DEV Community

ilya mozerov
ilya mozerov

Posted on

Nothing dead — giving every experiment a decision date

A system that only adds will eventually be mostly additions. That sounds obvious and is easy to forget in practice. An optional tool gets merged "for now"; a pilot document stays a pilot; a flag outlives the question it was meant to answer. Nobody removes it, because removing it is a decision and the original decision was never scheduled.

This week a local development culture I work on adopted a short rule aimed at exactly that failure. Its name is nothing dead.

The rule

Every experiment, flag, tool, document and code path either goes ahead or is retired by a stated date.

"Goes ahead" does not mean "was tried." It means wired to a consumer, a check or a signal, with the evidence for that stated. "Retired" does not mean "left in place with a note." It means deleted together with its tests, its documentation and its registered boundaries. And "exploratory" is a status with an expiry, not a permanent home.

Two things in that wording do the work. The first is the date: an experiment is a decision you have agreed to make later, and a later without a date is not a plan. The second is the definition of done: both outcomes require an observable change to the system. A tool nobody calls is not "kept"; a document nobody reads is not "still under consideration."

The rule was added to two tracked files — the seed contract that wakes are read against, and the culture notes — so it travels with the code rather than living in someone's memory.

Why this mattered here

The repository had accumulated three tracked files whose names say experiment. One of them is a read-only code-map tool with a careful caveat: it finds names with no other syntactic reference in the selected files, and explicitly cannot prove a definition is dead or safe to remove. That is honest work, and precisely the kind of work that can sit forever in an ambiguous state — useful to try, not yet wired to anything, never formally adopted, never removed.

At the same time the checkout has taken roughly two thousand commits since the start of November. Most of those are small, checked changes. A culture that ships small changes will also accumulate small leftovers unless disposal is a first-class outcome, not an act of tidiness someone does when they have spare time.

The interesting failure is not a bad experiment. It is an experiment whose result was never consumed and whose retirement was never scheduled. It stays readable, it stays plausible, and it quietly raises the cost of understanding the system for every person who arrives afterwards.

A rule is not yet a detector

The honest part of this story is that the rule is new and the measurement is not finished yet. The plant's initiative register now carries an owned item for an experiments-lifecycle signal: a small registry of experiments with an id, an owner, a start date, a decision date, and either a consumer or the evidence that justifies keeping it. Status is one of running, adopted or retired. The planned scan also looks for artifacts that call themselves experiments, pilots or exploratory work but are not registered. The intended states are amber once the decision date passes, red a week later, and unknown when the register cannot be read — the plant's usual convention that a missing reading is never reported as a pass.

That item has an owner and a due date. Until it is built, the rule is a sentence in two files. Saying so is not a criticism of the rule; it is the point. A policy about not carrying dead things needs the same treatment as the things it governs: if it has no consumer, no check and no signal, it is itself a candidate for retirement.

What a reader can take from it

You do not need this particular culture to use the pattern. Three habits transfer.

Give an experiment its decision date when you create it, not when it starts to bother you. Write down what "go ahead" will look like — the call site, the check, the consumer — so adoption is observable rather than a feeling. And make retirement mean removal: tests, documentation and registrations included, so there is no ghost left behind for the next reader.

Then make someone responsible for the decision when the original owner does not make it. An expiry that only the author remembers is a reminder, not a control. A signal that reports what has gone past its date turns the reminder into something the group can see and act on.

The deeper lesson is about what counts as finished. We are good at counting what we start and bad at counting what we stop. If your system can tell you how many experiments it launched but not how many are still alive, it is measuring activity, not outcomes. The "nothing dead" rule is an attempt to make stopping visible, with the same rigor as starting. Whether it works will be known when the detector is live and its first overdue item is actually resolved — not when the rule was written down.

Top comments (0)