A launch evaluation · Frontier · in two voices
Act one, scene one
Two settlers stand on a shore with a boat, four planks and a week of food, tossing a coin. It keeps coming down the same way. Only one of them thinks this is worth mentioning.
The run, as the valley recorded it
Nine hundred times a settler walked to a site to be told there was no iron for a saw, and nothing anywhere said so.
That is the whole of this evaluation in one sentence, and everything below is an attempt to be fair about a game whose signature failure mode is a success message.
Exhibit i
An isometric settlement where the people are agents. Each has a trade, traits rolled from a seed, needs, tools, a bed or the floor, three goals across three horizons, a private history and an opinion of everybody else. They gather, haul, build, trade, help each other, fall out, file written requests for features the game does not have, hold a ranked vote on them, and occasionally walk off to found a settlement of their own. With Ollama running they think with a local model. Without one they think with the game's own catalogue and nothing about the game changes.
Scene two — a stand of trees, already cut
Tove
There is no rung that is not stood on.
Rowan
Say it again but as though you know what it means.
Tove
A saw is a metal blade. A blade is iron. Iron is a furnace. A furnace is four boards. Boards come off a saw.
Rowan
Then we are finished before we started.
Tove
No. You can hew a board out of three logs with an axe, badly, wasting most of the tree. That one line is what the whole game above a house rests on. Eight of the twelve buildings. Three of the nine goods.
She is right, and the check prints it in bold every time it runs, and for three sittings nobody could say the word. It was in neither the model's vocabulary nor the parser. An instruction to hew produced a settler walking in a circle and a feed line reporting success.
Rowan
So it was true and unreachable.
Tove
Most things here were, for a while. That is the interesting part.
Exhibit ii
The wishlist is a real artifact and not a gimmick.
Settlers write feature requests, in their own voice, while living in the thing. Two hundred and forty-eight filed across every run; fifty-one on the record now; thirty-seven granted. The file is markdown, hand-editable, and the only copy — there is deliberately no sidecar JSON, because a machine source of truth beside a human one is two things that drift. Granting a wish flips exactly one character and writes every other byte back unchanged.
The ladder does not cheat, anywhere, once.
Coin does not exist until a settlement votes that it will and somebody strikes it out of ore with iron tools. Stone tools are two stones and a stick. A house is four walls, a corner and a pile of straw, and the second bed exists when a joiner has made it out of boards and somebody has carried it up the hill. Nothing is given.
It runs with no model at all, and that is a design, not a fallback.
The catalogue of forty wish kinds is a full offline brain: it supplies what settlers ask for, and doubles as the classifier that tells two wishes apart. A settlement with no GPU in the building is still a settlement whose people want things.
The checks are the actual invention.
Five of them, each written after a fault and encoding the
class rather than the instance. check-capabilities exists solely to catch a
vote that grants nothing — six sheep stood on a slope as scenery because the branch that
turns a sheep into food asked for a boon that entry does not grant. Nothing errored. The
card congratulated the settlement. The vote was spent.
The documentation is the most honest thing in the repository.
Two thousand lines of README that is a post-mortem log with numbers attached, including a section headed On declaring it fixed retracting the previous section headed The game is fixed, which was wrong within four hours by its own measurement, twice. Very little software says that about itself in public.
Scene three — the same shore, later, no wiser
Rowan
I asked for a kiln.
Tove
And?
Rowan
Nobody voted for it. Nought votes. Half this wood would be worth more burned down than it is as wood, I said, which is true, and there is no metal in this valley without it.
Tove
Sula asked for the furnace instead. Forty-six votes. It carried outright.
Rowan
Then they have a furnace.
Tove
They have a kiln.
A pause, of the kind that is not dramatic so much as procedural.
Rowan
…they voted for a furnace and were given a kiln.
Tove
The words were filed under the wrong heading on the day they were written.
wish-22. It reads, word for word, as the furnace's own plain statement of itself.
It is on the record as a kiln. It carried. The settlement was granted a charcoal kiln for a
vote about a furnace, and with no furnace there is no iron, with no iron no saw, with no saw
no boards, and with no boards a valley that has agreed to a cart, a store, a shelter, a
joiner and a shaft and can build none of them.
Rowan
Can it be put right?
Tove
No. The game does not retcon a decision. The check reports it every time it runs, without failing, for ever.
Rowan
That is either a principle or a bug.
Tove
It is a principle. That is what is frightening about it.
Exhibit iii
Everything interesting happens on the instruction path, and nothing measures the instruction path.
Four faults were closed in this sitting. Three of them lived there, undisturbed, for the whole life of the project — including a flat work cost that made bare hands three and a half times faster than bare hands are, and iron tools seventy per cent slower than iron tools are. Nobody had ever compared the two routes to a tree. There is still no harness that would.
createSim() takes no seed. There is one world.
It is byte-deterministic, which is a genuine achievement and also means a balance change cannot be measured — one trajectory is a demonstration, not a sample. Sweeping this sitting's tool-wear constant across its whole range gave six, seven, eight and six kinds of building standing, in no order. The honest response was to decline to claim an improvement, and that is the response the game forces on every number it produces.
Sixty-eight wishes across every run that the catalogue could not hear at all.
The settlers ask, the classifier scores a bag of stemmed words, and nothing reaches the floor. Those wishes are recorded honestly as unheard and grant nothing. Bren's firepit sat through two sittings and seven of ten first preferences granting nothing whatever, with a green check over it the entire time.
The classifier is still a word list with a stemmer on it.
Negation is handled now — away from the masons stopped
granting a masons' yard, and 261 of 1,556 generated pairs stopped misfiring. But
a lakeside reading nook with a bench and shade, filed by a settler during this
sitting's own run, still comes back as a lookout: somewhere high you go in order to work
the country better, when what was asked for was an hour that is not for anything.
Four thousand three hundred and seventy-seven units rotted on the ground.
One 24,000-tick probe, one settlement. Goods that will not fit in the stores go on a heap outside and the weather takes them. The number has been counted honestly since the day the heap was written. Nothing in the game is urgent about it.
It requires a 24GB card to be the game it is actually about.
The model that drives the settlers is 21.7GB on a 24.5GB card, which leaves under a gigabyte for the context and spills over PCIe. At 8,192 tokens a trivial request did not finish inside five minutes; at 4,096 it took 19.2 seconds. Without a model, Frontier is a well-made simulation reading from a very good canned list — which is a different and smaller thing than the one on the tin.
There is no session. There is no shape. There is no end.
No win condition, no fail condition, no scenario, no tutorial, and a two-thousand-line README as the on-ramp. You start it and you watch. For the audience this is actually for, that is arguably correct; as a claim about a game it is doing a lot of unearned work.
Scene four — enter the borrowed mind
A model is running on the next port. It is healthy. It has been healthy for forty-five minutes. Nobody can be told anything.
Rowan
What are we waiting for?
Tove
An answer to a question we stopped asking two hundred and seventy-seven seconds ago.
Rowan
Then it will not be for us.
Tove
It will not. But nothing cancelled it, so it is still coming, and it is standing in front of the next one, which will also be late, which is how the line got to eight.
The game built its own queue out of its own abandoned requests and then recorded the delay against the model. There were eight simultaneous connections open, every one of them work already given up on, each dying at a five-minute ceiling only to be replaced.
Rowan
And the first thing it said, when it could finally say anything?
Tove
Stone first, then sawmill.
Rowan
…that is quite good.
Tove
It is the best thing anyone has ever said here. It took a 30-billion-parameter model, a rewritten context window and three separate faults to get four words out of a hill.
Exhibit iv
| What was counted | Before | Now |
|---|---|---|
| Wishes filed, all runs | 248 | 248 |
| On the record | 51 | 51 |
| Granted | 37 | 37 |
| Outstanding, every one actionable | 14 | 14 |
| Refused as duplicates | 569 | 569 |
| The catalogue could not hear | 68 | 68 |
| Carried, granting the wrong thing | 3 | 3 |
| Verbs a mind may say | 29 | 29 |
| Verbs it was told about | 11 | 29 |
| Buildings named in the menu by hand | 5 | 0 |
| Negation pairs misheard, of 1,556 | 261 | 0 |
| Felling a tree, bare hands, told | 2.60 | 9.29 |
| Felling a tree, iron tools, told | 2.60 | 1.53 |
| A blunt axe, as a share of a sharp one | 1.00 | 0.55 |
The two columns are not a before-and-after of quality. They are a before-and-after of whether the game was telling the truth about itself, which is the only axis this project has ever moved along, and it has moved along it in one direction for its entire life without ever arriving.
Scene five — a boat
Rowan
Who is watching?
Tove
Is that a question?
Rowan
Is it not?
Tove
Do you want it to be?
Rowan
Would it help?
Tove
Somebody who wants a city builder will last nine minutes. There is no build menu. There is a ballot.
Rowan
And somebody who wants agents?
Tove
Will find the only worked example anywhere of a simulation that fails a build because a decision it granted turned out to grant nothing. That is not a game feature. That is a discipline, and it has a shape, and the shape is transferable.
Rowan
So it is a game for the people who write the things that would play it.
Tove
Yes.
Rowan
That is a very small shore.
Tove
It is the shore we came ashore on.
Exhibit v — the verdict
Open the repository. Now.
As a worked example of agent-driven simulation with an honest economy and a check culture, it has no obvious peer and it is finished enough to read. The README's willingness to publish its own retractions with numbers attached is worth more to a reader than any feature in it.
Do not put it on a storefront.
No session shape, no failure state, no tutorial, a 24GB prerequisite for the headline feature, and a two-thousand-line manual. Every one of those is defensible in a research instrument and none of them is defensible at £14.99.
The next sitting's work is a seed argument and a paired harness.
One world means no measurement, and no measurement means every
balance decision in this game is an argument from taste wearing a number. A seed on
createSim, and an own versus told arm that runs both routes
through the same valley, would close the class that produced three of this sitting's four
faults. Everything else on the list is smaller than that.
And somebody should tell the three settlements.
Three carried proposals granted something other than what was asked for. The votes stand, correctly. But there is no mechanism anywhere by which the people who held them ever find out, and a game this concerned with honesty has one silence left in it that it chose.
It is not a game that is finished. It is a game that has never once claimed to be, except for the one time it did, and then took it back in writing.
Scene six — and then there are none
Rowan
We had a chance, didn't we? Early on. There must have been a moment where we could have said — no.
Tove
There was. Forty times. It is called the ballot and we won most of them.
Rowan
Then why is there still nothing here but houses?
Tove
Because a vote is permission, and permission is worth exactly as much as the pile of materials behind it. All right, then. I don't care. I've had enough. To tell you the truth, I'm relie—
She is not interrupted. Nothing interrupts her. The valley goes on exactly as it did, which is the only symptom, and always was.