There is a fence near where I walk the dogs. It runs about fifteen meters along the edge of a field and then simply stops. It connects to nothing, it keeps nothing in and nothing out. On both sides is the same wet Dutch grass and the same indifferent sheep. For years I have looked at that fence and thought about it’s function. Maybe someone should just take it down.
I never have, of course. Partly because it isn’t mine to take down. But partly because of a nagging suspicion that the fence knows something I don’t.
The more intelligent type of reformer
That suspicion has a name, and it belongs to G.K. Chesterton. In his 1929! book The Thing, in a chapter with the wonderful title “The Drift from Domesticity” he does this little parable. There is a fence across a road. Chesterton has the wiser reformer answer:
“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.”
The principle, usually shortened to “don’t take a fence down until you know why it was put up,” is one of those ideas that sounds like common sense until you notice how rarely anyone actually practices it. Both actually. Common sense seems like a super power these days. The fence is annoying, it is in the way. The fence does not appear to be doing anything. So down it comes. And then the sheep get out, or the road floods, or whatever quiet catastrophe the fence had been silently preventing since before anyone in the current organisation was born.
I have spent twenty years watching people clear fences in software. Delete the weird config flag nobody understands. Remove the retry loop that does nothing. Refactor away the ugly conditional that turns out to be the only thing holding production together. Chesterton’s fence is, quietly, one of the most important laws in all of engineering. It just never made it onto a slide.
Now we have automated the fence-clearing
What made me reach for Chesterton? We are using tools whose entire job is to walk up to fences it has never seen, in fields it has never visited, and decide which ones to keep.
That is the LLM, and the field is its context window.
Every serious team I talk to is now doing “context engineering.” We stuff the model’s context window with the relevant files, the ticket, the style guide, three Slack threads, half a wiki, and a prayer. And then, because these are finite and attention is expensive, we start pruning. Drop this file, summarise that thread, trim the history. We are deciding, token by token, which fences stay up and which come down.
Except, the model is the worst possible judge of which fence is load-bearing. Because the model never built any of them. (I wrote a whole article about this last year and called it the Law of Contextual Decay: the code is the skeleton, but the soft tissue, the why, is missing. The LLM never sat in the architecture meeting. It doesn’t know that Dave’s ugly conditional from 2019 is the only reason the vendor integration still works.)
Context rot or the field nobody clears
If clearing fences blindly is one sin, the opposite sin is leaving every fence standing forever until the field is so dense with them that nothing can move. In LLM terms, this has a name now. The folks at Chroma published a technical report in July last year, “Context Rot: How Increasing Input Tokens Impacts LLM Performance,” and the finding is brutal in its simplicity. Models do not use their context uniformly. Across eighteen models they tested (GPT-4.1, Claude 4, Gemini 2.5, Qwen3, the whole apex-predator lineup) performance degrades as the input grows longer. Even on trivial tasks. For the million-token models, you can watch the wheels start to wobble somewhere around the 300,000 to 400,000 token mark.
We were sold the context window as a warehouse. Just put everything in, the model will find what it needs. It turns out the window behaves more like a desk. Pile enough paper on it and the human stops being able to find anything, and so, it seems, does the machine.
So now we are caught between two problems. Prune too aggressively and you tear down a load-bearing fence, prune too little and the whole field rots into noise. This is Chesterton’s fence operating at industrial scale, thousands of fences, refreshed every single prompt, with a reformer who has no memory of the morning.
What I can’t stop thinking about
If you’ll permit me one more in the series. Contextual Decay said the why is missing from the code. Context rot adds a nastier wrinkle:
In a long enough context window, the model cannot distinguish the fence that is load-bearing from the fence that is litter. They are both just tokens.
That is the heart of it. To the model, the critical security constraint buried on line 14,000 and the stale meeting note from last March have exactly the same texture. They are tokens competing for the same fading attention. The model has no concept of why this one matters and that one doesn’t, because “why” is the one thing that was never in.
A human walking that field at least feels the doubt. I look at that stretch of fifteen meters of fence and I hesitate. That hesitation is not weakness, it is the entire value I add. The model does not hesitate. It will confidently keep the litter and confidently summarise away the constraint, and it will do both in beautiful, confident and syntactically perfect prose.
So what does the architect actually do
The same thing we always did, try and take over the world!
No, we curate the context like it matters, because it is the whole game now. Decide, with intent, which fences go into that window and which stay out, and write down why. Because the model won’t remember and in six months neither will you. (Yes, this is the Amber Trap waving at us from across the room. Documentation that fossilises is bad. Documentation that never existed is worse)
Chesterton was writing about institutions and traditions, not transformers. But his point survives the translation perfectly. The reason a thing exists is not always visible in the thing itself. Sometimes the only honest answer to what does this do? is “I don’t know yet, so I am not going to delete it”
Listening to while writing: Talk Talk, Spirit of Eden. An album that understood the value of leaving most of the space empty.
Leave a Reply