A prediction ·
The Fence That Worked
Every failure writes itself into the record. A prevented failure writes nothing. We are now handing the pruning of civilization’s rules to minds that learn only from the record — and the better a fence has worked, the less evidence survives that it does anything at all.
Tuesday, 9:14 a.m.
The optimization report is waiting when you sit down. Your agent has read every workflow in the company — something no employee has ever done — and found eleven steps serving no function it can detect. Step seven of the deploy checklist: wait ten minutes. A second signature on invoices under a thousand dollars. A three-day hold before any new vendor is paid. The report is lucid, costed, and correct about the price of each. Nobody in the meeting can say what step seven is for. The vote to remove it takes ninety seconds.
In 1929, G. K. Chesterton described a fence built across a road for no reason anyone present could name. The modern reformer, he wrote, goes gaily up to it and says, “I don’t see the use of this; let us clear it away.” And he gave the answer that has carried his name ever since:
“If you don’t see the use of it, I certainly won’t let you clear it away. Go away and think. Then, when you can come back and tell me that you do see the use of it, I may allow you to destroy it.”
For a century that advice was easy to keep, because tearing out a fence took longer than wondering about it. What Chesterton could not have imagined is the reformer we just built: one that reads everything ever written about the fence, in under a minute, and still cannot see what it is for. Not because it reads poorly. Because nothing was written.
Prevention is the job that leaves no evidence
Underneath Chesterton’s manners sits a principle harder than etiquette: the record of the world contains only what happened. Failure is a prolific author. Every incident writes timestamps, tickets, invoices, case files, scars. Prevention is silent. The embezzlement the second signature made impossible never occurred, so it was never investigated, so it was never written down. A safeguard’s costs are all recorded — minutes, salaries, delays, the annual complaint about step seven — while its entire benefit takes the form of a particular kind of bad day continuing not to arrive.
Figure 01
Two authors, one archive
During the Second World War, the statistician Abraham Wald, working with the Statistical Research Group at Columbia, showed that studying damage only on the bombers that returned was misleading: the aircraft hit in the truly fatal places were the ones that never came back. His analysis implied that protection belonged where the survivors showed no damage. The planes that mattered most were missing from the data, and a working fence has the same geometry. Its whole contribution is a missing hole.
The reader changed
Institutions have always leaned on a guardian the record never shows: memory walking around inside a person. The veteran who says, leave that one — we learned it the hard way in ’09. Her protection was never in the files. It was a private copy of the bad year, held in a body, and it retires when the body does. Every organization older than a generation runs on fences whose reasons now exist nowhere but in people, and in fewer people every year.
The new reader is thorough in exactly the way the veteran was not, and absent in exactly the way she was present. An AI agent is a perfect student of everything the institution wrote down — and prevention is precisely the work whose success is not written down. I have made this point about people and companies before, in another vocabulary: expectation is trained by the record, only by the record, and there is no other road in. A fence that works perfectly sends nothing down that road. No mind that learns from the archive alone — human or machine — will ever expect the thing the fence is holding back.
And one more thing changed, quietly, with the reader: scale. When a single manager tears out a single fence, the lesson, when it comes, is local. When the same model advises a hundred thousand companies in the same quarter, the fences come out everywhere at once — and the bad years, when they come, arrive as a class.
Figure 02
The life of a working fence
- The bad year. Something goes wrong. It costs dearly, and it writes itself into the record in full.
- The fence. Someone builds a rule so it cannot happen again. The rule works.
- The silence. Years of nothing. The people who remember the bad year retire, one by one.
- The audit. The record now shows a step with visible cost and no visible function.
- The removal. An optimizer trained on the record deletes the step, holding the strongest case anyone has ever assembled against it.
Month eight, a Friday, 6:40 p.m.
Something happens at your company that has not happened in thirty-one years. In the postmortem, someone finally finds it: a memo from 1994, scanned sideways, written by a man who retired before the person now reading it was hired. The last paragraph says the new procedure should prevent recurrence. It did.
What I think happens next
The Great Smoothing.
Over the next three years, agents will remove more institutional friction than the previous fifty years added. Most of it deserves to go — genuinely dead rules, ritual without referent — and the gains will be real and immediate. Which is exactly why the removals will be approved a little faster every quarter.
Solved problems come back.
A new class of incident appears: the recurrence of something that last happened before anyone’s tenure. Its signature is a postmortem that ends at an old memo. And because the same few models will have pruned the same fences across thousands of firms at once, these incidents will arrive in cohorts, not one at a time.
Deletion gets the paperwork that creation never had.
For a century, adding a rule required review while removing one was celebrated as reform. That inverts. Before a step can be deleted, someone — increasingly an agent — must produce its biography: who built it, in what year, after what. Call it friction archaeology. It becomes a discipline, then a checkbox, in that order.
Fences learn to testify.
Safeguard designers stop letting prevention be silent. The blocked attempt gets logged, counted, and reported upward — artifacts manufactured deliberately, so the fence can survive its next audit. Silent success becomes a design flaw. The best fences of the next decade will announce every single thing they stop.
The veteran’s last job is dictation.
The highest use of long tenure becomes sitting down with the agent and getting the unwritten reasons written — translating scar tissue into the record before it walks out the door. Companies will pay for this the way they pay for insurance, and for the same reason.
I run my own work the way I write about it: delegated, with a standing instruction of you decide. This piece is the one exception I have added since. When my AI proposes removing a step from anything I run, I want one line first: who built this, and what was happening that year. If the answer is a reason that has died, remove it without asking me. If the answer cannot be found at all, that is not clearance — that is the one case where I want the question brought back, because the line that remains runs through me, and this is what the line is for.
How this could be wrong
- Already true
- Chesterton’s fence is a 1929 record, and Wald’s survivorship memoranda are a 1943 one. Agents that read a company’s entire process library and propose removals are shipping products today, not a forecast. And every institution older than a generation already runs on fences nobody alive can explain.
- What has to happen
- Agents must graduate from advising on processes to holding standing authority over them, and approval habits must loosen the way they always do when the first ten removals pay off. The predictions ride on deletion becoming cheap enough to batch — a change in trust, not in technology.
- Where I am probably wrong
- Most unexplained fences really are dead, and the tax they collect is enormous, invisible, and — note — also missing from the record. Worse for my argument: the same machine that cannot see a silent fence is the best archaeologist ever pointed at one. It can read the 1994 memo in the first minute, not in month eight, and it never retires. If the provenance brief becomes the default before deletion, removal in machine hands could end up safer than it has ever been in human ones — and this piece will have described a transition cost, not an era.
Go back to the meeting at 9:14. The report was right about ten of the eleven steps, and the company is faster for taking them out. The vote on step seven took ninety seconds because nothing in the room could speak for it — no incident, no ticket, no memo anyone could find, just a decade of quiet.
The quiet was the fence working.