Feedback, Control and the Cybernetic Loop

Prediction can be right and still be powerless

A prediction is a conditional statement about a future: given this state, these actions and these assumptions about disturbance, this trajectory is expected. It can be extraordinarily accurate and still lack any power to maintain a target. Causal influence depends on wiring, not descriptive quality.

James Clerk Maxwell's analysis of centrifugal governors made stability, rather than mere mechanism, a mathematical problem in the nineteenth century.[1] Norbert Wiener later gave the broader field its cybernetic name, joining control and communication across machines and organisms.[2] The durable idea is neither the steam engine nor the thermostat. It is circular causation: action changes the world, observation returns information about that change, and the returned information changes subsequent action.

Three causal wirings that should not be collapsed

Open-loop control chooses an action sequence from an initial state and a model, then executes it without using the realised trajectory to revise the sequence. Feedforward control measures or predicts a disturbance before it has produced error, then compensates in advance. Feedback control measures the state or output after the world has responded, compares it with a target and changes later action.

The three can coexist. A building controller may use a weather forecast as feedforward, a thermal model for planning and room sensors for feedback. A robot may predict a trajectory, estimate contact forces and replan after each camera frame. The useful question is not whether prediction or feedback is superior. It is which new information can enter the action policy, and when.

Åström and Murray define the engineering core as sensing, computation and actuation, with stability and disturbance rejection as central design concerns. They also stress the other side of the bargain: feedback can create instability, inject sensor noise and couple components into a more complex dynamical system.[4]

The counterfactual test for a real return path

A circular arrow on an architecture diagram does not establish feedback. Telemetry may be collected only for reporting. A critic may comment after the final action. An outcome may update a dashboard while the policy continues from its original plan. The causal test asks what would happen to the next action if the observed consequence were different while the target, model and prior history were held fixed.

In intervention notation, compare the action under the realised observation with the action under an imposed alternative observation: does changing y_t change u_t+1? The notation is less important than the experiment. Replay the same episode, freeze the outcome channel at its nominal value, and inspect the next action. If nothing changes, the channel is observational rather than controlling. If action changes for every fluctuation, including irrelevant noise, the loop exists but may be badly designed.

The test also separates feedback from a clock-driven update. A planner may issue a different action every minute because the schedule contains time-indexed commands. That is dynamic behaviour, yet the realised plant state has no influence. Conversely, a simple threshold rule can be genuine feedback because crossing the threshold changes action. Control sophistication is not measured by model size; it is measured by the causal use of consequence.

For a learned agent, the clean ablation is to remove or scramble the post-action observation while keeping the model, prompt, tool set and action budget constant. Compare task recovery, redundant actions and unsafe continuation. This does not by itself prove that the observation is correct or sufficient. It proves whether the claimed return path contributes to behaviour.

The return path changes the dynamics

Consider a deliberately small discrete-time plant. Let x_t be deviation from the target at step t, u_t the control action, d_t an external disturbance, a the plant's natural persistence and b the effect of the actuator:

Suppose the controller measures the deviation and applies proportional negative feedback, u_t = −k x_t. Substitution gives:

In this scalar, delay-free, unsaturated case, deviations decay when |a − bk| < 1. If the factor is close to zero, error is corrected rapidly. If it is negative, the state alternates around the target. If its magnitude exceeds one, error grows. The controller has altered the law of motion of the coupled plant-controller system.

This compact result carries the central mechanism. A prediction can estimate x_t+1. Feedback makes observed x_t a cause of u_t, which in turn changes x_t+1. The system no longer follows the plant's open-loop dynamics. It follows closed-loop dynamics whose stability must be analysed as a whole.

Worked example 1: one room, three controllers

The prediction-only schedule keeps issuing 0.733. It does not fail to execute its plan; the plan fails to absorb new evidence. A proportional controller adds heat in proportion to current error. A proportional-integral controller also accumulates persistent error, allowing it to counter a sustained bias that proportional action alone may leave.

The common PID form combines proportional response, accumulated error and error trend. Integral action is powerful because a constant disturbance keeps adding evidence until the controller produces enough opposing action. It is dangerous because the integral can continue growing when the actuator is saturated, producing overshoot after the constraint clears. The simulator uses a simple conditional anti-windup rule.

The comparison is not a universal ranking. Different gains, noise, delays or actuator limits can reverse it. Its role is to isolate one causal feature: after the door opens, measured error contains information that was absent from the original schedule. The feedback controllers can use it. The schedule cannot.

W. Ross Ashby's language of requisite variety sharpens the intuition. A regulator needs enough distinguishable responses to counter the relevant variety of disturbances.[3] More actions are not automatically better, and the theorem is not a recipe for arbitrary complexity. It is a warning that a controller whose sensors or actuators collapse materially different situations into one response cannot regulate them all.

What the curves establish, and what they leave open

Maximum deviation answers one question: how far did the room move from the target? Integrated absolute error answers another: how much discomfort accumulated over the episode? Heater effort approximates a third: what corrective resource was consumed? A controller that wins one measure can lose another. High gain may reduce peak error while spending more energy or producing oscillation after the door closes.

The target itself may be a region rather than one number. Occupants might accept 20.5°C to 21.5°C, making small corrections economically pointless. A dead band can prevent rapid switching, but it also permits bounded error. The design choice depends on consequence: a half-degree deviation in an office is unlike a half-degree deviation in a chemical reactor.

The simulation also holds measurement accuracy, actuator health and outside temperature fixed. A credible controller comparison varies these features separately. Seed model bias to test robustness, sensor delay to test stability, actuator limits to test windup and observation corruption to test false correction. A feedback claim is about a disturbance distribution and a control envelope, not one attractive trajectory.

Prediction becomes stronger when it stays inside the loop

Feedback is sometimes caricatured as reactive and prediction as intelligent. Modern control dissolves that contrast. Model predictive control observes the current state, solves a finite-horizon optimisation problem, applies only the first action and repeats at the next sampling instant. The prediction supplies foresight; repeated observation prevents that foresight from becoming a frozen commitment. Mayne et al.'s review makes this receding-horizon structure explicit and highlights its ability to handle hard constraints on states and controls.[8]

Diffusion Policy is one modern illustration. Its authors combine a generative action policy with visual conditioning and receding-horizon control, reporting results across twelve tasks from four manipulation benchmarks.[9] The benchmark does not prove that feedback alone caused every improvement. It shows how a learned predictor can be embedded in a loop that repeatedly conditions action on current visual evidence.

This distinction matters for language-model agents. Generating a ten-step plan is not the same as controlling a ten-step process. A closed-loop agent checks whether each action changed the environment as expected, updates world state and replans when the preconditions of later steps no longer hold. Planning chooses a possible path; feedback decides whether the world is still on it.

The serious negative control: a world that exactly matches the model

A causal claim earns credibility by facing a case where its preferred mechanism is unnecessary. Remove the door disturbance. Make the plant deterministic, the model exact, the initial state known, the actuator faithful and every future disturbance available at planning time. The nominal open-loop schedule now keeps the room exactly at target.

Feedback receives no new information because realised state never departs from predicted state. Its sensors, computation and actuator updates add cost without improving the trajectory. In some settings they can add measurement noise or failure points. Under perfect knowledge and perfect execution, an open-loop plan can equal the closed-loop result.

The negative control locates the real advantage. Feedback earns its cost when observation carries action-relevant information that was unavailable or unreliable at planning time. The benefit grows with disturbance, drift and model mismatch, then falls when sensing is too noisy, delayed or weakly connected to useful action.

Feedforward may still be better for disturbances that can be measured early. If a weather service predicts a cold front accurately, pre-heating can avoid the error that feedback would wait to observe. The strongest design often combines anticipation with correction, then tests which source of information actually improved the outcome.

Feedback is a bargain, not a free upgrade

Negative feedback opposes deviation only when the signal's sign, timing and scale are right. Reverse the sign and a small error produces an action that increases the error. Increase gain without accounting for delay and the controller reacts to a state that no longer exists, overshoots, sees the opposite error and overshoots again. The loop oscillates or diverges.

Sensor noise creates another trade. A controller cannot respond to measured error without also responding to measurement error. Filtering can suppress high-frequency noise but adds delay. Derivative action can anticipate trend yet is especially sensitive to noisy differences. Controller design is therefore an allocation problem among speed, noise sensitivity, robustness and control effort.

State may also be hidden. A camera sees pixels, not object pose; a service exposes latency samples, not the complete queueing state; a patient monitor observes proxies, not physiology in full. Kalman's classic formulation made recursive state estimation from noisy measurements a central companion to control.[5] Estimation does not remove uncertainty. It turns observations and a process model into a state estimate whose error must remain visible.

Actuator saturation establishes a harder boundary. A thermostat can recognise a 10°C deficit and still lack enough heater capacity to correct it. An agent can detect a policy breach and lack authority to reverse the transaction. A reviewer can notice an error after a letter has already been sent. Observed error is useful only while an admissible action can still change the consequence.

The internal model principle adds a qualified point. In specified linear regulation settings, persistent rejection of a class of reference signals or disturbances requires the controller to embody the relevant generating structure.[7] Conant and Ashby's good-regulator result likewise linked successful, simple regulation to a model under its assumptions.[6] Neither justifies the casual slogan that every capable agent must carry an explicit, human-readable world model. Recent work on embodied agents argues that the classic result does not generalise straightforwardly outside its restricted setup and that an observer's interpretation can make the model trivial or perspective-dependent.[10]

A model output is not yet a cybernetic loop

Machine-intelligence systems often use feedback language loosely. A dashboard records results, but nobody or nothing changes action. A user corrects a response, but the correction is not validated or scoped before entering memory. A model critiques its own answer, but no independent signal discriminates the original error from a more persuasive rewrite. These are information flows. They become control only when they alter a later action in a way that is causally connected to the target outcome.

The system boundary matters. A language model can propose an action, estimate consequences and explain alternatives. Identity, permission and effect limits should remain enforceable outside the model. After execution, an outcome observer must establish what changed in the external system. That evidence updates world state and determines whether the next step should continue, compensate, retry or escalate.

Five distinctions prevent category errors:

The distinction between control and learning is especially important. A thermostat can regulate for years without learning new parameters. A model can be fine-tuned from historical outcomes while remaining open-loop during each live action. Learning changes a procedure across episodes. Feedback control changes action within an episode or process because the realised state differs from the desired state.

Fast loops, slow loops and target ownership

Compound AI systems usually contain several loops with different clocks. A millisecond scheduler controls token batches. A seconds-level router controls queue and latency. A case workflow controls evidence completeness over hours or days. A release process controls model and prompt changes over weeks. Treating them as one loop hides delay, authority and ownership.

The fast loop should not optimise a target owned by the slow loop unless the relationship is explicit. A latency controller can choose among approved routes, but it should not redefine the quality threshold to make its service-level objective easier. A model-improvement loop can propose a new routing policy, but release and rollback remain separate decisions with broader evidence. Each loop needs its own setpoint, observer, actuator and escalation boundary.

Proxy feedback is especially dangerous. User clicks may be observable and frequent, yet they do not uniquely identify truth, safety or long-term value. Rewarding clicks can create a stable loop around the wrong variable. The system may become excellent at preserving the proxy while moving away from the institution's actual objective. The remedy is not simply a richer reward. It is to state which target is operationally controllable, which outcome is only measurable later and who has authority to trade them.

Nested loops also explain why a local controller can be healthy while the whole system fails. Autoscaling may hold latency while spend grows beyond budget. A verifier may reduce unsupported claims while human review queues become unmanageable. Local stability does not guarantee system-level acceptability. Cross-loop constraints and slower supervisory control are part of the design.

The site's separate article on reflection and self-correction develops the downstream implication for generated artefacts: revision becomes dependable only when error-sensitive evidence and terminal verification enter the loop.[11] The present argument supplies the more general foundation. A second pass is feedback only if its signal can distinguish and correct a property that matters.

Worked example 2: controlling an inference service

A demand predictor estimates the next hour's arrival rate and schedules capacity. Under the expected pattern, the schedule is cheaper than constant overprovisioning. At 14:07, a product event creates a burst the forecast did not contain. A provider also slows one model endpoint. Queue depth rises before p95 latency fully reflects the problem.

An open-loop scheduler keeps the original capacity plan. A feedback controller samples queue depth, request age, actual service time, timeout rate and cost. It can add bounded capacity, reduce batch size, shed low-priority work or route eligible requests to a smaller model. The predictor remains useful as feedforward. Consequence data closes the loop.

A naive controller can make the incident worse. p95 is delayed and noisy. If the controller doubles replicas whenever one interval breaches the threshold, new replicas arrive after the burst has begun to fall. The next sample looks healthy, so it removes them. Repeated gain on delayed evidence creates capacity flapping. A better design uses a faster leading signal, hysteresis, cool-down, hard cost limits and an explicit degraded mode.

The controlled variable also needs care. Optimising latency alone may route every request to the cheapest fast model and reduce task quality. Optimising average quality alone may violate an interactive service objective. The controller therefore operates over a constrained target set, not one scalar vanity metric. Hard safety, entitlement and data-residency constraints remain outside the learned router.

Command acceptance is not consequence

Suppose the orchestrator sends “increase replicas from 12 to 20” and the API times out. The effect may have failed, succeeded or partially completed. Immediate blind retry can create 28 replicas or trigger competing reconciliations. The correct control state is unknown outcome. The controller reads authoritative deployment state using the action identifier, reconciles the result, then decides whether to retry.

This is where cybernetics meets operational evidence. The loop needs an observation whose semantics match the target. A successful HTTP response may prove that a request was accepted. A deployment status may prove that capacity changed. Telemetry may show that latency improved. Business outcome may still require a different observer. Close each loop at the level of consequence it claims to control.

The consequence-closure test

A system diagram with a circular arrow can still be open in every consequential sense. The following instrument tests whether the return path can genuinely regulate the claimed target. It is design inference derived from control principles and operational failure analysis, not a published theorem.

Failure at the first gate means the system is blind. Failure at the second means it sees activity without diagnostic value. Failure at the third means it is monitoring or advising. Failure at the fourth means it may be too late or dynamically unsafe. Failure at the fifth means it can issue commands but cannot establish that correction succeeded.

How to evaluate a feedback claim

Begin with matched systems. Hold the plant, model, actuator, compute budget and initial conditions constant. Compare an open-loop or feedforward baseline with the proposed feedback design. Introduce disturbances whose onset, magnitude and recoverability are known, including a no-disturbance negative control. Random variation alone is not enough because an easier episode can make any controller look effective.

Measure more than final success. Record peak deviation, accumulated error, recovery time, overshoot, control effort, constraint violations and false correction when the initial state is already acceptable. For an agent, add redundant tool calls, repeated side effects, unnecessary escalations and cases where the outcome channel was unavailable. Stratify by disturbance type so one easy family does not conceal a dangerous one.

Then inject faults into the return path. Delay observations, corrupt one sensor, provide a stale effect receipt, saturate the actuator and make one action acknowledgement ambiguous. A robust controller should degrade within stated bounds, enter a safe mode or escalate. It should not silently convert missing evidence into success. The negative control should also test whether feedback perturbs an already correct trajectory.

Finally, ablate the causal pieces. Freeze observation, remove the corrective action, disable readback or replace the controller with matched-cost resampling. The result should reveal which component earned the improvement. The strongest evidence for feedback is differential recovery under controlled disturbance without an unacceptable rise in instability, resource use or false correction.

Executable artefact: a disturbance-response laboratory

The laboratory below runs the synthetic room model used in Figures 2 and 3. It compares a fixed prediction-only schedule, proportional feedback and PI feedback. The default positive case opens the door. The negative-control button removes disturbance and model bias so the three approaches coincide.

The artefact assumes a first-order room, fixed outside temperature, a bounded heater, exact temperature measurement apart from the chosen delay and a temporary additive door loss. Its output is illustrative. It reveals the logic that prose hides: the fixed schedule cannot react to disturbance; feedback trades error reduction for control effort; delay can turn aggressive correction into oscillation.

function stepRoom(temp, outside, heater, doorLoss) {
  const alpha = 0.02;   // heat leakage per minute
  const beta = 0.30;    // heater contribution per minute
  return temp + alpha * (outside - temp) + beta * heater - doorLoss;
}

// Prediction-only: heater is fixed from the nominal model.
// Feedback: current or delayed measured error changes heater.
// PI: accumulated error adds correction, with anti-windup at limits.

Compact glossary

Plant

The process whose state an action changes, such as a room, robot or inference service.

Setpoint

The desired value or admissible region the controller tries to maintain.

Disturbance

An influence on the plant that is not the controller's chosen action.

Gain

How strongly the controller changes action in response to measured error.

State estimate

A bounded inference about hidden plant state from observations and a process model.

Closed loop

A causal cycle in which realised consequence can change later action.