Every feedback loop runs on the same four operations
You already use feedback loops. You just do not name the parts, which means you cannot diagnose why some loops work and others fail.
Every feedback loop — whether it runs inside a thermostat, a military fighter cockpit, a manufacturing line, or your own nervous system — executes exactly four operations in sequence: act, observe, evaluate, adjust. The labels change depending on who formalized the model. The structure does not. Once you see the four parts, you start noticing which part is missing every time a system fails to learn.
This lesson names the four parts, shows you the same structure across four independent formalizations, and gives you a diagnostic frame you can apply to any loop you encounter.
Part 1: Act — generate output that can be measured
The loop begins with action. Not intention, not planning, not thinking about what you might do — action. Something moves into the world and produces a result.
This is where Norbert Wiener started when he laid the theoretical foundations for cybernetics in 1948. In Cybernetics: Or Control and Communication in the Animal and the Machine, Wiener defined the core problem: any system that needs to maintain a goal state must first produce output that interacts with its environment. A ship's steering engine turns the rudder. A thermostat activates the heater. A person speaks a sentence in a negotiation. Without action, there is nothing for the rest of the loop to operate on.
The action does not need to be correct. It does not need to be optimal. It needs to exist. The entire value of the feedback loop is that it can correct for imperfect action — but only if the action happens. This is why perfectionism is the enemy of learning. A person who refuses to act until they are certain they are right has disabled the mechanism that would make them right over time.
W. Edwards Deming called this step Do in his cycle. John Boyd called it Act. Miller, Galanter, and Pribram called it Operate. In neural network training, it is the forward pass — the moment data flows through the network and produces a prediction. Every formalization starts the same way: put something into the world.
Part 2: Observe — capture the result without filtering it
Once you have acted, the loop requires you to observe what actually happened. Not what you expected to happen. Not what you hoped would happen. What happened.
This is harder than it sounds. Boyd placed enormous weight on the Observe step of his OODA loop (Observe, Orient, Decide, Act), developed in the 1950s and 1960s while studying why American F-86 pilots dominated Soviet MiG-15s in the Korean War. Boyd concluded that the F-86's superior visibility — its bubble canopy versus the MiG's restricted cockpit — gave American pilots a faster observation cycle. They saw what was happening sooner. Not better planes. Better observation.
Observation means measurement. Donella Meadows, in Thinking in Systems (2008), defined a feedback loop as "a closed chain of causal connections from a stock, through a set of decisions or rules or physical laws or actions that are dependent on the level of the stock, and back again through a flow to change the stock." The critical word is closed. If you act but do not observe the result, the chain is open. No information flows back. The system cannot learn.
In neural network training, observation is the output of the forward pass — the prediction the network actually made, compared to the ground truth. The network does not get to decide what its prediction was. The numbers are the numbers. The same discipline applies to you: observation is data collection, not narration.
The failure here is selective observation — seeing what confirms your expectation and filtering out what contradicts it. You launch a feature, three people praise it in Slack, and you observe "success." Meanwhile, the usage metrics show 12% adoption and falling. You observed the signal you wanted, not the signal that existed.
Part 3: Evaluate — compare the result to a standard
Observation alone is insufficient. Data without interpretation is noise. The third part of the loop is evaluation: comparing what you observed against some standard, target, or expectation.
This is where the TOTE model — Test, Operate, Test, Exit — makes its contribution. George Miller, Eugene Galanter, and Karl Pribram introduced TOTE in Plans and the Structure of Behavior (1960) to bridge the gap between behaviorist stimulus-response models and cognitive processes. Their key insight was that behavior is governed by comparison: the system continuously tests the current state against a desired state. If there is a discrepancy (what they called an "incongruity"), the system operates to reduce it. If there is no discrepancy, the system exits the loop.
TOTE made evaluation explicit. The "Test" step is a comparison function: Does the current state match the goal state? Yes or no? This is not a subjective feeling about how things are going. It is a binary or gradient measurement against a defined standard.
Boyd embedded the same logic in his Orient step — which he called the schwerpunkt (center of gravity) of the entire OODA loop. Orientation is where raw observation gets filtered through your mental models, cultural context, prior experience, and analytical frameworks to produce meaning. Every feedback path in Boyd's full diagram flows through Orient. Without it, you have data but no understanding of what the data means.
Deming called this step Check in the version that Japanese industry adopted, or Study in his preferred PDSA variant. He specifically objected to the word "check" because it implied a binary pass/fail, when what he wanted was deeper investigation — studying the results to understand why the outcome matched or diverged from the prediction. The distinction matters. Checking tells you that something went wrong. Studying tells you what went wrong and generates a hypothesis for fixing it.
In neural network training, evaluation is the loss calculation — the mathematical function that quantifies how far the network's prediction landed from the correct answer. Cross-entropy loss, mean squared error, whatever metric fits the task. The loss function does not just tell you "wrong." It tells you how wrong, in what direction, at every parameter in the network. That precision is what makes the next step possible.
Evaluation requires a standard. If you have no target, no definition of success, no metric that tells you whether the action worked, you cannot evaluate. This is the most common point of failure in personal feedback loops. People act, they vaguely observe, and then they skip evaluation entirely because they never defined what "working" looks like. The loop collapses into activity without learning.
Part 4: Adjust — change the next action based on the evaluation
The final part closes the loop: take what you learned from the evaluation and use it to modify the next action. Without adjustment, the entire cycle is academic. You acted, you observed, you evaluated — and then you do the exact same thing again. That is not a feedback loop. That is repetition.
Deming called this step Act (confusingly, since Boyd uses "Act" for the first step). In Deming's PDCA, the fourth step is where you standardize what worked or revise the plan based on what you learned. Shewhart, whose 1920s work at Bell Labs laid the statistical foundation that Deming later popularized in Japan, emphasized that this step must feed directly back into the next cycle's Plan step. The adjustment is not a conclusion — it is the starting condition for the next iteration.
In the TOTE model, adjustment is implicit in the loop structure: if the second Test reveals continued incongruity, the system returns to Operate with modified parameters. Miller, Galanter, and Pribram described this as hierarchical — a single TOTE unit can nest inside another, so the adjustment at one level becomes the action at a subordinate level. A carpenter testing whether a nail is flush (Test) strikes it again (Operate) with a different angle. The angle change is the adjustment.
Boyd's formulation places adjustment in the transition from Decide to Act — but with a critical nuance. Boyd drew a line directly from Orient to Act that bypasses Decide entirely, labeled "implicit guidance and control." When you have internalized the loop deeply enough, the adjustment happens without conscious deliberation. An expert pilot does not evaluate the airspeed indicator, consciously decide to pull back on the stick, and then act. The observation-evaluation-adjustment sequence has been compressed into a single reflex. The four parts still operate. They just operate below conscious awareness.
In neural network training, adjustment is backpropagation and weight update — the algorithm propagates the loss backward through every layer, computing the gradient of the loss with respect to each weight, and then nudges each weight in the direction that reduces the loss. The adjustment is mathematically precise: each parameter moves by an amount proportional to its contribution to the error, scaled by a learning rate. This is what makes neural networks learn. Not the forward pass. Not the loss calculation. The weight update — the adjustment — is where learning physically occurs.
The same is true for you. Learning does not happen when you observe a bad outcome. Learning happens when you change your behavior in response to a bad outcome. If your retrospective identifies that you underestimated the project timeline by 40%, learning has not occurred. Learning occurs when you multiply your next estimate by 1.4.
The same structure, four different names
Here is the mapping, so you can see the convergence:
| Part | Generic | Deming (PDCA) | Boyd (OODA) | Miller et al. (TOTE) | Neural Networks | | ---- | -------- | --------------- | ----------- | -------------------- | ------------------------------- | | 1 | Act | Do | Act | Operate | Forward pass | | 2 | Observe | Check (partial) | Observe | Test (second) | Output / prediction | | 3 | Evaluate | Study (PDSA) | Orient | Test (comparison) | Loss calculation | | 4 | Adjust | Act | Decide | Operate (modified) | Backpropagation + weight update |
The labels differ because each model was built for a different context. Deming was optimizing manufacturing quality. Boyd was optimizing fighter combat. Miller, Galanter, and Pribram were modeling cognitive behavior. Neural network researchers were training statistical models. But the underlying structure is identical: produce output, capture result, compare to standard, modify approach.
Wiener saw this in 1948. The word he chose — cybernetics, from the Greek kybernetes, meaning steersman — captures it perfectly. A steersman watches the heading, compares it to the destination, and corrects the rudder. Act, observe, evaluate, adjust. The ship reaches port not because the steersman set the perfect initial course, but because the loop ran continuously.
Why loops fail: missing parts
Now that you have the four-part model, you can diagnose any broken loop by asking which part is missing.
Missing action: Analysis paralysis. You plan endlessly, evaluate hypothetically, but never produce output that the world can respond to. Common in perfectionists and over-planners.
Missing observation: You act repeatedly without measuring outcomes. The entrepreneur who launches feature after feature without looking at usage data. The manager who gives the same feedback style to every report without noticing that it works for some and fails for others.
Missing evaluation: You collect data but never compare it to a standard. Dashboards full of metrics that nobody interprets. Weekly reports that get filed but never discussed. You have observation without meaning.
Missing adjustment: You evaluate accurately — you know what is wrong — but you do not change your behavior. This is the most painful failure mode because awareness without change is just sophisticated suffering. You can articulate exactly why your meetings are unproductive, but next week's meeting has the same agenda.
Each missing part produces a different symptom. When you learn to diagnose which part is absent, you stop saying "this isn't working" and start saying "the observation step is broken" or "we have no evaluation standard." That precision turns a vague problem into a specific repair.
The loop is the unit of learning
The feedback loop with four parts is not one model among many. It is the structural minimum required for any system — biological, mechanical, cognitive, or social — to learn from its own behavior. Remove any of the four parts and learning stops.
The previous lesson established that action generates information. This lesson names the complete circuit that transforms that information into improved action. The next lesson examines what happens when you compress the cycle time — when the loop runs faster.
For now, take one process you are currently running and name the four parts. If you cannot, you have found the part that is missing. Fix that part first. The loop will do the rest.
Sources:
- Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the Machine. MIT Press.
- Boyd, J. (1976). Destruction and Creation. U.S. Army Command and General Staff College.
- Miller, G. A., Galanter, E., & Pribram, K. H. (1960). Plans and the Structure of Behavior. Holt, Rinehart and Winston.
- Deming, W. E. (1986). Out of the Crisis. MIT Press.
- Meadows, D. H. (2008). Thinking in Systems: A Primer. Chelsea Green Publishing.
- Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). "Learning representations by back-propagating errors." Nature, 323(6088), 533-536.
- Shewhart, W. A. (1939). Statistical Method from the Viewpoint of Quality Control. Graduate School of the Department of Agriculture.