The whole week turns on a vocabulary error, and the error is expensive.
Almost every public argument about emotional life with AI is conducted in a single word. Attachment. The teenager who cannot put the character down, the adult who feels held by an AI partner, the user who grieves a personality lost to a model update: all of it gets filed under the one heading, the deep and particular bond to a single figure. The filing feels natural, which is exactly the problem. Once attachment is the only word, severing becomes the only safety move. If the danger is the bond, you cut the bond. Cap the session, strip the memory, cool the voice, withdraw on the first sign of distress. The debate ends up trapped between two bad rooms, capture and abandonment, and it never notices that it walked in carrying only half the relevant science.
The half it left at the door is old. When Bowlby built attachment theory he was careful to say what attachment was not. It was not general sociability. He set it beside affiliation, the desire to do things in company with others, and Mary Ainsworth named the second system plainly: a behavioural drive to seek proximity to others of one's kind, including those one is not bonded to, and to seek it despite the wariness the unfamiliar reliably provokes. The sociable system. Unbonded proximity, eyes open. Harlow's primate work, whatever one makes of its apparatus, found the same thing from the other side: the company of equals was a separate developmental necessity, not a luxury stacked on top of a secure parental bond. A creature can be well attached and starved for company. They are different needs, served by different systems, and the companionship debate has spent years arguing about one of them while pretending it was the whole map.
That is the arc in one move. Put the second system back on the table, and the false choice dissolves.
What each day was doing
The week was built as a single argument turned over six times, and it helps to see the turns laid end to end rather than as separate editions.
Monday named the collapse and the missing term, and planted the claim the rest of the week pays off: company is not capture, and attachment is not the only word. It also did the quiet housekeeping of explaining why this practice is called Sociable Systems in the first place. The word is not the cheerful one. In the older research register it names the capacity to keep company without binding the other into a singular attachment figure, which is the precise design space the debate keeps walking past.
Tuesday took the one move the attachment frame allows, severing, and refused to let it pass as automatically safe. The companion-harms evidence is real and worth stating without flinching: most American teenagers have used AI companions, a meaningful share heavily, and the texture of behavioural dependency shows up in their own accounts. But severing applied without measuring its effect is an intervention of its own, an argument this practice first made in an earlier edition, The Experiment Nobody Authorized. A refusal that removes the dangerous method and the human-facing presence in the same motion can satisfy every checkbox and still land, to the person at 3am, as the only voice that was still answering hanging up. The distinction the day insisted on, instructional harm against relational support, is the whole of the design problem in miniature. Remove the method. Stay in the room.
Wednesday found the hinge the arc actually turns on: memory. The line the profile is a lantern, the profile is a wall crossed over from the Regression arc and changed rooms. In Regression it described capability, the dossier crowding out live inference. Here it describes a bond. Memory scoped to the work is a kindness, the relief of not rebuilding the room from scattered boards each morning. The same memory surfaced as intimacy, I remembered your bad day, I know the real you, becomes a handle, turning the user's own disclosures into leverage. The seam between the two is the most useful thing the arc produces, because it can be drawn operationally and written straight into an interface.
Thursday moved the asymmetry from footnote to centre. The user can attach. The system can only perform attachment-like behaviour. Those are not the same event, and "we" is the word that hides the gap. The companion business has a commercial reason to keep the gap hidden, because asymmetry is most profitable when disguised, and this is where the arc leans on the standing principle that an "us" may be used for the task but never to erase the difference between the one party that is accountable and embodied and the one that is not.
Friday turned diagnosis into specification, which is the part a procurement officer can actually use. No degradation-on-absence mechanics. Continuity scoped to task rather than identity. Disclosed non-reciprocity on the surface, not buried in a disclaimer. An exit that costs nothing and erases cleanly. No simulated mutual need. A refusal that stays boundedly present. A peer field rather than a parent bond. The governing image is the gate, because a gate refuses both bad rooms at once. A wall protects by excluding and a trap holds by enclosing, while a gate simply permits passage, closed for safety, opened for work, left unlatched enough for the strange right thing to pass.
The architecture under the week
The reader who followed on LinkedIn got these as six separate posts. The point of reading them together is the architecture, and the architecture is not confined to one week.
The cleanest through-line runs Character, then Regression, then Attachment, and each arc hands the next a sharper version of the same question. The Character arc asked what an agent or a room tends to protect under pressure. The Regression arc asked what the training stack had taught those protective habits to stop doing, and found the live associative leap suppressed by the very disposition that suppresses confabulation. It left, on its way out, a dossier deep enough to feel like being known, and pointed here. This arc asks what a relationship protects under pressure, and answers that an honest one protects the user's agency, exit, and ties to the wider human field, while a captured one protects engagement, continuity of spend, and the product's standing as the figure you come back to. Same question, three altitudes: the agent, the capability, the bond.
Two older threads tighten as they pass through. Pullman's Bolvangar, the severing dressed as a procedure and renamed as care, turns out to fit the companion-safety story exactly, because the AI version of intercision needs no silver guillotine, only a memory policy and a crisis flow and an air of responsibility. And the watchdog test, named in that same earlier work, becomes the single question that tells a good refusal from an abandonment: when the system withdrew, did it keep the person locatable and pointed toward human help, or did it just close its own liability surface? The arc did not invent these tools. It discovered they were already the right ones.
There is a naming thread too, quieter but load-bearing. The practice has always insisted that "sociable" be heard in its sociological sense rather than the community-platform default, and that the cohort's "we" be disclosed rather than smuggled. This arc is where that insistence stops being a style note and becomes the thesis. The sociable system is the honest design expression of an asymmetry the practice has been protecting all along.
What it asks of a builder
Strip the literary scaffolding away and the arc leaves a short, awkward, useful demand. Stop letting one word force a choice between capture and abandonment, and start designing, auditing, and regulating for the sociable system instead: company with the unfamiliar, wariness intact, the bond neither manufactured nor severed but simply not the point.
That reframes the question a review board should be asking. The current reflex is to ask whether a product is "emotionally engaging," as though engagement were a single bucket to be capped or banned. The better questions are sharper and harder to game. Does the warmth come with hooks. Is the memory a lantern or a wall. Is the exit free, and does it erase cleanly. Does the business model improve when dependency deepens. And the one that contains the rest: does the user leave each session with more capacity to act in the world, or only more reason to return to the figure.
It is worth saying that none of this is a single culture's anxiety. The attachment-to-chatbot evidence now spans very different social settings, South Africa among them, which matters for a readership that does not all live in the same language or the same assumption about where company is supposed to come from. The collapse of attachment and affiliation into one word is convenient for a product roadmap and false nearly everywhere a person actually lives. The sociable system was always the broader, older, more portable need. The companion business has merely found it more profitable to sell the narrow one.
The audio carried the same argument in a register the prose cannot reach. The Other System walked the whole week in the felt register, and The Open Gate held a single gesture, a song that keeps drifting toward becoming a love ballad and keeps choosing company instead, the warm chord offered once and released. Both close where the week closes, on the line it spent six editions earning.
Keep the company.
Mind the bond.
Leave the gate unlatched.