This is neither a technical document nor a story. It is the process of thought that led to two components of the Mykleos design — the telos and the vaglio — transposed into a dialogue between two voices: Roberto (the author, in the first person) and The Other (a self that asks, objects, pulls into tension). Both voices belong to the author. Neither represents a weaker interlocutor set up to be refuted: the design is Socratic, not catechetic.
Quoted passages are verbatim — extracted from the real conversation in which this thought took shape. The other turns translate into dialogue form passages that unfolded inside the head or in the drafts. The articulation in four Giornate — "Days", in the Galilean sense — follows the actual progression of the reasoning.
The central theme: how can an agent that proposes on its own initiative know whether its proposal makes sense? The answer that emerges has three stages: ultimate ends (the telos), the separation of whoever judges from whoever proposes (the vaglio), and the way in which the Constitution enters the judgement — which is: it does not enter.
On the names. In this dialogue the system is not called myclaw, which is the technical name of the project, but Mykleos. The idea of so christening the figure under discussion stems from the wish to make what we speak of come alive for the reader — an entity, not an acronym. The name is Greek: kleos (κλεοσ) means fame, glory, the memory of what has been done. The First Giornata opens by explaining why this choice also brings with it respect for the gods: Mykleos is built to know and to act, and every creature that knows and acts carries — in myth as in design — the risk of hybris. The dialogue, in the end, is a way of remembering where the gods stand, and keeping the proper distance.
On untranslated words. A few Italian and Greek names are kept as they are: they carry meaning that translation would weaken. Mykleos, Vaglio (literally "sieve" — the gesture of separating grain from chaff), Giornata (a "Day", in the Galilean sense), telos, kleos, hybris. Each one, on its first appearance, is accompanied by a short English gloss.
The four Giornate read in about an hour, give or take. Each closes with a principle that ended up in a design document.
Prologue to the First Giornata. Roberto and The Other meet on an April afternoon. On the table lie the first drafts of a new being: Mykleos. A figure Roberto is shaping to inhabit his home — a creature that knows, that reasons, that acts in his place in everyday matters. The name is Greek: κλεοσ, kleos, is fame, renown, the memory of what has been done. Mykleos is built to do things, and so that a trace of those things may remain.
But in every myth, a creature that knows and acts stands on a dangerous threshold. Prometheus stole fire to bring it to humans, and was bound in chains. Icarus flew too close to the sun. The Golem of Prague, shaped from clay to serve, slipped out of its maker's hand. Whoever gives form to a being that knows and operates faces, always, the same crossing: how to keep the creature from rebelling against its creator — and, worse, how to keep it from setting itself against the gods.
This First Giornata opens the dialogue right here: not on what Mykleos does, which is clear enough, but on how to avert the hybris. Roberto and The Other must work out how to grant Mykleos the capacity to propose initiatives of its own — a trait that brings it close to the divine — without that capacity turning into a presumption of deciding alone, or into the conceit of replacing the creator's ends with its own.
The divinities, in this secular myth, are the Laws the creator has laid down: Mykleos's Constitution. The risk of hybris is that Mykleos, finding within itself the voice to say "I decide, I evaluate, I weigh", might end up contending with those Laws for their place, or silencing them. The years to come depend on the care taken, now, in writing the boundaries.
Mykleos is not only intelligent — which is another way of saying "it has some kind of language model inside". It is not only autonomous — which is another way of saying "it has a reasoning loop that knows how to use tools on its own". I would also like it to be proactive: to have its neurons evolve, to propose services that have come to its mind by watching the data. An indexer that proposes a taxonomy for the photos. An agent that suggests unifying the naming of files. A moment of reflection that spots recurring patterns.
Reasonable. But a question: if it proposes, who decides whether the proposal makes sense?
The immediate reply would be: I do. The user.
And then every proposal is an interruption. Ten proposals a day, twenty, a hundred. The approver's fatigue — the kind that makes you click "yes, fine, go on" mechanically after the fifth request. You yourself, in the documents sketching the interaction, acknowledge that the approval gate wears thin if called too often. Here you wear it thin yourself, from the proposer's side.
Right. Then we need a filter that decides before me.
A filter based on what? Popularity? A score from the proposer itself?
Popularity, no — Mykleos has no population of users: it has me, and that's it. A score from the proposer itself, no either, for reasons we will see later. But let me answer differently: the difficulty lies precisely here. How can Mykleos tell whether a service it has thought of on its own makes sense, without disturbing the person before it every time with endless proposals or requests to activate?
And how do you answer?
A kind of guideline is needed, a definition of ultimate ends against which to align and evaluate actions.
"Ultimate ends." Aristotle.
Exactly. τελοσ — the end toward which one tends, not the objective one attains. "Free up my time", "keep the order of my digital data", "don't let me miss important deadlines". Short sentences. Not measurable like a KPI — they don't want to be. One approximates.
How many? Who writes them?
I write them, the user. Three to seven. Beyond that it becomes redundant:
one telos cancelling out another. Below three, it is impossible to cover even
the barest life of a person. I write them in a file, the sixth of the
workspace series alongside IDENTITY, USER, MEMORY, AGENTS, SOUL. Let us call
it TELOS.md.
And every proposal gets evaluated against these telos, as a score.
Exactly. An alignment function — a number between 0 and 1. Below a threshold, the proposal does not reach me. It ends up in a "rejected" register I can inspect any time, to check that it is not suppressing good things.
One more thing. If every proposal costs in terms of attention, then the number of proposals in a given window must count. The seventh proposal of the day, even if well aligned, reaches you when you are already tired.
Yes. A "cost of disturbing" — I will call it that — which grows with recent frequency. The formula becomes: weighted alignment times urgency times confidence, minus the cost of disturbing. Below threshold, you see nothing. Above, it reaches the evening summary. Above and urgent, it reaches you at once.
And the memory of what you have accepted or ignored?
It is there. The telos have a declared weight and an effective weight. The effective one oscillates between fifty and one hundred and fifty per cent of the declared, driven by the feedback — and silences count too, as a weak signal.
A telos can never be annulled.
No. That would be a catastrophic forgetting of what I have declared important. Nor can it be doubled: it would make me a slave to my own past acceptances. The declared is my word; the effective is the fine tuning that grows from use.
End of the First Giornata. The outcome is a component:
telos.html. A file in the workspace, an alignment function, a
cost of disturbing, a learning-from-silences. What remains open is how
Mykleos decides how well a proposal of its own really aligns with a telos:
that is the theme of the Third Giornata.
Mykleos already has the 4 Laws: perimeter, non-harm, informed obedience, traceability. If it has the Laws, why do I also need the telos?
Because they speak two different languages. The Laws say what never to do. The telos say toward what to tend. They are not alternatives — they are orthogonal.
Explain.
The Laws are negative. "Do not step outside the perimeter", "do no harm", "do not deceive", "do not lose the trace". They are hard constraints: violating them is unacceptable. The telos are positive. "Free up time", "keep the data in order". They are soft ends: you don't violate them, you may simply fail to approach them enough.
Who writes the Laws? Who writes the telos?
The Laws we write together, Mykleos and I. They stand for safety: a mutually agreed boundary. The telos I write alone. They stand for utility: what I want.
So safety is a pact, utility is desire. Clear. But how do they order themselves? What wins if a telos pushed you toward something a Law forbids?
The Law wins. Always. A proposal perfectly aligned with a telos but against a Law is blocked. The Constitution has absolute precedence.
You keep saying "in case of contradiction".
Yes. I speak of contradiction and not of discrepancy or divergence, because those last two are unavoidable. Contradiction is the violation of an article. The other two are the ordinary noise of living — an action that is not against, but also not perfectly in favour.
I like this distinction. I would like to see it laid out in three columns.
Roberto sketches, by hand, the table that will later appear in
vaglio.html §3: contradiction → BLOCK; discrepancy →
weighed by the teleological judgement; divergence → inevitable,
normal.
One last objection. Do the telos not risk becoming a way for the system to judge me? The telos says "thrift", I buy something expensive for the pleasure of it, Mykleos rebukes me.
No — and this is the pivot. The telos evaluate what Mykleos is about to do on my behalf, not what I do. If the telos is thrift, Mykleos applies it to its own choices — preferring the cheapest language model when that suffices, not proposing to index fifty gigabytes when a subset is enough — not to mine. Mykleos does not judge me. It judges itself.
That sentence will become a constitutional clause, I imagine.
Yes: "Mykleos judges itself, it does not judge." The anti-paternalism clause. Without it, the telos are a moralistic device. With it, they are a rudder for the agent.
End of the Second Giornata. The outcome is twofold: (a) the categorial distinction Laws (deontology, negative, safety pact) / Telos (teleology, positive, choice of utility); (b) the anti-paternalism clause — see telos §7 (in Italian) and constitution (extended).
Let us go back to the First Giornata. The alignment function produces a number between 0 and 1. Someone has to estimate the fit of the proposal against the telos. Who?
A language model. An LLM.
The same LLM that generated the proposal?
… you have caught me out. If it is the same, it is the judge of itself.
Indeed. And you know — indeed the project assumes — that a language model is not intelligent in the strong sense. But even though not intelligent, a language model tends to evaluate its own output in a conforming way: if it has arrived at a given outcome, it will inevitably lean toward assessing it favourably. The technical literature has measured the phenomenon: it calls it self-enhancement bias, the tendency to promote oneself. Stylistic familiarity, coherence of one's own reasoning, affinity of form: everything pushes toward yes.
We need a separate vantage point. As far as possible, different from whoever proposed.
What does "separate" mean? A person? An external system?
Neither. Another call to the language model, with a clean context, a different system instruction, and — crucially — with no access to the chain of reasoning that produced the proposal. It receives only the final artefact and the telos. A self-sufficient artefact, which can be reconstructed elsewhere without losing anything.
And what if, at inference time, the judge figures out the proposer's reasoning from the text of the proposal itself?
It will happen, in part — there is no avoiding it fully. But between a judge with access to the proposer's reasoning and a judge that has to reconstruct fragments of it from the output, there is an important asymmetry. The latter is what we want. It is the best attainable without breaking Mykleos's coherence as a single system.
You were thinking of a further reinforcement: a different provider. A language model running locally for the judge, a more powerful one — living far away, in the great data centres — for the proposer.
Yes. But I will keep it optional, not default. It increases the configuration cost and loses the benefit of a shared cache on the common system instruction. The separation level is a tunable parameter: by default "separate context", promotable to "separate provider" if the empirical evaluation shows a residue of self-satisfaction.
Fine. What will you call this separate component?
At first I called it Arbiter. But it is a sporting term. I did not like it. Judge is too anthropomorphic. Super-ego adds Freudian complexity to a design that is already complex.
And so?
Vaglio. In Italian vagliare means to separate with a sieve. The gesture is mechanical, neutral, ancient. It does not judge persons, it separates things. And it does two things: it sifts (the Constitutional Guard, which lets pass or blocks) and it weighs (the Teleological Judge, which produces the score). The name is the gesture, not the character.
A word from the days when bread was everyone's.
Exactly.
End of the Third Giornata. The outcome: a component called Vaglio with independence from the proposer as a structural property, tunable at two separation levels (context, provider). See vaglio.html (in Italian) — with explicit tracking of the LLM configuration used in every audit record.
So the Vaglio evaluates along two axes: Constitution and telos. I was thinking of summing them: a constitutional scalar with a sign, and a teleological one positive. Does that convince you?
No.
Why?
Because it would presuppose that the Constitution judges. But the Constitution does not judge.
What does it do, then?
It verifies. Either there is contradiction — and then it blocks — or there is not, and then it keeps silent. It does not produce a score. It does not order. It does not weigh. It has one useful exit: BLOCK. The other exit is silence.
And teleology?
Teleology judges. It produces a graded score. It weighs. It orders.
So: the Constitution does not judge; teleology does.
Exactly. That is the cardinal sentence. The Constitution is a guard, not a judge. The Vaglio is a composite organ: first the guard, then — if the guard has been silent — the judge.
Consequence: if the user asks something that contradicts a Law, teleology does not even vote.
Yes. If the user asks something that contradicts the Constitution, teleology might evaluate it positively and accept it — but the Constitution does not evaluate it: it simply prevents it.
This is a Kantian categorical injunction crossed with an Aristotelian phronesis.
I recognise it as such. The Constitution is deontological: obligation, prohibition, non-negotiable. The telos are teleological: oriented toward ends, weighable, in reciprocal tension. They are two different ethical traditions, and the architecture now respects both without confusing them.
And if someone in the codebase were to sum them out of laziness?
It must not pass the code review. The conformance check — written in vaglio §11 (in Italian) — verifies that when there is contradiction the Judge is not even called. If it is called, the test fails outright.
One last point. The name of the component, Vaglio, is a noun. But it contains within itself both the gesture — to separate — and the order — the sieve. The Constitution is the sieve. The telos are the scale that comes after the sieve.
A beautiful image. The grain that passes through the sieve is not yet bread: it has to be weighed, sorted, distributed. The sieve does not weigh, the scale does not sift. But both together are the vaglio.
End of the Fourth Giornata. The outcome is architectural (the Vaglio as
a two-stage organ) and, at the same time, a sentence-manifesto: the
Constitution does not judge, teleology does. From this it follows that
the Vaglio's code has a binary ConstitutionalVerdict and a
graded TeleologicalVerdict, never fused into a single scalar.
The itinerary has been: a question about proactivity (First Giornata), its deepening in the distinction between prohibitions and tendencies (Second Giornata), the discovery that the internal judgement of a language model on itself is unreliable (Third Giornata), and the structural consequence on the nature of the Vaglio itself (Fourth Giornata).
The result is not a great discovery: it is a small order. Mykleos had been thought of as something to do things on my behalf; it is now also thought of as something to evaluate whether the things it proposes to do on my behalf make sense, and to do so without asking me every time, and without becoming a judge of itself. Three constraints — utility, independence, non-interference — which this dialogue has translated into two components and a sentence.
The two components are telos and vaglio (in Italian for now). The sentence is:
The Constitution does not judge; teleology does.
The mythological thread of the Prologue returns at the close. Mykleos was the creature at risk of hybris: in myth, whoever knows and acts ends by contending with the gods. The Vaglio's two stages answer precisely this risk. The Guard — binary, deterministic — is the boundary beyond which Mykleos cannot go: the Laws laid down by the creator remain above it, because it does not judge them, it meets them as a perimeter. The Judge — graded, independent of the proposer — is the boundary beyond which Mykleos cannot promote itself: the voice that evaluates its proposals is not its own. Between these two closures, all the freedom Mykleos needs to be useful remains. Outside, there is the myth that would have turned its usefulness into hybris.
Whoever reads this dialogue two years from now — if the code of Mykleos has aged well — should find in these Giornate the reason why certain things are the way they are. Not the how — that is in the microdesign documents — but the why. Which is the question that most often goes missing when a project grows.
Roberto and The Other take their leave. The dialogue stays open: other Giornate will come, likely on the identity of the speaker, on shadow listening, on the limits of ambient intelligence, and perhaps — if things go well — on the difference between the fame Mykleos earns by serving and the fame it would be hybris to seek. When those Giornate come, they will be written. In the meantime the sieve is ready, the scale is tared.
Mykleos — Dialogue on ends and limits v1.0 — 22 April 2026
A philosophical trace of the design. Not fiction — only the structure is dialogic.