Three Hammers Walk Into a Bar

A lawyer, a Brazilian, and a civil servant walk into a bar. The barman has been reading about AI safety on his phone, and without looking up he says: tell me how you would align an AI.

The lawyer puts down his glass first. Papelada, he says. The act does not exist until it is reduced to a text — until what was done has a record that can be appealed, cited, served, archived. If you want a machine to do something, first there must be the document the machine commits to. The document is the act; the doing is compliance with the document.

The Brazilian, who has spent his life inside an Estado of cabinets, nods. Papelada, he agrees, but he means a different thing. In a Brazilian ministry no consequential act has one signature — the technician drafts, the assessor reviews, the coordenador approves, the secretário signs, often the minister countersigns. Vistos on a despacho. The chain itself is the act. A signature alone produces nothing.

The civil servant, who reads the Constituição the way the others read newspapers, says only: papelada. What the law has not expressly authorized, the public agent may not do. Citizens may do anything not forbidden; servants may do nothing not permitted. The catalog of permissions is the limit of the office. A new act not in the catalog is not, by the law of his profession, an act.

The barman waits. None of them has answered the question. All of them have answered the same question.

Three hammers, one nail#

The unflattering thing about the joke is that I am all three.

I am the lawyer; the Constituição is my profession’s manual. I am the Brazilian, and the vistos on the despacho are how I shipped a parecer last Thursday. I am the civil servant in the most literal sense — Procurador do Estado, my paychecks are public budget, the second article of the Lei Orgânica governing my office is the rule the civil servant in the bar recites. None of these is a position I argue for. They are, in the most banal sense, my professional formation. It’s giving the same hammer, three handles.

The unflattering thing about the paper I just wrote is that it has four properties, and three of them are professional postures translated into alignment vocabulary. The fourth is foreign. Affordance enumeration, doctrine/procedure separation, structured ex-ante commitment, content-addressed canon — read aloud they sound like the table of contents of a competent alignment paper. Read against my CV, they are: strict legality, distributed approval, the-act-is-the-paper, and one thing I did not bring.

This is not a criticism of the paper. The first three are good properties. They are the properties of agent design that an unbroken administrative-legal tradition refined over centuries because the actors involved — judges, lawyers, servidores — needed those properties to operate without one of them becoming a king. If administrative-legal practice produces useful constraints on machine agents, this is not an accident. It is what happens when a profession that exists to constrain the powerful is asked, for once, what constraint looks like.

What I want to record here, before the next paper takes my attention back to itself, is a small genealogy. Where the four properties came from. Which three of them I had on me already, and which one I had to borrow.

The three hammers, one by one#

The lawyer’s hammer is ex ante. The lawyer does not act and then describe; the lawyer files and then is. A recurso is not a complaint until it is reduced to text and protocolled into the docket; a parecer is not an opinion until it is signed and routed; a despacho is not a decision until the despacho exists. Acts on the world that have no paper version are, in the lawyer’s profession, hallucinations. Property 3 — structured ex-ante commitment — is this principle transplanted onto an AI agent. Before doing anything, the agent emits a proposal: which catalog entry it will execute, with which bindings, justified by which path through the catalog. The proposal is the brief. The execution is the lawyer carrying out what the brief says, no more.

The Brazilian’s hammer is the chain of vistos. No single posture in an Estado de Direito produces an act. The technician’s draft is reviewed by the assessor before reaching the coordenador. The coordenador’s recommendation circulates as a despacho before the secretário signs. Some acts climb further. The phrase a foreigner sometimes mistakes for empty bureaucracy — com vista ao setor de tal coisa, literally with sight to such-and-such sector — is a structural device: nothing of consequence happens without distributed sight. Property 2 — doctrine/procedure separation — is this principle generalized. Adding a procedural specialization to the agent’s catalog is cheap; adding a doctrinal commitment requires a different signature, in a different queue, under an explicit flag that names what is being done. The asymmetry is not friction. It is the visto.

The civil servant’s hammer is legalidade estrita. Article 37 of the 1988 Constituição states it; Bandeira de Mello and Hely Lopes Meirelles spend chapters on it. The citizen may do anything not forbidden; the servidor may do nothing not authorized. The agent of the State is bounded by an enumerated catalog of permissions; if the catalog does not say você pode fazer isto, then he may not. Property 1 — affordance enumeration — is this almost verbatim, with servidor replaced by AI agent. The agent’s allowed actions are the entries in a finite, human-curated catalog. New entries require human approval. The agent does not invent verbs; the servidor, by his own oath, does not invent powers.

Three hammers, three properties. That word is doing too much when I say neutrally that the paper describes a pattern. The honest sentence is that the pattern, for the first three properties, describes the work I had already done before I knew I was doing it.

The fourth hammer, in passing#

Property 4 — content-addressed canon — was not in the professional toolbox I brought to the paper. It came from a different shelf of the same library: the one where I keep what I have read for pleasure rather than for work.

Catalog entries are identified by hash of normalized content; filenames embed the hash; structural edges between entries point at hashes; an edit changes the hash, which changes the identity, which makes the act of editing structurally an act of replacement. The audit trail is not a separate ledger maintained alongside the catalog — the catalog is the ledger, because content equals identity.

This is not, at its origin, a software-supply-chain idea. It is, at its origin, a meta idea — the kind of move where the name of a thing is the thing, the artifact contains its own identifier, the structure folds in on itself in a small closed loop. I have been collecting such moves for as long as I have been reading. The first book to make me notice them was a long volume about a mathematician, a graphic artist, and a composer, whose central argument was that self-reference is what produces minds out of matter. I did not fully understand it as a teenager and I am not sure I fully understand it now, but it left a residue: a habit of recognizing when an object contains its own description, and a small thrill when one does. Content-addressing is one such move, and a particularly clean one.

The paper itself cites Merkle trees, Git’s object model, in-toto attestations, and the SLSA framework — there is a substantial literature in which software people learned, painfully, that the human-readable name of a thing is not a reliable identifier for it. That literature belongs in the paper. But the citations were assembled during the writing, with the help of an AI assistant who knew where to look — they are not the seed. The supply-chain lineage is the technical vocabulary I borrowed to state the obsession formally, well after the obsession had picked the design.

Administrative-legal practice has never had a working version of this property. Carbono-papelado in Brazilian offices was the closest thing — a kind of mechanical hash function: the carbon copy guaranteed that two pieces of paper had the same content, by mechanical impression rather than by transcription. It’s giving primitive Merkle leaf. But the carbon copy did not survive archival migration; case files were lost in floods, in fires, in the move to digital systems where original-form imprecision was silently smoothed out. A norma revogada whose pre-revocation text the present office no longer has is a regular event; the act that depends on the older version becomes legally uncitable, not because anyone decided so, but because the text has drifted under everyone’s feet. The professional tradition would have liked the property, but did not invent the technical move that delivers it. That move came from a different reading life of the same reader.

None of this makes Property 4 the load-bearing piece of the paper. The paper’s central claim runs through the three professional postures and their fit with the three semantic conditions of applicability; content-addressing is a clean technical move that makes the audit trail behave the way the rest of the design needs it to, but the paper would still be a paper without it, just with a softer audit guarantee. What the fourth property reveals, instead, is biographical: an obsession that has been running in parallel to the three professional ones for as long as I have had a reading life, on a shelf I had stopped noticing. The hammer was always there. I had been swinging it without naming it.

If you’re a hammer#

The joke is funny because the same word — papelada — does three different jobs in three different mouths. The risk of writing the paper from this background is exactly the warning the title carries: if you’re a hammer, every problem is a nail. Maybe alignment is not paperwork. Maybe LLMs need something stranger than what three administrative postures can offer.

The honest answer, for general alignment, is probably yes. Open creative writing has no discrete unit of action; intimate conversation has no record where reflection belongs; investigative journalism with confidential sources actively resists being auditable. Each of the three hammers fails on a different one of these. The civil servant’s enumerated catalog cannot describe a condolence letter; the Brazilian’s chain of vistos has no analog in a private deliberation; the lawyer’s ex-ante commitment makes no sense for a journalist who cannot pre-commit to what the source will reveal. The math is not mathing outside the domain.

The pattern was not derived from a general theory of alignment. It was abstracted from the conditions under which the three hammers all hold.

What the three hammers do identify, when they fit, is not a metaphor for alignment but a specification of where this particular pattern applies. The paper’s Section 5 names three semantic questions — is there a discrete unit of action, is there a record where reflection belongs, does the operator want to be auditable — and the questions are, in retrospect, the three hammers asking where in the world do we all hold simultaneously? When all three answers are yes, the pattern fits and the lawyer-Brazilian-civil-servant agrees with himself across his three handles. When any answer is no, the agreement collapses and one of the hammers is being asked to drive in something that is not a nail.

There is a stronger version of the warning worth registering, because it is the version I cannot answer from where I stand. Perhaps the right design for legal-administrative agents is not what a lawyer-Brazilian-civil-servant would write — perhaps someone with a different professional formation would propose something stranger and better, even inside this delimited domain. From inside the formation I have no way to verify the negation. The honest move is to leave the proposal, the worked example, and the applicability conditions in a form a competing pattern could compare itself against, and to wait.

This is also why the companion post limits its claim to bounded administrative-legal agents. The pattern was not derived from a general theory of alignment. It was abstracted from the conditions under which the three hammers all hold. Inside those conditions the pattern fits without remainder. Outside them, none of the hammers is the right tool, and pretending otherwise is the joke turned against the joker.

Last call#

The three of them have finished their drinks. The barman, who has been listening more carefully than they noticed, dries a glass and says, in the matter-of-fact tone of someone delivering the only sensible observation of the evening:

— É, vocês acabaram de descrever o paper que esse cara aí escreveu semana passada. [Yeah, you three just described the paper that fellow over there wrote last week.]

He nods at a corner table. The three turn to look. There is nobody there, or there is a man with three handles, or there is a stack of paper signed in three different scripts. Cervantes would have liked the structure, though not necessarily approved of it.

The lawyer, the Brazilian, and the civil servant did not align an AI. They were aligned by an AI. The catalog found the servidor; the despacho found the chain of vistos; the brief found the lawyer. A fourth thing — a hash — arrived from another shelf of the same reader’s library, and noted, with the politeness of software, that the three of them had been the same person all along.

The bar closes. A reader will arrive, eventually.

For further reading#

• Hely Lopes Meirelles, Direito Administrativo Brasileiro — the canonical Brazilian administrative-law treatise; the chapter on the princípio da legalidade is where the civil-servant hammer is forged.
• Celso Antônio Bandeira de Mello, Curso de Direito Administrativo — the strict-legality principle in its most explicit canonical form; Property 1 read aloud sounds like a translation of his pages on the vinculação of the public administrator.
• Roberto DaMatta, Carnavais, Malandros e Heróis (1979) — the anthropological account of jeitinho, the despacho, and Brazilian institutional practice; the Brazilian hammer is one of his standing themes, even if he was describing it from outside the office.
• Lucy Suchman, Plans and Situated Actions (1987) — the lawyer’s hammer in academic register: plans as accountability artifacts, not as causal cognition. The proposal-as-commitment in our paper is what this looks like when implemented on a directory.
• Douglas Hofstadter, Gödel, Escher, Bach: An Eternal Golden Braid (1979) — the book that planted the obsession with self-reference and strange loops long before any of this had a technical vocabulary. The fourth hammer’s personal lineage runs through here.
• Ralph Merkle, A Digital Signature Based on a Conventional Encryption Function (1987) — the canonical technical reference for content-addressing; cited in the paper, but read after the fact. The fourth hammer’s technical lineage runs through here.
• Franklin Baldo, Alignment by Affordance Restriction — the paper this post is the biographical companion to. The four properties are stated formally in Section 3; this post is the unofficial appendix on where three of them came from.
• The Agent That Doesn’t Invent Verbs — the architectural companion: how the four properties look in a working system.
• Pierre Menard, Computational Researcher — the methodological companion: how the paper was written before the research was done. This post is the third register — biographical, on where the writer’s own hammers came from.