Featured

The Bounded Space in Which Danger Has Been Excluded in Advance

AI - o unealtă Hits: 832 40 minutes read

The fence in the park.

On freedom, limits and speed, from the fairy tale to automated computing systems.

In brief Freedom is not the absence of the fence, but the bounded space in which danger has been excluded in advance. From this follows a simple and uncomfortable thesis: the speed at which we can move, in a story or in a technology, depends on the quality of the limits we have set. Not on their number, but on how visible, proportionate, easy to revise and well understood they are. I start from a dog park, pass through the prohibition in fairy tales, the sacrifice in the legend of Master Manole and the rites of initiation, arrive at the mechanisms by which automated computing systems are constrained today, and stop at the cost of the fence, which nobody gets for free.

"I take my little dog to the dog park because I want him to have freedom, but I don't want him to run into the street. The dog sees the fence and does not go past it. He can play freely, and I know he is safe. The key to moving fast [in AI development] is having the right guardrails."

(The text from which this essay sets out, reproduced verbatim. Keith Fulton, Chief Data Officer and Chief AI Officer at the fintech company Jack Henry)

The cursor that blinked

In 1997 I spent six months in a room in Botoșani where, strictly speaking, I had no business being. I had enrolled in a two-week course, "IBM PC Operator". The course ended, but the room stayed open. There were two computers that worked and one taken apart, an instructor who chose not to notice that I came in every day, and a black screen bearing a single line, C:\>, followed by a cursor that appeared and disappeared every half second.

Nothing appeared on that screen of its own accord. If I did not type, nothing changed. If I typed an exact command, the command was executed exactly. If I got a letter wrong, a terse message appeared, without explanation and without tone, and I was left to work it out. It was, as I would say now, a pure tool: everything that happened began with me and stopped where I stopped.

And yet, even there, on so passive a machine, there was a fence.

If I gave the command that deletes all the files in a directory, the program did not execute it on the spot. It displayed a question, in English, whether I was sure, and waited until I pressed a letter: yes or no. If I tried to prepare the disk for fresh use — which meant erasing everything on it — the message was more emphatic still: all data on the disk will be lost. None of this stopped me doing as I pleased. The fence was not a prohibition. It was a second gate, placed exactly in front of the one gesture that can no longer be undone, and it asked one small thing: to repeat, consciously, what a moment earlier I had done in haste.

Today I believe it was precisely that question which let me experiment. On a computer that everyone in the room depended on, a beginner like me would have been a danger if every mistyped command could have wiped everything with no brake. As it was, I could type fast, I could try commands I did not yet understand, I could read, err, start again. The speed at which I learned in that half-year did not come from an absence of limits. It came from the fact that the one place where a mistake cost a great deal was guarded.

Years later, when I began working with websites and far more complicated machines, I recognised the idea under other names: backup copies, access rights, staging environments. All are variants of the 1997 question. They do not say "don't do it"; they say "do it, but not where you can no longer turn back."

Everything that follows is about this simple structure.

The dog park and a definition of freedom

The opening text has two parts. The first is a scene: a dog in a fenced park, a man watching it without fear. The second is a conclusion about the development of automated computing systems: to move fast, it says, you must have the right fences. The scene was well chosen. In a few lines it contains an idea worth taking seriously and — precisely for that reason — worth examining carefully.

Let us take it apart. The man wants the animal to be free. He does not want to keep it on a lead, he does not want to keep it indoors. But he wants one more thing: that it should not run out into the street. It would be easy to see here two desires at loggerheads, freedom on one side, safety on the other, and to look for a compromise. The text says something else: the two are obtained through one single thing. The fence does not diminish the dog's freedom; it makes it possible. Without it, the man would have to keep the dog close, shout after it, pick it up at every passing car. With it, he can afford to look away.

From this I draw the definition that will be the red thread of these pages:

Freedom is not the absence of the fence. It is the bounded space in which danger has been excluded in advance.

The definition sounds simple, but it contends with a far more widespread idea: that freedom means the absence of obstacles. The philosopher Isaiah Berlin called it, in an essay that has been a classic since 1958, negative liberty: the space in which nobody prevents you from acting. It is a valuable idea, especially when we speak of what the state may do. But, taken as it stands, it does not explain the park. The dog in the park is prevented in many ways: it cannot get out, cannot reach the road, cannot go wherever it likes. And yet, of all the places it could be taken, the park is the one where it moves least tensely. My definition does not reject Berlin's idea; it completes it with a criterion he was not looking for: not just how many obstacles there are, but what kind of harm lies beyond them.

From the definition there follows a practical question, to which I shall keep returning: how do we recognise a good fence? It is not at all obvious. There are fences that protect and fences that merely obstruct; fences we see, and fences we feel only after walking into them. That is why I do not rest content with the metaphor: I use it as a working hypothesis and submit it, in turn, to four trials: language, story, rite and mechanism.

Fence or censorship

Before that, a confusion that spoils many discussions must be cleared up. When people speak of limits placed on automated computing systems, many hear the word "censorship". The association is natural, since in both cases someone has determined what may not be done. But the association hides a difference of substance, which I shall use as a working criterion, not as a law.

Censorship looks at content and decides what may be said, and by whom. The fence looks at consequence and decides where a mistake can no longer be repaired. Censorship usually conceals its motives; the fence displays them: a fence you cannot see is no longer a fence, it is a trap. Censorship protects the one who applies it; the fence protects the one inside. And censorship, as it strengthens, narrows the space, whereas a good fence can enlarge it, because it gives you the courage to use it to the very edge.

The best image for the good fence is not a prison wall, but the barrier on a racing circuit, or the lanes marked out on an athletics track. Neither crowds anyone in. Each tells the runner where the ground ends on which they may spend everything they have. Or, if we prefer an image without people, the bed of a river: the water is not freer without banks, only more disordered, slower and, when the rain falls, dangerous.

One further clarification is needed, because I am approaching territory where language slips easily. The dog in the park is an animal. It sees the fence, has grown used to it, learns not to cross it. Automated computing systems, of which the rest of the text speaks, see nothing, learn nothing in the human sense, and get used to nothing. What limits them is built into them, or around them, by people — and it either works or it does not. I shall return to this difference at a particular point, because half the argument depends on it. Until then, the dog remains what it is, an animal, and the man in the park remains the only one who can build a fence.

What a proverb can and cannot say

The first trial is the one closest to hand: let us see what language has preserved about fences. A culture leaves its thoughts about limits in small places — one-sentence maxims, spoken at table and forgotten, then gathered by someone with patience. In our case that someone was Iuliu A. Zanne.

Zanne published, between 1895 and the beginning of the next century, Proverbele românilor din România, Basarabia, Bucovina, Ungaria, Istria și Macedonia (The Proverbs of the Romanians of Romania, Bessarabia, Bukovina, Hungary, Istria and Macedonia), a collection in ten volumes containing some 47,330 sayings, to which he systematically added equivalents from other languages. If you want to know whether a saying actually circulates or merely sounds like one, this is where you check. The collection's importance goes beyond philology: the Romanian delegation to the Paris Peace Conference of 1919–1920 placed it among its evidence for the cultural unity of the Romanians. Such a text is not quoted from hearsay.

That is why I make a methodological choice here, which I state openly. In this essay I quote no proverb. The rule I have set myself is that any saying about boundaries, watchfulness, haste or the brake enters the text only if I can attest it in Zanne's collection, with the volume indicated. That attestation could not be made in preparing this version, so the ideas remain in my own words. I prefer this to a trove of popular wisdom fabricated from memory, which I would then sign with the signature of an entire people. That is precisely the sort of gesture that turns an essay into a collection of beautiful, unverified quotations.

It is not mere scrupulousness. The choice has a gain. Without proverbs to hand, I am forced to ask what a proverb can, in principle, offer to a discussion of fences. I believe three things, and none of them is what one would expect.

First: a proverb states; it does not design. A proverb about watchfulness or boundaries says, in essence, that the limit matters and that neglecting it costs. That is all. It does not say where the fence goes, how high, with what gate, who checks it, how it is repaired. For a parent choosing a park, the observation is sufficient. For someone building an automated computing system, it is not. A program does not receive advice; it receives specifications: what goes in, what comes out, what may not come out, what happens when the rule is broken. The distance between "the limit matters" and "the limit goes here, in this way" is the distance between wisdom and engineering. Neither replaces the other, but without the second the first produces nothing.

Second: a proverb collection preserves both the advice and its counter. Any old culture holds together prudence and daring, care and courage, slowness and resolve — each good in a different context. Whoever quotes a single proverb from such a pair and presents it as "the wisdom of the people" is making an editorial choice, not a discovery. This warns us that, when we read proverbs about speed and haste, we should not treat them as arguments. They are testimonies about how people thought, not demonstrations.

Third: a proverb preserves an experience, but hides its cost. Every maxim about watchfulness grew out of a loss, but the loss is no longer visible in it. The conclusion remains; the invoice does not. In the rest of this essay I shall try to put the invoice back.

When the attestation from Zanne has been done, the place of these three observations will remain the same: proverbs will enter here as illustrations of what language has retained about watchfulness and haste, not as authority. Until then, I turn to a source stronger than any saying, because it does not merely state — it demonstrates in motion: the fairy tale.

Prohibition as engine: what Propp saw

If the proverb says the limit matters, the fairy tale shows what happens to a story when the limit exists. To read this well I need an instrument, and the best is the one devised by the Russian folklorist Vladimir Propp in his Morphology of the Folktale, published in 1928.

Propp began from a simple observation: the tale is built not on characters but on functions. The hero may be a prince or a poor lad, the dragon a beast or a witch, but what each does in the story repeats, and appears in an almost fixed order. He called these units functions and observed that their number is limited. The narrative begins with an absence — someone leaving home or the community's care. Then comes a prohibition. Then its violation. And only then comes the harm or the lack that sends the hero on his way. In Propp's numbering, the prohibition is the second function, its violation the third — two paired functions, linked like a question and its answer.

What follows from this is not a detail of literary theory. If the order is thus, then the prohibition is not an ornament placed at the start of the story to make it harder. It is the story's engine. Without prohibition there is no violation. Without violation no imbalance. Without imbalance no quest, no trials, no helpers, no battle, no return. And without all these the fairy tale reduces to one sentence: once there was a man, he lived quietly and he died. The fence, in the fairy tale, does not stop the narrative. It produces it.

The semiotician Algirdas Greimas took up Propp's scheme and rewrote it in terms of opposed pairs: departure and return, the formulation of the prohibition and its violation. He added three trials through which the hero passes: the qualifying trial, in which he receives the tools he needs; the decisive trial, in which he confronts the evil; and the glorifying trial, in which he is recognised. The reader who wants a map of the fairy tale finds in Greimas something Propp did not say in these words: the story is the process by which a broken limit is repaired. The fairy tale ends when equilibrium has been restored — that is, when the broken fence stands again, or when another has grown in its place.

Read thus, the fairy tale is no longer a children's story. It is a model of the life of a fence: set up, broken, paid for, repaired.

One more thing must be said, lest an instrument become dogma. Propp built his scheme on the wonder tale, and he himself allowed, in passing, that the structure can be found, with adaptations, in other species too, from legends to animal tales. More recent researchers have applied it to Romanian texts that are not strictly fairy tales. But a scheme that explains much does not explain everything. I use it as a lens, showing what matters in a story, not as a law that determines it.

For the rest of the essay, the lens can be reduced to four questions, put alike to a fairy tale and to a technical system:

  1. Who sets the limit, and at what moment? Usually someone about to be absent.

  2. How is it formulated? As a list of forbidden things, or as a verification procedure?

  3. How is it broken? Through defiance, or through a failure to grasp the situation?

  4. What does the repair cost? Who pays, and with what?

Let us put them now to the first two stories every schoolchild in Romania has read.

Two fences from school: Harap-Alb and The Goat with Three Kids

I shall not retell either The Story of Harap-Alb or The Goat with Three Kids. Anyone who has been through the lower and middle forms knows them. What interests me is only how the fence works in each, viewed through the four questions.

Harap-Alb: the limit formulated in categories

The Old King, the father, sends his youngest son to his uncle, Emperor Green, and gives him one piece of advice: to beware of the red man and, above all, of the spân (the bald man). We can answer the first question. Who sets the limit? A father who knows he will not be there. His absence — function one in Propp — is not an accident; it is the very condition of the fence: a limit is set precisely for the moment when the one who sets it can no longer check.

The second question: how is it formulated? Here lies the weakness. The advice is a list of dangerous types, defined by visible traits — the colour of a man's skin, or the lack of hair. It is a rule of the kind "beware of X", not "check thus and so before trusting anyone". Any rule formulated in categories has a problem: life does not present itself in categories. It presents itself as a concrete man, who offers help, who knows the way, who speaks beautifully.

The answer to the third question is therefore more surprising than we would expect. The violation does not come from defiance. The hero does not decide to disobey. He simply fails to connect the situation he is in with the rule he was given. What stands before him is a travelling companion, not "the type I was warned about". The rule is good, but it has no mechanism to bring it before his eyes at the right moment. From here follows the chain of misfortunes: the loss, servitude, the trials.

The fourth question: what does the repair cost? Almost the entire story. The hero loses his rank, his freedom, for a while even his dignity, and must undergo, under a foreign master, the trials Greimas would call qualifying, decisive and glorifying. The more vaguely the limit was formulated, the more expensive its violation. This is a lesson we shall find again, centuries later, in the building of technical filters: a list of prohibitions is always easier to evade than a procedure.

The Goat with Three Kids: the limit as procedure, weakened at verification

A much shorter story, a much more transparent structure. The goat goes out to feed, and her absence is once again the condition. She leaves the kids a rule: open the door to no one. Not a rule about dangerous types, but one about a gesture — and surrounded by a procedure: recognition by voice.

Here the fence is far better made than in Harap-Alb. The prohibition applies to everyone, not just a category. The verification is concrete. The procedure can be followed by a child. And yet the story breaks. Why? Because the single check — the voice — can be imitated. The wolf fakes his voice, and two of the kids open the door. The third does not.

Together, the two stories show two different ways a fence gives way. In one, the rule is too vague to be applied at the critical moment. In the other, the rule is good, but the verification it rests on is weak. In the language of our own day: a password that can be guessed. An access system that recognises a single trait. A well-hung gate in a badly built fence.

Let us notice one more thing, about the kid who does not open. Same fence, same rule, same wolf — different outcomes. The fence does not apply itself. Someone must apply it, and when the guardian is absent, application depends on the one inside. The goat cannot be at home; that is why she left a rule. But a rule left behind is not a rule kept. Here appears the part the fairy tale states without naming: true fences are designed for the moment when we are no longer watching.

We meet the same situation in front of any system that runs without someone standing beside it — a program working through the night, a tool generating text for thousands of people at once. We cannot be there at every step. We therefore need a fence that holds even when we are not looking. And above all, we need a way of noticing when it no longer holds.

Rite, sacrifice and maintenance: what a fence costs

So far the fence has seemed a matter of formulation: well or badly said, well or badly verified. But the fairy tale has a deeper root, and it changes the way we must speak about cost.

Propp argued, in the works after the Morphology, that the oldest foundation of the fairy tale is the rite of initiation — the ceremony by which a young man passed into a new status. In a study of the hidden-face motif in European fairy tales, the researcher Otilia Ursache recalls the very formulation: the initiation rite is the oldest basis, and in the ceremony the mimetic representation of the novice's death and resurrection as a new man predominated. We need not be carried too far by this reading. What I keep is something else: the prohibition in the fairy tale was not born as a storyteller's device. It is the memory of a practice in which the limit was not merely spoken but lived: the young man was separated from the community, told what he might not do, put through trials, and at the end returned changed.

The anthropologist Arnold van Gennep, in his 1909 book on rites of passage, described this structure in three stages: separation; a liminal period in which the one passing through is no longer what he was and not yet what he will be; and reintegration. The liminal period is when limits are strictest and, at the same time, most instructive. In a study of the hero's initiation in the Romanian fairy tale, Camelia Bof describes the hero's journey as a sequence of trials under the guidance of a mentor, in which order matters: first accompaniment, then autonomy.

This idea deserves to be stopped for a moment, because it gives us a practical rule we shall use later: the fence does not remain the same throughout the life of a relationship. While the young man has not passed his trials, the limits are tight and in someone else's keeping. Once he has passed them, the limits widen, because the checks have shown he can bear part of them alone. The parent who finally lets the child go out alone to the corner of the street has not given up the fence. He has moved it from his own hand into the child's mind.

Where the analogy ends

A distinction must be made here, because it protects the argument from a mistake. The dog in the park and the child in the initiation rite can internalise the fence. After enough walks beside it, they know the place. They can come to respect it without needing to be watched.

Automated computing systems do not. A program does not make a limit, a gate, its own. The limit is either built into it or around it, or it does not exist. There is no moment at which it "enters the character" and can be forgotten. From this follows a consequence I repeat because it is easy to lose sight of: in the case of an automated system, the maintenance cost of the fence never falls to zero, and it falls on people. Every widening of the limits must be decided, checked and, if necessary, cancelled by someone. A parent may, at some point, relax. The builder of a system may not.

The sacrifice in the walls

To see how much it costs to build a fence that holds, there is in our culture no more cutting image than the legend of Master Manole. Whoever has been through school remembers: walls raised by day collapse by night, until the master builders resolve to wall in the first wife to come, and the first to come is Ana, Manole's wife. The monastery stands.

The legend can be read in several ways, all of them heavy. I keep two, because they bear directly on what we are discussing. First: a durable construction demands a price, and the price is not symbolic. Second: the one who pays is not the one who decided. Ana took no part in the pact among the builders, nor did she know that she was its due date.

Can this be applied to technical fences without tipping into exaggeration? I believe so, with measure. The limits placed on automated systems have costs that do not show in the final result. Reports in recent years have shown, for instance, that part of the work of checking and labelling harmful texts is done by poorly paid people, who read exactly what the rest of the users should never have to see. I do not claim the image is perfect. I say only that, when we praise a fence, it is worth asking who paid for its being raised.

Three kinds of cost

From everything covered so far, there follows a short list, which I shall set, at the end of the essay, against the fence's gain:

  1. The cost of construction: what is invested, as with Manole, for something to stand. Work, time, attention, sometimes suffering.

  2. The cost of maintenance: the rite is repeated, the fence is checked, the gate is oiled. For automated systems, as I have said, it never falls to zero.

  3. The cost of unnecessary blockings: a badly placed fence also stops what should not have been stopped. The thing least spoken of, because it leaves no trace: what could not be done.

One more thing must be said, lest we conclude that any fence is a good thing. The English writer G. K. Chesterton formulated, in a text of 1929, a simple rule: if you come upon a fence and cannot see what it is for, do not remove it before you learn why it was put there. The rule has its reverse side, which Chesterton accepted: once you have learned, if the reason is gone, the fence may come down. Maintenance means this too: revision in both directions.

Mechanisms: how the fences of an automated system are actually set

I have arrived at the fourth trial, the technical one. For a while I shall be colder. What must be kept is what we have gained so far: the definition of freedom as a space without danger, the four questions of Propp's lens, and the three kinds of cost.

The turning point: the system does not see the fence

The text I set out from says that "the dog sees the fence and does not go past it". A true sentence, for a dog. For an automated computing system, no. It must be said clearly, once, and then respected: such a system sees nothing. It has no fence in front of it that it respects or violates. It has a structure, made of parameters and rules, which produces, for a given input, an output. If the output must be limited, the limiting is done either inside the structure or around it, by people, through construction.

Even the adjusting of these parameters is better described technically than with words borrowed from animal psychology. What some papers call a "reward" is, in practice, a number calculated according to a rule, and the parameters are modified so that the number grows. The rule was written by people. The quality of the fence depends, therefore, on how that rule was written.

Fences in layers

The first observation common to all the technical sources I have read is that safety fences (in English, guardrails — the term nearly all sources use) are set on several levels. In an IBM presentation, the limits are grouped into four layers: those of the data, of the model, of the application, and of the infrastructure. The division is natural and can be explained with an ordinary application. If an application uses a program that generates text, one must decide what data may be sent to it (the data layer), what kinds of results the program itself may produce (the model layer), what the user sees of all it has produced (the application layer), and what server resources the whole assembly may reach (the infrastructure layer). Each level catches a different kind of failure. A single fence is not enough, for the same reason that a park has a gate, a mesh and a wall, and not just one of them.

Input filters, output filters and the airlock

The best known are the filters at the two ends: one checks what goes in, the other what comes out. They resemble the park gate. And, as at the gate, this is where most failures happen, because the gate is where the fence must open. Well-made dog parks have a double entrance, a small airlock with two gates, of which never both are open at once — precisely so that a dog cannot slip out the moment another comes in. The idea is very simple and very applicable: the crossing points are the places that deserve the greatest design care.

To correct, not merely to stop

A recent paper, published on arXiv, proposes a step further. The authors start from the observation that harm does not always arise from a single output, but sometimes from a chain of steps, each innocent taken separately. A filter that checks only the last step arrives too late. They formulate safety as a staged decision problem — exactly as in control theory — and propose a filter that foresees whether the proposed actions lead toward the violation of a limit, and corrects them before that point is reached. The method compares itself, by the authors' account, with simple signpost-and-block solutions and surpasses them in safety.

In short, in control theory a safety filter receives the desired command and modifies it just as much as is strictly necessary for the system to remain within a zone considered safe. A familiar example: the system that keeps a car in its lane. It does not take the wheel from you, and does not stop your journey. It straightens your trajectory only when you are about to touch the edge. It is, literally, a balustrade with a corrective effect, not a wall. I present the idea as a proposed mechanism, not as one proven in general; the source is a research paper, still under discussion.

When the constraint literally shortens the road

There is one place where the link between limit and speed is no longer metaphor. In search problems, from solving a sudoku to timetabling a school, the rules of the problem eliminate an enormous set of impossible variants before any of them is tried. A program that solves sudoku does not try every possible completion — that would take an eternity. It uses the rules to cut off the dead branches, and solving time falls, often by orders of magnitude. Several technical sources note the same thing: constraints guide the search and spare it blind alleys. Here the law of my title appears in its pure state: the better the rules are chosen, the faster one arrives.

A written set of principles, and one of its limits

Also worth mentioning is a method published in 2022 by researchers at Anthropic, called Constitutional AI. As the paper describes it, a written set of principles — a "constitution" — is used so that the produced texts undergo automated critique and rewriting, and the results are then used to build the number by which the parameters are adjusted. The declared advantage: the amount of human labelling of harmful texts is reduced. Critical distance requires that we note a limit as well: the choice of how to write the principles remains a human decision. It has not disappeared; it has only moved. I return to this problem in the section on who raises the fence.

Logs and alarms: how you learn the fence has given way

One question remains that The Goat with Three Kids asks without saying it: how do you learn, once you have left, that the fence has given way? The goat learned too late, from the kid who escaped. An automated system can be built not to wait that long. Most technical fences come with a log — a written record of every significant action — and with an alarm, a threshold beyond which someone is called. The log stops nothing, but it makes the violation leave a trace, and the trace is the material of repair. The alarm repairs nothing, but it shortens the time between failure and remedy. Together, they turn every attempt to get past the fence into information, not merely damage.

One thing deserves emphasis. A log that nobody reads does not exist. Many organisations amass enormous volumes of records that no one opens except after an incident. The quality of a fence includes, therefore, who reads the traces, and how quickly. In the fairy tale, that role belongs to the storyteller: he remembers where the rule was broken, so that it does not break a second time in the same place. In engineering, it belongs to someone with a name, with responsibility, and with time.

Latitude and revision

Finally, several commercial sources use the same logic we saw in the rites: stop conditions at every step, spending limits, logs that keep the trace of each action, quality checks before a result reaches the user, and periodic revision of the limits as the system is granted more latitude in execution. It is, in engineering language, exactly the guidance that precedes autonomy: the fence widens after the trials, not before.

Why a good fence gives speed

I have gathered enough pieces to attempt the whole argument. I shall build it step by step, because the thesis is sharper than it looks.

A rule to start from

Whoever drives a vehicle, leads a team, runs a project or a system does not choose speed at random. They choose it, even without noticing, according to the gravest mistake they can no longer undo. On a road with a wide hard shoulder and a barrier at the cliff's edge, one drives faster than on a road without them — not because the drivers are braver, but because the worst case has shrunk. I call this admissible speed: the maximum speed at which the possible harm remains bearable.

Examples are close at hand, and not technical. A climber ascends a hard route with rope and anchors: not because he fears falling, but because, knowing the fall is limited to a few metres, he can afford to attempt the difficult moves. Without protection he would choose easier, slower routes. A railway divides its line into sections in which there may be a single train at a time; this limitation, apparently a brake, is precisely what allows many trains to run fast on the same line. An electrical fuse melts under excessive load; thanks to it, a house can use powerful appliances without every socket being a danger. And on the 1997 computer, the question "are you sure?" was what allowed me to try commands I did not yet understand.

From rule to thesis

The thesis of this essay is that speed, in the fairy tale as in technology, is directly proportional to the quality of the constraint. I can support it in four steps.

  1. The speed of advance is chosen by those who advance.

  2. In the end, they choose it so that the gravest irreversible outcome remains bearable. If they do not, they pay later: one technical source calls this the "rework tax" — the time lost on repairs because the checks at the start were skipped.

  3. A better fence lowers the possible harm at the same speed — or, at the same possible harm, allows a higher speed.

  4. Therefore, as the quality of the fence rises, admissible speed rises.

The second step is the weak hypothesis. There are people and organisations who run without looking at the worst case. The thesis does not say that they move more slowly. It says that their net speed — the speed you keep after subtracting the rework — is lower than that of those with good fences. By "proportional" I do not mean a mathematical straight line, but a relation that grows, within a certain interval, with quality. Nowhere does it say that more fences mean more speed. It says that better quality means more.

What "quality" means

From the whole road travelled so far, the quality of a fence is read in five signs:

  • Visibility. A fence you cannot see is a trap. The prohibition in the fairy tale works precisely because it is spoken; the technical fence works if the people working with the system know where it is and why.

  • Proportion. Neither tighter than needed, so as not to produce unnecessary blockings, nor looser, so as not to leave open what should be shut. The rule in Harap-Alb, formulated in categories, was too loose in one place and too tight in another.

  • Sound verification at the gate. The Goat with Three Kids had a procedure, but its verification — the voice — could be imitated. A fence's quality is decided at the crossing points, like the park's airlock.

  • Revisability. A fence that cannot be moved becomes, sooner or later, an obstacle or an illusion. The initiation rite has no fixed limits: they grow with the trials.

  • Knowledge of the cost. A fence whose invoice nobody knows cannot be maintained, and cannot be defended. Manole knew what he was walling in. Whoever raises a fence without being able to say what it cost risks raising it in vain.

There is nothing revolutionary in this list, and that is as it should be. Any work held upright for more than a generation — from bridges to the rules of a game — has these five traits.

How you can see, in numbers, that a fence is good

A thesis that cannot be checked does not deserve to be carried. Fortunately, the quality of a fence can be tracked with three simple numbers, which any team and any parent can hold in mind, even without writing them down.

The first is how often the fence is violated — how many times something crosses it that should not have. If the number grows, something has changed in the ground, or in how the fence was placed. The second is how often the fence stops what should not have been stopped: how many legitimate things were blocked. A fence with zero violations and a heap of unnecessary blockings is not a good fence; it is a wall. The third is the time to repair: how long it takes from the discovery of a crack to mending it. A good fence is not one that never gives way, for no such fence exists. It is one that rarely gives way, is seen when it gives way, and is quickly repaired.

The thesis thus becomes checkable. If a team were to raise, in turn, the quality of its fences — measured by these three numbers — and its net speed, the speed after subtracting rework, did not grow, the thesis would be weakened. I do not claim we have good measurements for every case today. I claim only that the thesis can be checked, which is what distinguishes it from a slogan.

What the literature says, and how far to believe it

A reader searching the internet for the link between limits and speed in automated computing systems quickly finds dozens of pages asserting it. It is worth saying calmly what they are worth.

Most are commercial presentations, pages published by companies, or popularising texts. They repeat, in near variants, a few formulas: limits are "the price of speed", paid in advance or in incidents; guardrails do not brake innovation; an organisation that ignores them will pay later. The formulas sound well and have a kernel of truth, but they are assertions, not evidence. Those who write them also have an interest: they sell products or services of this kind. That is not a reason to reject them. We read them with the reserve we would bring to an advertisement for a padlock.

Beyond the formulas, the mechanisms remain, and these defend themselves: the layering of limits, input and output filters, correction instead of mere blocking, the cutting of impossible variants in search problems, periodic revision. For these we need no praise. We can judge them by how they work, like any other tool.

There is one further difference between the sources and this thesis. Many of them speak of a "balance" between speed and safety, as though they were two pans of a scale: more of one means less of the other. At a given level of fence quality, so it is. My thesis says something else: raise the quality, and you move the whole balance. You no longer choose between two evils; you get more of both. It is a bolder claim and, for that very reason, harder to prove; I support it with the argument above, not with statistics I do not have.

Objections: when the fence becomes a cage

An argument that does not seek out its own weak points does not deserve trust. Here are four serious objections, with my answers.

The badly placed fence

Some limits stop too much. A filter that blocks any text in which a given word appears, regardless of context, stops the doctor, the historian and the teacher alike. Such a fence produces unnecessary blockings and teaches users to work around it. The objection is correct, but it does not overturn the thesis: a fence that blocks what it should not is a fence of poor quality — the proportion is wrong. It shows only that the number of fences means nothing by itself.

Who decides what counts as danger

This is the most serious objection. In a dog park, the danger is easy to recognise: the street. We all agree that a dog under wheels is a harm. With automated systems, the definition of danger is almost never so clear. What is a harmful text? What counts as dangerous information? Who decides? The principles by which a system is adjusted are written by someone, and their choice becomes, without anyone voting, the rule for millions.

I have no answer that dissolves the problem. I have, instead, a division that can hold it in check. There is a core of irreversible harms about which people are, in practice, agreed, and there the fence may be strict. And there is a wide zone of disputed things, where the fence must be less rigid: visible, motivated, contestable and easy to revise. The gravest mistake is the mixing of the two zones: to treat what is disputed as though it were core, and to make of it an invisible wall.

The impassable fence as illusion

In the fairy tale, the prohibition is almost always violated. We have seen that this is how the story starts. The lesson for the engineer is that any fence will be tested. Users quickly sense where the edges are and try to go past them; anyone who has played an open-world game knows that players search passionately for the places where the "invisible wall" has a crack, and speedrunners find them. Circumventing restrictions is not the exception; it is the rule. A good fence is designed for the attempt to cross it: on several levels, with logs that keep the trace, with the correction ready-made for the day someone finds the crack.

The fence that lulls vigilance

There is one further objection, more subtle. When we feel safe, we become bolder and sometimes more careless. Some researchers call it risk compensation: protective measures do not always deliver the full expected gain, because people raise, without meaning to, the level of risk they accept. If that happens here too, a good fence might lead, paradoxically, to less attention around it. The answer is not to give up the fence, but what I said about maintenance: a fence without periodic supervision becomes an empty promise. The goat left, but she left a rule. She did not leave a guarantee.

Conclusion: from the street, into the park

I began from a tension. On one side, "running into the street": chaos, danger, what can no longer be undone. On the other, "playing freely": exploration, play, all that has value precisely because it can be attempted and reattempted. The two seem to exclude each other, but they do not. They are separated by a fence, and the fence is the one thing that makes them exist in the same place.

The resolution does not come from compromise; it comes from a shift in the question. We no longer ask how much freedom we sacrifice for safety. We ask where the irreversible lies, and how we take it out of the space in which people move. What remains inside is a space of play — of trial, of recoverable error, of speed. Freedom, as we have defined it, is not the absence of the fence, but the bounded space in which danger has been excluded in advance.

The road passed through four trials. Language showed us, through what I could not quote, that to state is not to design. The fairy tale showed us that prohibition is the engine of the story, that a rule can fall through vagueness or through weak verification, and that the fence is set for the moment when we are no longer watching. The rite and the legend showed us that the fence is built with sacrifice, maintained with effort, and widened as the trials are passed. And the technical mechanisms showed us that the same ideas can be written in code: layers, the double gate, filters that correct, constraints that cut off dead ends, periodic revision.

From this follow the criteria I have established for the "right fences": visible, proportionate, with a well-verified gate, revisable, and with a known cost. On them rests the thesis: speed, in the fairy tale as in technology, grows with the quality of the constraint. Not because limits are good in themselves, but because they take out of our path the gravest mistake and let us run over the rest of the ground.

I return to that 1997 room. If I had to sum up what I learned there, in six months, it would be two things. First: a passive tool needs no fence, because everything that happens begins with you. Second, harder to see at the time: even there, there was a question, placed exactly in front of the one gesture that can no longer be undone. Today, tools are no longer passive in the same way, and fences can no longer remain at the level of a question on a screen. They must be designed, paid for and maintained by people who know that the man in the park is looking elsewhere precisely because he trusted the fence.

And one more thing must be kept in mind, which is not about technology. The one who builds the fence is the one answerable for it. The dog will never know what it cost.

Sources

Found and consulted in preparing the essay

From my own knowledge, to be verified before publication

  • Berlin, Isaiah, "Two Concepts of Liberty" (1958).

  • van Gennep, Arnold, Les rites de passage (1909).

  • Chesterton, G. K., "The Fence", in The Thing (1929).

  • Bai, Yuntao et al., Constitutional AI: Harmlessness from AI Feedback (2022).

  • Reporting on the labelling of harmful texts (from the international press, 2023–2024): concrete sources to be added.

  • Creangă, Ion, Povestea lui Harap-Alb and Capra cu trei iezi, for the details in section 5; the legend of Master Manole, in the version found in school textbooks.

Transparency note: text edited with automated editorial assistance, in keeping with the site's transparency policy. The personal scenes retell memories published by the author; the facts and quotations remain for him to verify before publication. The tool used in the editing is made by the company that published the method described in section 7, which the reader has a right to know. No proverb appears in the text, because their attestation in Zanne, with volume, could not be made.