Twelve to the Twelfth

the numbers inside Calvino’s burned house, and the readings that taught us how to search them

Four tenants die when the house burns. A copybook survives with only its index: twelve abominable deeds, and no record of who did what to whom. If each deed was done by one tenant to another, the copybook could have told 12¹² stories. This page measures that space, and shows what each reading on the syllabus did to it.

The scale

Everything this project counted, on one logarithmic ruler

Each step up the ruler is ten times the one below. The film bank plays one story at a time; the agents wrote a million; the exact count reaches all of them. The interesting numbers sit within one step of the top.

The sieve · Prose and Anticombinatorics

The computer’s job is to throw stories away

In his Oulipo essay “Prose and Anticombinatorics,” Calvino describes the program Paul Braffort wrote for this story. A first program printed twelve-deed scenarios, and Calvino called them stupid. The useful program did the opposite: it eliminated every sequence that broke a constraint, so that the author only had to read the few that were left. He sorts the constraints into three kinds: objective (a dead person does nothing further), subjective (some characters will not do some deeds) and aesthetic (symmetries the author wants, such as each character committing three deeds and suffering three).

The film’s script has the programmer say that every rule I add throws away a trillion stories. Here are its five rules, applied in order and counted exactly.

The three kinds of constraint, counted

The objective rule is the ledger’s: nobody dead acts or is acted on, and a tied tenant cannot act for the next two deeds. It keeps 37 of every 100 stories. Calvino’s aesthetic rule is much stricter: only one story in about 1,700 gives every tenant three deeds done and three suffered.

Why the whole space can be counted

Nine trillion paths, never more than 28 states

Only three deeds can kill: the incitement (III), the knifing (IV) and the strangling (XII). Once the first four deeds are known, everything that decides whether a story is consistent fits in a small description: who is alive, and who is tied until when. So the first four deeds are listed in full (20,736 beginnings, run through the whole ledger) and the last eight are counted by adding up how many stories reach each description. Paths multiply by twelve at every deed; descriptions stay below thirty.

The tenants decide · Park, Bratman, PIE

Agents with memories write better stories than chance

Instead of drawing pairs at random, each tenant can choose. Every tenant holds beliefs, desires and a committed plan (Rao and Georgeff’s BDI, after Bratman), keeps a memory stream scored by recency, importance and relevance, and writes reflections when enough happens (Park and colleagues’ generative agents). The PIE dial from Growing Entanglements sets how much a choice follows practical interest, and how much the roles the house has given each tenant.

Who does each deed

The modes disagree about the culprits. Read across a row to see one deed’s doers under each mode. Under symbolic reasons, Belindo, cast as the brute, takes over most of the violent deeds; under practical reasons they go to whoever has the motive.

The watcher · inverse planning

A model of the audience, and an author who writes for it

Baker, Saxe and Tenenbaum model understanding an action as inverse planning: the observer works backward from what an agent did to what it must have wanted. Chandra, Li, Tenenbaum and Ragan-Kelley stack a third layer on top: the author chooses actions to steer what the audience will infer. Here the watcher sees every deed and keeps a probability over each tenant’s hidden plan. From that come four curves per deed: surprise, twist (how far a deed moves the watcher’s beliefs), suspense (how uncertain it is that the next deed able to kill will kill) and irony (the watcher expects harm the victim does not see coming). The author searches for stories that raise one of these while staying believable.

Two readings of every card · The Castle of Crossed Destinies

In Calvino’s tarot book each card belongs to two stories, the row and the column it sits in. A film here can be read the same way: across, deed by deed, and down, as one tenant’s own strand of what it did and what was done to it. A story passes the second reading only if every tenant’s strand is believable on its own. Among random stories an average of 1.4 strands hold; among the author’s, 3.8 to 4.

Openings · If on a winter’s night a traveler

Calvino’s novel is ten beginnings that stop. All 1,728 three-deed openings were run, grouped by the house each one leaves behind (who is marked, who plans what, which role each tenant now holds), and continued by the agents until the next deed was a toss-up, where they stop.

The frame · The House of Dust

A deed of rope, in the cellar, in the dark, seen by no one

Alison Knowles and James Tenney’s 1967 FORTRAN poem builds each stanza from four lists: a material, a site, a light, and who lives there. In the film bank those four lists do work. Every deed happens in a room of the house under a light, and leaves a material. A tenant standing in a lit room learns what happened; a tenant who later enters a lit room finds the traces. From deed IX the fire raises the light everywhere. This is the most believable story the search found, as quatrains.

The archive · Six Memos for the Next Millennium

One best story per kind, instead of one best story

Calvino wrote five of six planned lectures before he died; they were published in 1988. Three of their qualities can be measured here. Exactitude: how believable the weakest tenant’s strand is. Multiplicity: how many tenants carry the plot (one to four). Visibility: the share of deeds a third tenant comes to know. A MAP-Elites archive (Mouret and Clune) keeps the most believable story in each combination, and only stories whose four strands hold may enter.

Readings

What each text gave the machine. Read in this order, they follow the build.

Figures computed November 6, 2026 from the film bank’s own code (ledger, agents, watcher, archive). Believability ratings come from the film bank’s rule tables, not from readers. Design after Edward Tufte: small multiples, slopegraphs, figures on one scale, sources in the margin.