True Once

Almost everything written down in your systems was true once. The page describing how the tokens work was right the day someone wrote it. The model you brought in to do the work was the best option the month you chose it, the brand guide still shows the company as it looked a year ago, and the strategy everyone cites has come apart from the strategy you're running so gradually that no one can name the day it happened. Each was accurate when it was made. Then the work moved on, the words stayed where they were, and at no particular moment true once turned into wrong now.

Every working system keeps two records of itself. There is the record that holds: the schema, the token file, the registry, the field that reads version: 2.7.4 and fails a check the moment it lies. And there is the summary: the page that explains the schema, the story about the color, the paragraph a person actually reads before they act. The first stays true because something downstream depends on the exact value. The second drifts, on its own schedule, because nothing downstream depends on the words. Nothing breaks when they go stale. The cost is real and the consequence is real, but neither is loud, so maintaining the summary always loses to work that has a deadline.

Now look at what you do about it. A serious team runs a thorough process for new work: code review on every pull request, design review at handoff, a retro every sprint, a business review every quarter, planning every year. Stack them up and the team is well-policed at the moment work happens. Every one of those checks faces the same direction. Each asks whether the thing being introduced right now is correct. Not one asks whether the thing introduced eight months ago is still true. You are guarded, thoroughly, against the new mistake, and unguarded against the old one still sitting on the floor.

Here is the uncomfortable part. Your picture of your own systems is getting less accurate, slowly, in a way none of your reviews will catch, because none of them is looking back. You already know this. You have opened an internal document, started reading, and three paragraphs in realized it was two versions stale. You have caught a contributing guide still describing a token workflow the team replaced months ago, and quietly corrected it. You felt the gap and fixed your one instance of it, and the structural fact went unnamed: nothing was ever going to make you look.

The same gap opens in rooms that have nothing to do with documentation.

It opens in the systems you have brought in to plan and act on your team's behalf. They execute well and learn nothing. The session ends, the work ships, and the pattern that would have made the next run faster evaporates, because no one runs the step that folds it back into the tools. You have bought capability and you are watching it start from the same baseline every morning. The model is current. Your use of it never compounds, and nothing breaks to tell you so.

Your brand keeps two records too. A design system can hold the hex, the type scale, the spacing, the corner radius. About half of what makes a visual language yours was never in the file to hold: the argument for this typeface and not another, the restraint that reads as confidence instead of timidity, the difference between a brand that has a quality and one that performs it. None of that reduces to a value, and a value is all a token can store. The config stays current because a pipeline reads it. The judgment that chose every value lives in one designer's head, written down nowhere, and people leave.

Your organization keeps them too. There is the strategy on the page, and there is the strategy you are actually running, the one that formed in the decisions nobody reviewed as strategy: who got hired, what got quietly killed, whose budget grew, which idea died for lack of a sponsor. By the time the stated strategy and the lived one have come fully apart, every quarterly review has signed off on the one that was true once. The signals that would have caught the drift early are faint by design, and faint signals trip no alarms. The options narrow first and announce themselves later.

Four systems that belong on four different shelves, and the same failure in each: the record held, the summary drifted, and the staleness made no sound. The standard response is to ask for more discipline. Add docs to the definition of done. Enforce it in review. Schedule a quarterly cleanup. It does not work, and not because people are careless. Telling someone to keep the summaries current is asking them to run faster to cancel out a property of the system, a request people accept in good faith and cannot satisfy, so it returns unmet every few months.

What works is small, structural, and aimed in the opposite direction from every review you already run. Take the short list of surfaces that are supposed to stay true: the source-of-truth files, and the summaries that paraphrase them for everyone who will not read the source directly. Compare them. Where they disagree, the record wins, and the disagreement becomes something you track instead of something you stumble on a year later. Run it once a year and you get a backlog no one wants to touch. Run it at the close of each piece of work, bounded to that short list, and it stays small enough to actually happen.

We went looking for this in our own work, on the theory that a studio whose own infrastructure has the problem has no business selling the fix. Eleven sessions, sixty units of capability, one shared instrument, named evidence required for every claim. We expected to catch the work out of date. The work was mostly current. What had rotted were the summaries: the page that explains a service to a client, the document that paraphrases the work for the website, the internal map people steer by. Six of eleven downstream surfaces lagged the underlying record by weeks. The fix we are building into the close of every session does one mechanical thing. It holds each summary up against the record and flags the gap before anyone reads the summary and acts on it. No one has to remember to run it.

The bill for forward-only review always comes due downstream, and it is rarely paid by the person who let the page go stale. It is paid by a designer in her first week, reading a confident page with no way to know it had stopped being true, building something good on a description of a brand that had already moved. She will learn what everyone senior already learned: don't trust the pages, ask a person. That is the workaround a whole industry runs on, and it is also a quiet confession that the map is known to be wrong and no one has scheduled the walk to correct it. Somewhere in your systems a paragraph is going stale as you read this, and nothing is breaking, and no one has gone to look.