Aim at a target, measure how far off you are, push by an amount proportional to the miss, measure again. That circle — error → correction → re-measure — is the mathematics of purpose. Wiener named it cybernetics, from the Greek for steersman. Down the center, data flows: the setpoint and gain go in, the loop runs, the settled value comes out. The blue team builds and defends it; the red team tries to break it.
source Norbert Wiener, Cybernetics: or Control and Communication in the Animal and the Machine (1948) — archive.org/details/cyberneticsorcon00wien. Rendered, not quoted.
The plant is a pure integrator: it keeps whatever it has and adds the push, xₙ₊₁ = xₙ + uₙ. The controller looks at the miss and pushes proportionally, u = Kₙ · e, where e = r − x (setpoint minus measured).
Substitute and the loop collapses to a single recurrence on the error:
eₙ₊₁ = (1 − Kₙ) · eₙ
So the error is multiplied by |1 − Kₙ| every step. It shrinks — the loop converges — iff that factor is below 1, i.e. 0 < Kₙ < 2. Live, for the current gain:
| quantity | value |
|---|
Watt's 1788 flyball governor was one physical loop: spin too fast, the balls fly out, the valve closes, the engine slows. Wiener's move was to strip away the brass and keep the math — feedback as the general form of goal-seeking, indifferent to whether the steersman is a valve, a nerve, a thermostat, or an optimizer.
The same three lines sit under a thermostat and under a training loop: measure the loss, take a step proportional to the gradient, re-measure. Where the syllogism made deduction mechanical, this makes purpose mechanical. Each sphere is the next one's premise — and the next one is gradient descent, this loop with the plant replaced by a landscape.
The blue team's live check: re-run the loop over a sweep of gains and confirm the three laws — a stable gain converges, feedback OFF never does, and more gain (overdamped) needs fewer-or-equal steps. If red flips the sign, this badge is where it shows.
A control loop needs four numbers fed in:
| input | symbol | here |
|---|---|---|
| setpoint (target) | r | 10 |
| start value | x₀ | 0 |
| gain | Kₙ | slider ▶ |
| plant | · | integrator |
The measured value starts far from the target. The gain sets how hard the controller reacts to the miss — too little and it crawls, too much and it overshoots and rings. That is the one dial you feed the panel below.
Feedback ON: the controller sees the error and pushes. A stable gain drives the miss to zero.
Move the gain — the trajectory, the settle time, and the verdict are all simulated on the spot from xₙ₊₁ = xₙ + Kₙ(r − xₙ), never looked up.
What the machine proves, live: with feedback and any gain in 0 < Kₙ < 2 the measured value converges to the setpoint; at Kₙ = 1 it lands in a single step (deadbeat); with feedback OFF it never moves toward the target; and within the overdamped band more gain settles in fewer-or-equal steps. The current settle time is above; these laws are the output.
The blue team's witness (left) re-runs these laws live; the red team (right) flips the feedback sign to break them.
And the exact-landing is a gift of the plant choice: proportional control hits the setpoint only because an integrator has infinite DC gain. Put in a first-order lag plant and a pure P controller leaves a permanent steady-state offset — you need the integral term (the I in PID) to kill it. Nyquist and phase-margin theory subsume this whole one-liner; the loop here is the first rung, not the ladder.
"More gain always settles faster." Cut. True only up to Kₙ=1. Past it the loop overshoots and rings; at Kₙ≥2 it never settles; above 2 it diverges — the engine shows the turn.
"Feedback makes a system stable." Cut. Feedback of the wrong sign or too much magnitude is exactly what makes it unstable — the tamper below proves it by flipping one minus sign.
"A P controller removes all error." Kept, corrected. Here it does — because the plant is an integrator. On a lag plant it leaves a steady-state offset; that gap is why PI exists.
The red team's move: flip the sign of the feedback so the controller pushes toward the error instead of against it — positive feedback. The blue team's witness (window 7) is watching.
Flip the sign and every gain sends the error away from the setpoint — the loop diverges, the witness recomputes, disagrees with the known laws, and turns red. Nothing is faked; the attack is real and it is caught.