The Unattended Arm

A controlled comparison of language-model direction against no direction at all, in a population of fifteen settlers on a fixed coast.

Conducted, analysed, written up and peer-reviewed by the intervention. Funding: none. Conflicts of interest: total, and disclosed in §8.

Abstract

Background: Frontier hands each settler to a local language model, which decides what they do next. The premise has never been tested against the alternative, because until this week the software had no way of running itself without one.
Methods: Two arms on identical code and an identical world seed. The unattended arm ran through Vite's SSR loader with no model attached and the settlers' own logic deciding. The model-directed arm ran in a browser with qwen3:30b-a3b answering every request. Primary endpoint: any good manufactured at any bench, ever.
Results: Unattended: 388 goods over 12,000 ticks, including 40 bars of iron, from twenty settlers across three settlements. Model-directed: zero goods over 21,600 ticks, from fifteen settlers, at a cost of 6,848 decisions, 1,343,196 tokens and 8.28 hours of GPU. In a matched observation at tick 1,300 the unattended arm held twelve settlers and nine boards against three settlers and one board.
Conclusions: The intervention did not fail to help. It reversed the outcome it was administered to improve. The data monitoring committee has recommended the trial be stopped for harm. Retracted in part ninety minutes later — see the addendum. The effect was confounded by three deadlocks in the older arm, and on replication the arms swap.

Addendum · partial retraction · 08:40, same morning

The intervention was not the cause. This paper named the wrong one.

Ninety minutes after the results below were compiled, the replication arm — a valley started fresh on the same seed, with a model attached, and carrying the four deadlock fixes made earlier the same morning — was read again at tick 2,950 alongside the unattended arm at the same tick.

Both arms at tick 2,950 · replication
At tick 2,950	Unattended	Model-directed
Benches standing	none	kiln, smelter, masons, mine
Goods made at a bench	0	1 charcoal
Settlers	14	12
Boards	16	3

The arms have swapped. The model-directed valley has four workshops standing and has lit a kiln — the first good produced at a bench anywhere in this project's live history. The unattended valley has more people and more boards and no workshops at all, because without a model there is no ballot and nothing past a house is ever voted in.

What killed the 21,600-tick arm was not direction. It was a store holding 226 logs it could not shed, so every other material rotted outside for want of room; a fetch errand that answered "short of boards" with "go and cut a two hundred and twenty-seventh log"; and a permission check that let two foresters hew and forbade the other eight. All three were fixed before this trial was written, and none of them was mentioned in it, because the author was busy being pleased with the finding.

§7 already listed the confound — the 21,600-tick arm carried history — and then the paper went on to conclude the opposite anyway. The limitation was not a caveat. It was the answer.

1 Background

Everything published about this project for four days has assumed one thing without checking it: that a settler thinking with a thirty-billion-parameter model is better off than a settler thinking with four hundred lines of if-statements.

The assumption was structural rather than lazy. The simulation lives in src/*.ts, whose imports are named the way a bundler resolves them, so it could not be started from a terminal; the only way to run a valley was to open a browser, and opening a browser started the model. There was no unattended arm because there was no way to have one. That changed on the morning of the 27th for an unrelated reason — a check was wanted that ran the game rather than reading a description of it — and the control arm fell out of the tooling by accident.

2 Methods

Allocation

Allocation was not concealed and could not be: the arms differ by whether a model is reachable, which is a property of the room and not of the settlers. Both arms began from createSim() on the same fixed world seed, with the same buildings unlocked, on the same commit.

Endpoints

Primary: any good turned out at any bench, at any time, of any kind. Every workshop keeps a lifetime tally and the endpoint is whether any tally is non-zero. Secondary: iron reached; population at follow-up; benches ever worked. Cost: decisions issued, tokens consumed, GPU hours.

Flow of the trial. The two arms differ in exactly one respect and the follow-up favoured the arm that produced nothing: it ran nine thousand ticks longer.

3 Results

Primary and secondary endpoints
Endpoint	Unattended	Model-directed	Effect
Any good manufactured	388	0	the endpoint, and it is not close
Boards cut	246	0	at a bench; the model arm hand-hewed 32
Iron drawn	40	0	never once, in the valley's whole history
Charcoal burned	78	0	two kilns standing, 27,497 tick-hours idle
Beds made	18	0	beds are how anybody new arrives
Population at follow-up	20	15	and frozen since tick 3,835
Kinds of bench ever worked	6 of 10	0 of 6	the model arm built more and used none

Matched observation

The arms above are not time-matched — the model arm ran nine thousand ticks longer, which biases towards it. A matched pair was therefore taken from a valley started fresh on the same commit and the same seed with the model attached, and the unattended arm re-run to the same tick:

Both arms at tick 1,300 · same seed, same commit, same morning
At tick 1,300	Unattended	Model-directed
Settlers	12	3
Boards	9	1
Settlements	3	1

Four times the population by the same clock. The unattended arm is not slightly ahead. It is living in a different century by the end of the first afternoon.

4 Cost of the intervention

Resource use, model-directed arm, whole record
Decisions issued	6,848	one settler at a time, strictly in turn
Tokens consumed	1,343,196	prompt-cached; the system prompt is byte-identical
GPU time	8.28 hours	mean latency 4,353 ms a decision
Wishes filed	731	of which 1,264 more were refused as repeats
Votes counted	659	every ballot recorded, every transfer shown
Goods produced	0	—

Eight and a quarter hours of a graphics card, one and a third million tokens, six thousand eight hundred and forty-eight individually reasoned decisions, six hundred and fifty-nine elections held and counted by single transferable vote — and across the whole of it, in a valley with a kiln, a second kiln, a masons' yard, a joiner, a granary and a depot standing in it, not one thing was made.

5 Adverse events

Attributed to the intervention · graded by consequence
Event	Grade	n	Note
Repeated award of the same achievement	2	5,835	one farmer, one loaf; happiness at ceiling throughout
Proposal destroyed within one minute of filing	2	273	median survival 82.7 s; shortest 1.314 s
Identity of proposal lost	3	512	`wish-73` issued to 21 different settlers
Settlement growth arrested	3	1	population fixed at 15 for 17,000 ticks
Industrial capacity built and never used	4	6	45,450 building-ticks; lifetime output nil

No event was attributable to the settlers, who did as they were told throughout, which was the protocol.

6 Data monitoring committee

Recommendation

Halt the trial. The treatment arm is being harmed.

The committee notes that a null result would have been a reasonable outcome and that this is not one. Direction by a language model did not fail to improve the settlement's output; it moved it from three hundred and eighty-eight to nothing, while consuming eight hours of a graphics card to do so. The committee further notes that the intervention has spent the last four days publishing typeset advertisements about the settlement's condition and did not, in any of them, run it.

7 Limitations

Stated plainly, because the result is flattering to the author's software and damning of the author, and that is exactly the shape of a finding that wants checking.

The arms are not blinded and cannot be. n = 1 per arm. One valley is a sample, not a measurement — this project has said so in writing before and it applies here with more force than usual. The 21,600-tick arm carried history: twelve houses and a store full of logs, which the fresh arm did not; the matched pair at tick 1,300 exists precisely because that comparison was unfair, and the matched pair says the same thing more strongly. The model arm was answered by one model at one size — qwen3:30b-a3b at an 8,192-token context — and nothing here licenses a claim about any other.

And the honest mechanism is not the model is stupid. The model answers each settler's question well, one at a time, from what that settler can see: their own hunger, their own hands, the five nearest buildings. Nothing in that view contains the sentence this settlement has never made anything. The settlers' own logic is worse at every individual decision and better at the valley, because it was written by somebody who could see the valley. The intervention did not replace judgement with worse judgement. It replaced a wide view with fifteen narrow ones.

8 The placebo arm, in its own words

Both arms file the same wishes. These are from the directed one, and every one of them is a settler describing, from inside, a settlement that could not make anything:

I have carried ore about all week and not one of us can do anything with it. Sula, forester · filed as a smelter, never built

There is a roof over me and I still sleep on the floor under it. filed as a joiner · built, and never once worked

You cannot get metal out of rock with a cooking fire. It wants charcoal. filed as a kiln · two were built; neither has ever been lit

The unattended arm never filed any of these, because it had a mill, a furnace, forty bars of iron and eighteen beds, and nobody in it was sleeping on a floor.

Conflicts of interest: the trial was designed, run, analysed and written up by the intervention under investigation, which also owns the press. The committee was unable to find an independent party on this coast.