Monday Myth: The Engine That Fought Its Governor

The machinery above the machinery

In 1788, a Boulton and Watt steam engine began driving metal polishing machines at the Soho Manufactory in Birmingham. Known as the Lap Engine, it served the works for seventy years and survives today in the Science Museum. Its endurance gives it the quiet authority of machinery that earned its place through use. Among its fittings, a pair of revolving weights performed a task quite different from polishing metal. They helped regulate the speed at which the work happened. [1]

Two balls revolved around a vertical spindle, suspended on jointed arms. As speed increased, they moved outwards and upwards, shifting a linkage that reduced the opening of the steam throttle. As speed fell, they descended and allowed the throttle to open further. The engine’s motion supplied the information used to adjust its own supply of power. No observer needed to translate a reading into an instruction. The measurement and the intervention travelled through the same mechanism. [2]

Consider the problem this addressed. A rotating engine driving workshop machinery encounters changing resistance as work changes. Maintaining useful operation involves more than supplying enough power for the heaviest demand. The supply must also accommodate lighter demand without allowing speed to rise excessively. The governor gave the engine a way to respond to what its connected machinery actually required. Its usefulness lay in the relationship between output, resistance and correction. [2]

Watching such a mechanism, one can easily imagine an improvement. Make it detect smaller deviations. Give it more authority over the valve. Allow it to correct more emphatically. If a governor brings speed under control, a more sensitive governor ought to bring it under better control. The proposition sounds particularly convincing when considered at a single instant, with the engine running too fast and the throttle waiting to close. But an engine does not operate at a single instant.

A correction with a future

In 1868, James Clerk Maxwell published On Governors, examining the motion of regulating mechanisms mathematically. His analysis distinguished disturbances that decay from oscillations that persist or grow. An apparatus could oppose a deviation and still fail to settle. Understanding its equilibrium did not establish the stability of its movement towards that equilibrium. Maxwell also examined resistance to oscillatory motion, including effects produced by viscous liquids. The regulating mechanism needed investigation as part of a moving system, rather than approval as an independently sensible attachment. [3]

That distinction survives in modern control engineering. A feedback loop has dynamics of its own, and delays constrain the performance that a controller can achieve. An instruction reaches a process whose response unfolds over time. Increasing the strength of correction without accounting for that response can undermine stability. Faster measurement helps reveal movement, but does not make the controlled process respond instantaneously. The timing and magnitude of action still matter. [4]

The unsettling possibility follows directly. Some of the movement seen in a regulated system may come from the regulator. A deviation provokes action, action produces an effect, and the developing effect provokes another action. A sequence of individually defensible corrections can form an indefensible pattern. Watching only the latest error conceals the contribution of earlier attempts to remove it.

An engine offers a relatively honest setting for this discovery. Its governor cannot commission a report demonstrating how promptly it reacted. The shaft continues turning, and the resulting motion remains available for inspection. The mechanism receives no credit for the number of corrections it performs.

The organisation that never has time to respond

Now imagine a software organisation whose delivery has slowed. In January, management moves experienced engineers towards a struggling product. The change creates immediate disruption while people learn unfamiliar constraints. February’s figures show little improvement, so a second intervention changes the team boundaries. By March, work started under the first arrangement reaches integration, but the people responsible now belong to different groups. The next review finds growing coordination costs and introduces another structure to reduce them.

Each decision can acquire a reasonable explanation. The product needed support. The boundaries looked wrong. Integration exposed dependencies. What remains harder to explain concerns the sequence. February’s measurement cannot contain the full consequences of January’s decision, and March’s measurement reflects several arrangements at once. The organisation keeps responding to a state partly produced by interventions whose effects have not yet finished arriving.

This resembles a delayed feedback problem, although people introduce complications that a steam valve never faced. A team remembers abandoned commitments, anticipates the next redirection and adjusts its willingness to invest. Repeated changes can therefore alter both current work and expectations about future work. The organisation does not merely wait longer for a response. Its responses themselves change as people learn the behaviour of the system directing them.

Suppose engineers begin postponing an architectural improvement because the next planning review may transfer ownership again. The delay then appears in delivery figures as insufficient initiative. Management intervenes to increase accountability, adding reviews that make uninterrupted work scarcer. What began as an attempt to accelerate progress now supplies further reasons to hesitate. The control mechanism has entered the behaviour it measures, while its records continue treating that behaviour as an independent condition requiring correction.

When the instrument becomes the assignment

Better visibility does not automatically resolve this difficulty. A dashboard can refresh every hour while the underlying outcome takes months to develop. Consider a product change intended to improve customer retention. Deployment provides early evidence that the change reached production. Usage offers a different signal. Retention requires customers to reach the point at which staying or leaving becomes meaningful. Displaying all three measurements in real time does not give them the same causal age.

Under pressure for immediate reassurance, an organisation may begin steering towards the measure that responds fastest. Completed tickets move before customer behaviour does, so completed tickets attract attention. The count offers something the actual objective cannot yet supply: a visible response before the next review. A measure selected to help observe progress acquires a new role in directing it.

The steam-engine comparison now reaches a revealing limit. A rotating shaft cannot divide one revolution into three reportable achievements, exclude difficult revolutions from its quarterly figures or transfer them to another department. An organisation has considerably more freedom. It can change the boundaries of a task, the definition of completion and the selection of work that enters the measurement, while recording every resulting number accurately.

Charles Goodhart examined a related difficulty in monetary policy: an observed statistical relationship may weaken once authorities use it for control. The relationship that made an indicator useful arose under particular conditions. Applying pressure through that indicator changes those conditions. [5] Completed tickets may once have accompanied useful delivery because teams divided work according to the needs of delivery. When the count becomes an object of pressure, the division of work acquires another purpose. The historical association no longer carries the assurance it appeared to offer.

This need not involve deception. A conscientious team can favour work that produces an observable response before the next review, while postponing work whose value takes longer to emerge. Difficult integration, preventative maintenance and unresolved customer problems remain beyond the instrument’s field of view. The numbers describe the selected activity faithfully, which makes their growing distance from the intended outcome harder to recognise.

There is an economic consequence beyond misleading reports. Frequent correction already makes it difficult to distinguish the effects of successive decisions. Now the relationship between the indicator and the outcome changes during the investigation as well. Later results mingle the consequences of intervention with adaptation to the measurement. The cost includes abandoned work and relearning, but also the knowledge that sustained observation might have produced. Another intervention looks attractive precisely because the previous ones have made confident diagnosis harder.

Several hands on the valve

The problem grows more intricate when several functions direct the same productive system. Product wants a faster release, finance wants a lower cost, security wants a smaller exposure and sales wants an immediate commitment. Each observes a genuine constraint through a different instrument. Each may judge its own intervention against a local result, although the engineering teams must absorb the combined movement.

An urgent customer request displaces planned work. The resulting delay triggers a delivery escalation. Extra coordination consumes capacity, and the cost review then questions why more people have produced so little additional output. Nothing in this imagined sequence requires a careless participant. The instability emerges through interactions that none of the individual reviews contains. The same engineers occupy several control loops, with different objectives and different expectations about response time.

Automation can intensify this arrangement. An AI system might detect deviations sooner and propose adjustments before the next meeting. That could help a controller that understands the process it directs, or reduce the effort required to issue another poorly timed correction. Faster recommendations do not resolve contradictory objectives or reveal which earlier instruction still has consequences travelling through the work.

Yet the alternative cannot simply consist of waiting. A failed deployment may demand immediate action, while an organisational redesign needs a longer period of observation. Treating both as reasons for patience would confuse very different processes. Modern control engineering also recognises systems that require sufficiently rapid control to remain stable. The relevant question concerns which dynamics govern the situation, rather than whether fast or slow intervention carries greater moral authority. [4]

Watching the valve

From inside such an organisation, the evidence can look reassuring. Reviews happen promptly, deviations receive attention and the reported figures improve. Yet each observation carries a question the reporting machinery cannot answer by itself. How much of the movement follows from changing conditions, how much comes from earlier corrections still working their way through the system, and how much reflects work reorganised to produce a more favourable reading? The instrument continues supplying numbers, but the relationship that once made those numbers informative may have changed under the pressure applied through them.

Success against the metric can then legitimise the pressure that weakened its usefulness. The dashboard improves, the intervention receives credit and the organisation strengthens the mechanism through which it now directs work. Consequences that take longer to emerge remain available for the next explanation and the next corrective programme. Those programmes can each claim support from the figures, even as their succession makes the underlying process harder to understand. The organisation acquires an increasingly detailed record of its interventions without acquiring an equally clear account of what they have done.

The Lap Engine spent decades turning workshop machinery. Its governor earned its place through the steadiness of the work it helped sustain. Imagine standing beside an engine whose valve keeps opening and closing while the shaft repeatedly gains and loses speed. Beside it, a counter records every corrective movement, and its steadily rising total offers reassuring evidence that the mechanism remains active. Beyond the engine, the polishing machines turn unevenly. The counter has missed nothing, although the metal still waits for steady motion.

Sources

[1] Science Museum Group, Rotative steam engine by Boulton and Watt, 1788. https://collection.sciencemuseumgroup.org.uk/objects/co50948/rotative-steam-engine-by-boulton-and-watt-1788-beam-engine-steam-engine

[2] Karl J. Åström and Richard M. Murray, Feedback Systems: An Introduction for Scientists and Engineers, introductory discussion of centrifugal governing. https://www.cds.caltech.edu/~murray/books/AM08/pdf/am08-complete_08Jan08.pdf

[3] James Clerk Maxwell, On Governors, 1868. https://en.wikisource.org/wiki/On_Governors

[4] Karl J. Åström and Richard M. Murray, Feedback Systems, Chapter 11, Fundamental Limitations. https://www.cds.caltech.edu/~murray/books/AM08/pdf/am08-loopsyn_28Sep12.pdf

[5] Charles A. E. Goodhart, The ECB and the Conduct of Monetary Policy: Goodhart’s Law and Lessons from the Euro Area, 2006. https://onlinelibrary.wiley.com/doi/10.1111/j.1468-5965.2006.00661.x

The organisational scenarios and extensions of the engineering analogy are the author’s analysis, rather than historical claims or findings attributed to the cited sources.