations. Acting directly on the signal can therefore produce the wrong intervention.

Second, the regulator treats action as if it had an immediate and isolated effect. Real systems contain delay, saturation, feedback, substitution, adaptation, and cross-boundary consequences. A locally corrective action may cause overshoot, oscillation, displacement, or delayed harm.

Third, the regulator lacks a disciplined way to learn from surprise. When an outcome differs from expectation, the discrepancy is called an exception, blamed on execution, or averaged into a performance metric. The underlying model remains unchallenged.

Model-Based Regulation makes the hidden model explicit enough to inspect, validate, and govern. It does not eliminate uncertainty. It organizes uncertainty so the regulator can choose when to act, probe, escalate, or fall back.

What counts as a model

The word model is intentionally broad. Depending on the domain, the decision-relevant model may be:

• a causal map showing feedback, delay, and leverage;
• a state-transition model;
• a set of conditional rules with explicit assumptions;
• a statistical response model;
• a stock-and-flow or agent-based simulation;
• a state-space model with an estimator;
• a digital twin;
• a structured expert model;
• an ensemble combining mechanistic and empirical representations.

The test is not whether the representation is mathematically sophisticated. The test is whether it answers the questions the regulator must answer:

• What state are we likely in?
• What disturbances or constraints matter now?
• What can we change?
• What is each action expected to do, on what timescale, and with what uncertainty?
• What evidence would show that this expectation is wrong?
• What should happen when the model is outside its validity range?

A simple model that answers these questions may be more useful than a high-fidelity simulation that is too slow, opaque, or weakly validated to guide action.

Core transformation

The archetype transforms a reactive regulator into a correspondence-tested regulator.

Before: observation → rule → action.

After: observation → state estimate → model-conditioned action → outcome comparison → model governance.

The additional steps matter because they separate distinct sources of error. The regulator can ask whether the observation was poor, the state estimate was wrong, the action model was wrong, the objective was misspecified, the action was executed incorrectly, or the environment changed. Without that decomposition, every failure looks like generic “bad performance.”

Component architecture

Regulatory Objective

The regulatory objective states what is being preserved, restored, or moved toward. It should name essential variables, target ranges, unacceptable states, priorities, and time horizons. An objective such as “reduce wait time” is usually incomplete because it can be achieved by denying service or shifting delay elsewhere. A usable objective includes the invariants that must not be sacrificed.

The objective is also where normative choices belong. A prediction model can estimate consequences; it cannot decide whose outcomes matter or what tradeoff is legitimate. Keeping those choices explicit prevents the model from laundering value judgments as technical necessities.

System Boundary

The system boundary determines what the model treats as internal state, external disturbance, controlled input, and outcome. Formal ownership is not a reliable boundary. A hospital unit may control beds but depend on diagnostics, transport, discharge destinations, and staffing markets. A platform may control ranking but affect creators, advertisers, and users who adapt strategically.

Boundary design should include material feedback and externalities even when the regulator lacks direct authority over them. Otherwise the regulator may appear successful inside its accounting frame while exporting harm.

State Observation Model

Regulators rarely observe the full state directly. They see measurements, reports, proxies, or delayed traces. The state observation model explains how those signals become an estimate of the condition that matters for action.

A good observation model records missingness, latency, uncertainty, and possible confounding. It also distinguishes “not observed” from “not present.” In high-consequence settings, uncertainty in the state estimate should change action authority rather than merely appear in a report.

Decision-Relevant System Model

This is the core representational component. It connects state, disturbances, interventions, constraints, delay, and outcomes. It should be as small as possible without collapsing distinctions that change the decision.

Decision relevance is tested counterfactually: would two states that the model treats as equivalent ever require different actions? Would two actions that the model treats as equivalent produce materially different consequences? If yes, the model is too coarse for that decision.

The model may contain several layers. A causal layer can identify structure; a statistical layer can estimate parameters; a simulation layer can compare trajectories. The layers should not be mistaken for independent truth. They are coordinated representations with different assumptions and error modes.

Policy Rule

The policy rule converts the state estimate and modeled consequences into action. It may select one intervention, rank options, set a threshold, recommend to a human, or determine escalation.

A model can be informative without being authoritative. The policy rule should therefore state whether the model is advisory, gating, or autonomous. It should also specify how uncertainty, reversibility, and consequence alter the level of authority.

Feedback Signal and Prediction–Outcome Residual

The feedback signal reports what happened after action. The prediction–outcome residual is the consequential mismatch between expectation and observation. This residual is the archetype’s main learning signal.

Residuals should not be reduced to one average score. They need segmentation by operating regime, time horizon, subgroup, action type, and consequence. A model that is accurate in ordinary cases but systematically wrong near a hazard boundary is not a good regulator model.

The residual also needs causal interpretation. A mismatch may arise from execution failure, observation error, an unmodeled disturbance, a bad state estimate, a wrong response model, or structural change. The investigation should preserve these alternatives until evidence distinguishes them.

Validity Boundary

Every model has a validity boundary: populations, scales, regimes, time horizons, input ranges, and disturbance classes for which its use is supported. The boundary must be operational, not a disclaimer buried in documentation.

Crossing the boundary should change behavior. The regulator may gather more evidence, reduce authority, switch to a robust rule, request human review, or enter fallback. Silent extrapolation is one of the most dangerous failure modes of model-guided systems.

Model Update Rule

The update rule defines when and how the model changes. It separates ordinary recalibration from structural revision and replacement. It should identify evidence thresholds, review roles, test requirements, rollback, and communication obligations.

Updating too slowly creates stale-model lock-in. Updating too quickly makes the regulator chase noise and can destabilize the system. The appropriate cadence depends on how quickly the regulated system changes, how much evidence arrives, and how dangerous a bad update would be.

Model Owner and Independent Challenge

The model owner is accountable for scope, evidence, operation, change control, and incident response. Ownership does not mean unilateral authority. In high-consequence settings, independent challenge should be structurally separate from model development and operational pressure.

The challenge function asks which assumptions are carrying the decision, what evidence is absent, which groups or regimes are poorly represented, how the model could be gamed, and what would falsify the recommendation. A challenge channel is valuable only when it can alter deployment or authority.

Safety Bound and Fallback

Safety bounds protect the system when the model is wrong. They may constrain actions, rates of change, exposure, resource depletion, or proximity to hazardous states. Wherever possible, their justification should not depend on the same model they constrain.

Fallback is the operating mode used when observability or model validity is inadequate. It may be conservative manual control, a simpler robust rule, reduced service, or controlled shutdown. Fallback should be designed and rehearsed before model failure; otherwise it is an aspiration rather than a component.

Decision Trace

The decision trace preserves the model version, state estimate, inputs, uncertainty, assumptions, predicted effects, rule applied, action, override, and observed result. It enables incident reconstruction and cumulative learning.

A trace should be proportionate. Recording everything without retrieval structure creates archival opacity. Record what is needed to reproduce why the decision was reasonable at the time and to diagnose why it later succeeded or failed.

Mechanism families

Representation and identification mechanisms

Causal-loop diagrams, state-space models, stock-and-flow models, agent-based simulations, and structured expert models are ways to represent the system. System-identification experiments and natural experiments estimate response relationships.

Choose the representation according to the decision. A causal map may be sufficient to stop an organization from treating a reinforcing loop as a one-time anomaly. A state-space model may be needed where hidden state must be estimated continuously. An agent-based model may help when strategic adaptation and heterogeneity dominate.

State-estimation mechanisms

Bayesian filters, Kalman-family estimators, particle filters, sensor fusion, and structured human assessment update beliefs about state as observations arrive. Their purpose is not to produce a single precise-looking number. It is to maintain a usable estimate and uncertainty distribution.

Estimation should degrade visibly when signals disappear, conflict, or become stale. A regulator that continues acting on an old estimate as if it were current has lost observability.

Predictive action-selection mechanisms

Model predictive control, simulation rollout, scenario comparison, and decision tables use the model to evaluate possible actions. They differ in computational demand and formality, but all should expose constraints and assumptions.

Optimization adds a special risk: it searches for regions where the model says performance is best, which are often the regions where model error is easiest to exploit. Robust constraints, uncertainty penalties, and out-of-distribution checks are therefore more important under optimization than under passive prediction.

Validation mechanisms

Historical replay, forecast backtesting, scenario testing, sensitivity analysis, digital-twin trials, and shadow-mode evaluation test whether the model supports the intended decision. Different tests answer different questions.

Historical replay asks how the model would have behaved on recorded conditions, but it can inherit selection bias and policy-dependent data. Shadow mode tests a model in the current environment without granting action authority. Scenario testing probes plausible but sparse conditions. Sensitivity analysis reveals which assumptions can flip a decision. None is sufficient alone.

Governance mechanisms

Model registries, model cards, approval gates, champion–challenger evaluation, residual dashboards, incident reviews, and red-team exercises turn model use into an accountable process. A registry establishes which version is authorized. A residual dashboard makes correspondence visible. A challenger creates an alternative. A red team searches for invalid regimes and strategic exploitation.

Governance should be tied to consequences. A low-impact recommendation model may need light controls. A model that allocates medical care, credit, liberty, or critical infrastructure authority requires independent review, contestability, and conservative fallback.

Tuning dimensions

Model-Based Regulation is not a single fixed design. Several parameters determine how it behaves.

Model scope

A narrow model is easier to validate but may export effects beyond the boundary. A broad model captures externalities but may become too slow or uncertain. Scope should expand when omitted relationships can reverse decisions, not merely because more data are available.

Fidelity

Higher fidelity can improve local prediction while reducing interpretability, speed, and robustness. Increase fidelity only when the added layer resolves a consequential error and can be validated. Otherwise use the simpler model and preserve the correction as uncertainty or a boundary.

Decision horizon

A short horizon improves responsiveness but can sacrifice slow variables. A long horizon includes delayed consequences but compounds uncertainty. Use multiple horizons when immediate safety and long-run viability differ.

Update cadence

Fast update is useful under real drift but can chase noise. Slow update preserves stability but can institutionalize obsolete relationships. Set cadence from evidence arrival, change rate, reversibility, and hazard.

Authority level

The model can inform, recommend, gate, or act autonomously. Authority should increase only with demonstrated correspondence, monitoring quality, reversibility, and fallback. Confidence in prediction alone is insufficient; the action path and remedy also matter.

Exploration rate

Learning causal response may require controlled variation. Exploration improves identification but imposes risk and fairness concerns. Use low-cost reversible probes, staged rollout, or natural variation where direct experimentation is unacceptable.

Centralization

A central model improves consistency and can see system-wide effects. Local models use fresher context and respond faster. Federated designs need shared invariants, boundary-state exchange, and conflict resolution.

Error tolerance

Not all residuals matter equally. Set thresholds by consequence and structure, not only frequency. A rare residual near a catastrophic boundary can deserve more attention than frequent low-cost error.

Invariants to preserve

The model must not become the principal.

• Accountability remains with the regulator and institution.
• Safety constraints cannot be optimized away by the model they constrain.
• The model version and decision rationale remain reconstructable.
• Validity and uncertainty are visible where action is authorized.
• A model can be challenged, suspended, and replaced.
• The regulatory objective cannot silently shrink to the available proxy.
• Evidence used for validation is not wholly self-generated.
• Affected parties retain proportionate routes for explanation, contest, and remedy.
• Fallback remains operable when the model or data path fails.

These invariants distinguish model-based regulation from model worship.

Target outcomes

A successful implementation improves more than prediction accuracy. It should produce:

• better action selection and timing;
• fewer oscillatory or symptom-only interventions;
• earlier recognition of regime change;
• explicit separation of observation, state, model, execution, and objective error;
• safer automation through bounded authority;
• more informative post-incident learning;
• cumulative improvement in the regulator’s correspondence with the system.

A useful evaluation asks whether decisions improved relative to a credible baseline, whether harm shifted elsewhere, whether uncertainty was handled appropriately, and whether failures caused model or policy revision.

Variants

Model-Predictive Regulation

This variant compares feasible future trajectories before selecting the current action. It is useful where delay, constraint, and coupling make immediate-error correction inadequate. Model predictive control is a common mechanism, but the variant also appears in planning and operational governance without continuous numerical optimization.

Its distinctive risk is optimization against model error. The implementation must be conservative in poorly validated regions.

Adaptive Internal-Model Regulation

This variant makes model revision part of ordinary operation. Prediction residuals trigger recalibration or structural review as the system changes. It is useful under real drift but must protect against noise chasing, adversarial feedback, and unaudited updates.

The update process needs its own safety case. A model that changes rapidly can be harder to validate than the system it regulates.

Distributed Local-Model Regulation

This variant places models near local information and authority while coordinating shared boundaries and invariants. It fits grids, federated organizations, distributed computing, and regional response systems.

Its challenge is not merely technical consistency. Local units may optimize different objectives or hide externalities. Federation therefore requires model-interface contracts, conflict escalation, and shared accountability.

Tradeoffs

The archetype imposes a model-maintenance tax. Observability, validation, traceability, independent challenge, and update governance consume resources. This is justified only where the decisions are recurrent and the consequences of blind regulation are material.

Interpretability and performance may conflict. A more complex model can predict better in-distribution while making challenge and incident diagnosis harder. Ensembles reduce single-model error but diffuse accountability. Local models improve responsiveness but can fragment system understanding.

The goal is not to maximize any one attribute. It is to create a regulator whose representational complexity is proportionate to the decision and whose errors are observable and repairable.

Failure modes and repair

Decorative model

The model is used to legitimize decisions made elsewhere. Repair by recording ex ante predictions and rules, then auditing consistency between model output and action.

Map–territory substitution

Internal scores become the objective. Repair by maintaining independent outcomes, countermetrics, and boundary review.

Model overreach

A model is used outside its validated regime. Repair by making boundaries machine- and operator-visible and linking them to authority reduction.

Residual suppression

Surprises are blamed on noise or execution to protect the model. Repair by assigning independent residual review and requiring structured-error investigation.

Feedback contamination

The regulator sees only data produced under its own policy. Repair with shadow policies, holdouts, safe experiments, off-policy evaluation, and explicit selection analysis.

Update instability

The model changes in response to transient error. Repair with staged updates, champion–challenger comparison, rollback, and distinct thresholds for parameter and structural change.

Stale-model lock-in

Integration cost and institutional ownership prevent retirement. Repair with sunset criteria, challenger budgets, model-risk gates, and accountable retirement authority.

False precision

Numerical output suppresses uncertainty and local knowledge. Repair by exposing uncertainty, preserving override, and recording disagreement.

Missing fallback

Model or data failure directly removes control. Repair by designing and rehearsing a bounded fallback before deployment.

Neighbor distinctions

Requisite Variety Matching

Requisite variety asks whether the regulator has enough distinct responses for the disturbance variety it faces. Model-Based Regulation asks whether it can recognize the relevant state and select the right response. Both may be necessary. Neither implies the other.

Explicit State Modeling

Explicit State Modeling defines states and transitions. Model-Based Regulation uses state representation inside a live decision and learning loop. A workflow state machine can exist without prediction or residual-driven revision.

Phase-Space Mapping

Phase-Space Mapping reveals regimes, trajectories, attractors, and transitions. It becomes a regulatory model only when action selection and correspondence testing are attached.

Control Surface Creation

Control Surface Creation provides levers. Model-Based Regulation provides the model-guided logic for using them. A system can be well instrumented and still be badly regulated.

Mental Model Mismatch Repair

Mental Model Mismatch Repair changes what people believe about a system. Model-Based Regulation changes the operating structure by which action is chosen and revised. Human understanding may be one representation inside it, but the archetype does not depend on cognition alone.

Policy Evaluation Before Deployment

Predeployment evaluation asks whether a policy should be released. Model-Based Regulation continues after release and makes live state estimation, residuals, update, and fallback part of operation.

Layered Model Validation

Layered Model Validation governs added model complexity. It supports this archetype, especially during model construction, but does not itself regulate the external system.

System Archetype Diagnosis

System Archetype Diagnosis provides a reusable explanation of feedback behavior. It can seed the system model. Model-Based Regulation adds action authority, outcome comparison, and revision.

Sequential Policy Optimization

Sequential Policy Optimization is a formal family for improving repeated decisions. Model-Based Regulation is broader: it includes qualitative and hybrid representations, non-optimization rules, explicit validity governance, and organizational regulators.

Cross-domain examples

Industrial process

A chemical plant estimates latent reaction state from imperfect sensors. A state-space model predicts temperature and pressure response to input changes. Hard interlocks constrain action independently. Residuals are monitored by operating regime. When catalyst aging creates structured mismatch, the controller reduces authority and enters a reviewed update process.

Hospital flow

A hospital combines arrival forecasts, patient acuity, staffing, bed state, diagnostic delay, and discharge probability. The model compares staffing, diversion, and prioritization actions over several hours. It logs the predicted effect and monitors displacement to other units. A new seasonal pattern produces residuals, causing a challenger model to run in shadow mode before deployment.

Financial risk

A lender estimates default and loss under changing economic conditions, but the regulatory model also represents selection effects, feedback from credit withdrawal, subgroup error, and portfolio concentration. Authority and pricing rules tighten outside validated ranges. Model-risk review distinguishes empirical prediction from policy choices about acceptable exclusion.

Cyber defense

A security operations center maintains a threat and dependency model rather than reacting to alerts independently. It estimates whether signals form a coordinated path, predicts containment side effects, and selects reversible actions first. The model is updated after incident reconstruction, while a red team probes blind spots and attacker adaptation.

Ecological management

A watershed authority models storage, inflow uncertainty, ecosystem thresholds, and downstream demand. Release decisions are constrained by safety and ecological bounds. Forecast residuals and observed impacts update parameters, while drought regimes outside historical support trigger conservative policy and independent review.

Non-examples

A dashboard is not Model-Based Regulation merely because it visualizes state. A forecast is not enough if it has no action relationship. A digital twin is not enough if it is a demonstration artifact. A machine-learning score is not enough if its objective, validity, authority, and feedback are undefined. A detailed model is not enough if surprising outcomes cannot reduce its authority.

The archetype is present only when a decision-relevant model is embedded in the regulator, tested through outcomes, and governed as a fallible source of action.

Adoption sequence

A practical adoption sequence is:

1. choose one recurrent, consequential regulatory decision;
2. write the objective and invariants before selecting a model;
3. reconstruct the implicit model currently used;
4. identify the smallest missing distinction that causes bad decisions;
5. create a minimal explicit model and state estimate;
6. run it in advisory or shadow mode;
7. record predictions and residuals;
8. define validity, authority, update, and fallback;
9. expand authority only after correspondence evidence;
10. review boundary effects and affected-party contestability.

This sequence prevents the common mistake of beginning with a modeling technology and searching afterward for a decision it can influence.

Review note

This draft should remain a distinct candidate archetype. It directly covers good_regulator_theorem without collapsing into Requisite Variety Matching or representation-only neighbors. Editorial review should focus on canonical naming, the boundary between qualitative diagnosis and action-conditioned prediction, and whether the adaptive and distributed variants later merit promotion.

Common Mechanisms

• Bayesian State Estimation — Infers the system's hidden state and its uncertainty by recursively updating a probabilistic estimate as each noisy observation arrives.
• Causal Loop Diagram — Draws the pressure behind a hazard, the feedback loops that regenerate it, and the delays between them, so a control can be aimed at the loop rather than the symptom it displaces.
• Champion–Challenger Evaluation — Runs the incumbent regulating model against candidate challengers on the same objective and promotes a challenger only when it beats the champion by a pre-set margin.
• Digital Twin Trial — Exercises a candidate policy against a synthetic, executable replica of the system — including conditions that have never actually occurred — before it is allowed to touch the real thing.
• Forecast Backtesting — Replays a predictor against withheld history — across time, segments, and regimes — to earn or deny the right to suppress its residuals.
• Historical Replay — Reruns a candidate policy over real recorded history to see what it would have decided, then measures those counterfactual decisions against what actually happened.
• Model Predictive Control — At each step, optimizes a whole sequence of near-term actions against a forecast of the moving target — subject to hard constraints — then commits only the first action and re-optimizes when the next observation lands.
• Model Registry — The system of record for every regulating model — its lineage, assumptions, owner, approvals, and deployment status — so any model in production can be traced, re-approved, or rolled back.
• Model-Failure Red Team — An independent team whose mandate is to make the model fail — hunting the conditions under which it gives wrong answers, mapping that failure frontier, and checking the system degrades safely past it.
• Residual-Monitoring Dashboard — Continuously tracks the gap between what the model predicted and what actually happened, so drift surfaces as a signal that triggers the model's revision.
• Scenario Testing — Checks the regulator against a curated set of plausible, extreme, and boundary situations, asking of each: does it stay within safe limits and degrade gracefully?
• Sensitivity Analysis — Sweeps the model's inputs and parameters across their plausible ranges to find which ones actually move its decisions — and whether the model's added complexity earns its keep.
• Shadow-Mode Evaluation — Runs a candidate policy silently on live inputs with zero authority to act, logging what it would have done so its divergences from reality can gate promotion.
• State-Space Model — Specifies the target as a hidden state that evolves by known dynamics and is seen only through a noisy observation equation — the source model an estimator later inverts to pull the state back out.
• System-Identification Experiment — Builds the system model empirically by injecting designed inputs into the real system and fitting the observed response, its disturbances, and the assumptions the fit rests on.

Compression statement

When effective action depends on hidden state, delays, coupled responses, or changing conditions, replace model-free rule following with a governed loop that estimates state, represents causal or dynamical relationships, predicts action consequences, acts within validity and safety bounds, compares predictions with outcomes, and updates or abandons the model when residuals show it is no longer adequate.

Canonical formula: estimate state → predict action-conditioned outcomes → choose bounded action → observe result → measure residual → retain, revise, or retire model

Related Abstractions

Abstractions this archetype builds on — directly (a source ingredient) or as a related pattern. Links follow the typed catalog namespace.

Built directly on (4)

• Feedback: Outputs influence inputs.
• Good Regulator Theorem: Every effective regulator of a system must be, or must contain, a model of that system.
• Observability: Infer internal state externally.
• Representation: Model complex ideas.

Also references 14 related abstractions

• Adaptation: Systems adjust to conditions.
• Agency: A system pursues representable goals through actions whose selection is sensitive to its beliefs about its situation, via a goal-representation, world-model, and action-selection coupling.
• Calibration: Aligning a system's output to a trusted reference by measuring deviation, adjusting to reduce it, and monitoring for drift.
• Causality: Cause-effect relationships.
• Controllability: Ability to steer system.
• Foreseeing (Prediction): Predict future states.
• Homeostasis: Maintain internal stability.
• Mental Model: Internal system representation.
• Monitoring: Continuously observing a system's state to detect deviation from expected behavior and trigger a response, separating genuine signal from routine noise.
• Requisite Variety: Match environmental complexity.

▸ Show 4 more

Variants

Narrower or domain-specific specializations that share this archetype's core structure. Recognized variants are established; candidate variants are provisional.

Model-Predictive Regulation · mechanism family variant · recognized

Use an explicit dynamical model to compare feasible future trajectories and choose the current action whose predicted path best preserves the regulatory objective and constraints.

• Distinct from parent: The parent requires decision-relevant modeling but does not require online trajectory optimization; this variant makes receding-horizon prediction the action-selection method.
• Use when: Interventions have delayed, coupled, or path-dependent effects; A sufficiently credible short-horizon model and state estimate are available; Constraints must be respected while choosing among several feasible actions.
• Typical domains: process control, energy systems, logistics, clinical operations
• Common mechanisms: model predictive control, simulation rollout evaluation, bayesian state estimation, residual monitoring dashboard

Adaptive Internal-Model Regulation · temporal variant · recognized

Continuously revise the regulator's internal model when prediction residuals show that system structure, parameters, or operating conditions have changed.

• Distinct from parent: The parent permits a stable model with periodic review; this variant makes online or frequent model adaptation structurally central.
• Use when: The regulated system changes on a timescale relevant to decisions; Outcome feedback permits model mismatch to be detected; A governed update process can distinguish drift from transient noise.
• Typical domains: adaptive control, fraud detection, inventory policy, epidemiology
• Common mechanisms: forecast backtesting, champion challenger evaluation, shadow mode evaluation, model registry

Distributed Local-Model Regulation · scale variant · recognized

Give edge regulators locally useful models and reconcile their state, assumptions, and actions through a shared coordination layer rather than forcing all regulation through one global model.

• Distinct from parent: The parent is neutral about model location; this variant addresses model partitioning, federation, and consistency across multiple regulators.
• Use when: Relevant information is geographically, organizationally, or temporally distributed; Local action must be faster than centralized model updates; Cross-boundary effects require model and policy reconciliation.
• Typical domains: distributed computing, federated organizations, power grids, public health
• Common mechanisms: federated state estimation, model registry, federation protocol, scenario testing

Near names: Good Regulator Design, Model-Embedded Regulation, Internal-Model Regulation, Conant–Ashby Regulation Design, Model-Based Control, Model Predictive Control.