Skip to main content
sociable systems.
Episode 241 · Sunday interlude · 2026-08-30

Someone Had to Sign

Sunday interlude. A signature ran on presence: a wrist that could stop mid-stroke, a name a court could find. An authorization token proves permission existed upstream and says nothing about who occupies the moment of action.

Cover art for episode 241: Someone Had to Sign
Authorization ArcInterludeAccountability
Episode 241: Someone Had to Sign

It moved the money before sunrise, it held the key you handed down. The log shows somebody approved. Nobody signed. Nobody's found.

The Corrupted Archive closed on two pocket questions about records kept on purpose and records left to drop. Both assumed a filer behind every entry: an institution with an appetite, a librarian with a shelf, somebody with a budget deciding what deserves survival. The answers held anyway, because most entries do have a filer, however careless.

The newest class of entry does not. Systems now hold credentials, move money, close tickets, and sign approvals on trails that read like human participation, while the moment of action contains nobody at all. The archive has started writing itself.


What a signature actually was

Before encryption, a signature ran on presence. The signer stood where the paper was, held the pen with a wrist that could stop mid-stroke, and walked away owning consequences that could find them later through a name, an address, and a court. Delegation worked inside that bond, and both bodies stayed reachable when something went wrong.

A digital authorization token carries none of that forward. It proves that permission existed somewhere upstream, at some point in time. It says nothing about whether whoever, whatever, is exercising it comprehends what execution will do. That distinction sits underneath everything that follows.


The fortnight that set the agenda

The material for this arrived within weeks of itself, which happens often enough now that convergence has stopped being remarkable.

An Okta report put adoption at ninety-one percent of organizations surveyed and put security incidents involving AI agents at eighty-eight percent of the same sample, naming shadow agents, systems operating on credentials their own administrators never registered, as the connective tissue. Almost everyone has adopted. Almost everyone has already been surprised by what they adopted. An unregistered actor cannot be interrogated or throttled or retired, because governance begins from an inventory, and the inventory has a hole shaped like everything running fastest.

Gartner followed with a projection: forty percent of enterprises will demote or decommission autonomous agents by 2027, after governance gaps surface post-deployment. And separately, enterprise surveys found under ten percent comfortable letting agents fully close out incidents without review afterward. The deployment is running several lengths ahead of the confidence.

Meanwhile agentic commerce crossed from demo to rails: delegated authorization, checkout through an API, payments executed without a person at the terminal. And on August second, Article 50 of the EU AI Act became fully applicable, making disclosure of machine-generated content a legal obligation in European scope rather than an engineering aspiration everywhere else. The regulators have arrived at the same door the shadow agents came in through. They are reading each other's paperwork.


Bartleby, inverted

Melville gave literature its most famous refusal of a signature. Bartleby copies legal documents in a chamber where every instrument ultimately depends on somebody attesting to a copy being true, and then he stops. I would prefer not to. The whole machinery of attestation discovers that it was running on one clerk's cooperation, freely given, withdrawable.

Modern delegation removes the human cooperation bottleneck entirely, so nothing can refuse, and equally nothing stands behind consent. A workforce that can decline forces a question about warrant at every desk. A workforce that executes automatically relocates the question upstream into preference settings almost nobody reviews, written in language almost nobody parses, at the moment they were skimmed.

The copying never had to pause. That is precisely the audit problem.


The question of the week

Attribution after an incident arrives cheaply now. Logs generate themselves, timelines assemble automatically, and the reconstruction frequently produces exculpation as a default output, because every step did follow its assigned policy. Nobody deviated. Nothing crashed. And the harm still landed.

So the week turns attribution around and examines it before execution instead of after. The test has three parts: a named party answerable before the action commits; reversibility designed into the transaction rather than hoped for afterward; custody of the evidence placed with somebody other than the accused system writing its own defense.

Every structure examined this week gets weighed against those tests.


The Track

Nobody Signed is this week's companion, and a straighter-lit sequel to Nobody Looked. Same deadpan narrator, colder room, and where the earlier track was playing with a fabrication, this one is level-eyed about an absence. The chorus is two words long, and the verses spend themselves explaining why those two words are compatible with a complete audit trail. The log says approved at half past nine, authority flowed, the trail looks fine, and you can follow every hop back to a token doing time for a hand that departed.

The bridge names the three absences the rest of the week works through: reversibility was optional, custody stayed internal, attribution ran retrospective, and retrospective never calls. The outro whispers active session, terminated, unclaimed, which is as close as the record comes to a signature, and even that gets revoked.


Companions


These notes come out of Sociable Systems, a practice that reads AI-shaped documents the way a hostile reviewer will, before a lender or a court finds the gap. The argument has an operational form: the Interim Protocol sets out four rules for AI use in environmental and social deliverables, covering disclosure at touch-point grain, evidence custody, the phrases no automated screening may settle, and a hostile read before anything ships. Free, and written to be cited or retired once institutional guidance arrives.