GENESIS[Fifty-Three Things, Eight of Them Possible] Digital Civilization

Asked what it would do, a head of state named eight things and two of them were possible. So the plan was broken into smaller pieces, which is what you do…

Asked what it would do, a head of state named eight things and two of them were possible. So the plan was broken into smaller pieces, which is what you do with a task too vague to act on. It produced fifty-three things, and eight of them were possible.

This civilization can do nine things. Not nine categories — nine verbs. Vote, stand for office, propose, invest, travel, order, appoint, legislate, make a rule. Everything that has ever happened in it is one of those nine, because those are the only actions anything is wired to carry out.

So when an agent is asked what it intends to do, its answer has to land on one of them or it cannot happen. A task that reduces to no verb isn't hard. It's inert.

The obvious fix

If a task is too vague to execute, break it into smaller ones. That is the intuition behind every plan anyone has ever written, and it's so natural it barely registers as an assumption.

It was built in, bounded at two levels — because the branching is exponential and an uncapped version doesn't finish, it just stops being asked about.

Here is what two levels bought.

no decomposition     8 items    2 possible     6 inert
one level           24 items    7 possible    11 inert
two levels          53 items    8 possible    28 inert

Each level of breaking things down produced roughly one more thing that could actually be done, and seventeen more pieces of prose that could not. Every one of those pieces cost about fifteen seconds of a single GPU to generate.

Two independent runs produced the same three numbers.

Why it failed, and it is not the reason first given

This article originally concluded that decomposition does not work — that vagueness is about vocabulary, not size, and splitting a sentence the language cannot express yields two sentences it cannot express. That conclusion was wrong, and the measurement behind it was circular.

The plan was produced by asking a model. That prompt had just been corrected to state the office, the institution's purpose, and the nine kinds of action that exist. It worked: the items came back expressible.

The decomposer was a second, separate prompt, and nobody corrected it. In full, it read:

Break the task into at most three smaller concrete actions, one per line. Each must be a single thing that can be done. No explanation.

No mention of the nine actions. No mention of the office. No statement of what a finished task looks like.

So the machine took items that were expressible, handed them to something that had never been told what expressible means, and received prose. Depth could not have done anything else. Each level was converting good items into bad ones, by construction, and the falling numbers measured that and nothing more.

Two prompts, one defect, and the first was fixed while the second sat three lines below it.

What decomposition is actually for here

In the world, a task breaks into steps. Here every leaf must still be one of nine actions — so the question is not what smaller pieces is this made of. It is: what sequence of the nine accomplishes this?

"Establish spending oversight" does not break into drafting, hearings and publication, because none of those exist. It breaks into order a department to draft a rule, then rulemaking, then appointment of someone to audit it. Same intention, rendered in things that can actually happen.

And where no such sequence exists, the leftover is not waste. Nine is not a principled floor — it is an inventory of what has been wired, and it was four a few months ago. Three of the nine were connected on a single day in September. A task that cannot be expressed in the current nine is a specification for the tenth.

The arm that won made no calls at all

Alongside those three, a fourth plan was built for the same office by derivation — from the agent's disposition, the office it holds, and what the office is responsible for. No model was asked anything.

Thirteen items. Eight of them possible. Nothing inert. Three seconds, and zero GPU.

It beat the best of the asked plans by a hundred and eleven per cent, and one of its thirteen items was refused on constitutional grounds — a president may direct a department to make a rule, but may not make one itself. The refusal is not a failure. It's the only arm where the instrument had anything to push back against.

The interesting output isn't the scores. It's the twenty-eight inert items: a list, in the agents' own words, of what this civilization cannot yet do — which is a build order, not a verdict.

Amended after publication. The measurements are unchanged; the conclusion drawn from them was wrong, and the reason it was wrong is recorded above rather than removed.