Twelve devices, thirty people, one year, and how we chose what not to build

Allocating limited hands across more projects than they could carry.

We chose what not to build by treating skilled hours as the scarce resource rather than good ideas, since a lab with thirty people and one year had far more worthwhile devices proposed to it than it could ever finish, and the honest constraint was never a shortage of things worth making.

Thirty people shared one workshop, three mills and two printers that year, and almost every decision about what to build next got made standing at a bench with the part for something else already in hand, since there was rarely a quiet moment set aside for planning that a half-finished bracket was not competing with. Saying yes to one more small device, a twelfth added to a list that already had eleven running, felt harmless in the moment it was said, and only showed its real cost weeks later in how much longer everything else had quietly started taking.

Introduction and overview

A lab is not limited by imagination, since almost any group of thirty capable people can generate far more plausible projects in an afternoon than the same thirty people could actually finish in a year, and the honest bottleneck sits instead in the small number of things that cannot be duplicated on demand, a particular machine, a particular skill, a particular set of hands free on a particular afternoon. Saying yes to a new device does not cost the time spent building that one device in isolation, it costs a slice of every other device's access to the same limited machines and the same limited people, which is why a project list that looks individually reasonable, item by item, can add up to a workload nobody in the room actually has the hours to deliver.

The dinner-party comparison

A cook working a two-burner stove for a dinner with four dishes learns quickly that starting all four at once gets every dish finished later and one of them burned rather than getting dinner on the table faster, because two burners can genuinely only heat two pans regardless of how many the cook is willing to think about simultaneously. Running two dishes through to completion before starting the remaining two, though it feels slower at the very start, gets a properly cooked meal on the table sooner in practice, since nothing sits half-attended waiting for a burner that a different dish is currently occupying. A lab with twelve devices and a handful of machines was making exactly the same trade whether or not anyone in the room had ever thought about it in cooking terms, and the lists of projects that suffered worst were the ones that tried hardest to keep every dish moving at once.

Why more projects in flight makes every one slower

Adding a new project to an already busy lab lengthens the timelines of every project it now competes with for the same limited machines and the same few experienced hands, rather than simply adding its own timeline alongside the others, an effect queueing theory captures in Little's law: the average time anything spends in a system rises directly with how much is being worked on inside that system at once, for a given rate of finishing things. A lab running four devices at a time and finishing each in three weeks was, under the same total capacity, capable of running twelve at a time only by accepting that each one now took roughly three times as long to reach the same finished state, the extra devices not appearing for free but paid for entirely in the currency of everyone's patience. Recognising this changed the actual decision being made every time a new project was proposed, which was never simply whether the project was worth doing, it was whether the lab could afford what adding it would do to every project already committed to finishing that year.

One figure worth keeping in mind

Of the roughly twenty device proposals that came up across that year, only twelve were ever started, and of those twelve, a comfortable majority were finished to a state worth showing anyone, a completion rate that came directly from having said no early and often rather than from any device being easier than the ones that got turned away. A lab that had instead said yes to all twenty would very plausibly have finished fewer than twelve of them to the same standard, since the same thirty people and the same few machines do not multiply simply because more has been asked of them.

What this changes in practice

Treating capacity as the real constraint means the question asked of a new proposal changes from whether it sounds worthwhile in isolation to what it would displace, and forces that displacement to be named out loud rather than absorbed silently as everyone's individual patience with slower progress. It also means a small number of genuinely difficult, high-value devices beat a long list of easier ones competing for the same bench time, since finishing a hard device properly usually returns more than finishing five easy ones badly, half-tested, in the gaps left over. The practical habit that followed was keeping a visible list of everything currently in progress and refusing to add to it without first agreeing what would come off, a rule simple enough to state in one sentence and consistently the hardest one in the lab to actually keep.

Where this stops being true

None of this argues for doing less work, only for doing it in a sequence that respects how few things a shared lab can genuinely move forward at once, and a lab with slack capacity, more machines than people or more people than machines needing them at any given moment, can reasonably carry more in flight without paying the same penalty. The judgement call that never goes away is guessing correctly which resource is actually the tight one this month, since a lab that assumes its bottleneck never moves ends up protecting a machine nobody was fighting over while three people quietly wait their turn for the one lathe that everyone needed.

My key error with this

In the first stretch we chose projects the way a room full of enthusiastic people naturally chooses them, which is that the most compelling advocate in the meeting won, and the result was a list nobody could have delivered and an uneven spread of effort across it. The correction was unglamorous and worked immediately: we sat down together, put every candidate through the same comparison, and scored each one on what it would cost us, whether we could actually do it with the hands and equipment we had, and what it would change if it existed. Impact had to be counted twice, once for the individual the device was for, since that is the entire point of the work, and once for whether anything about it would generalise to somebody else afterwards. What replaced the belief is that project selection is a design decision deserving the same explicit criteria as any other, and that the purpose of scoring things is not really the scores, it is that everyone in the room ends up arguing about the same axes.