The world
Layer three of the Embryo
A simple world where the work’s own claims about cooperation and defection can be run, and seen.
What it is
The world is made of operators. In the work’s own words:
An operator is any sovereign agent that applies control input to hold its structure against entropic drift — a person, a bacterium, a corporation, a civilisation. The definition requires no consciousness; it requires a budget, a drift, a corridor, and an exit.
From AP02 The Operator, “What this paper does”
AP and a number name a paper of the work. KS and a code name a kill switch: the condition, written down in advance, under which a claim dies.
In this world, each operator holds a stock. The stock runs down by itself, all the time. What holds it up is what comes in: what others give it, and in some worlds a little of its own. If it falls to a line, the operator is ruined, for good.
Operators are joined in pairs by links. On a link, each side can give to the other, and what one gives is worth more to the other than it costs the giver. So two who both give each gain. Every link also costs a small fee to keep open.
There are three kinds of operator:
| Kind | What it does |
|---|---|
| The careful giver | Gives on every link. Watches which way its own stock is going, and the record of every link, and cuts a link that costs it more than it brings. |
| The giver who never stops | Gives on every link, and never cuts. |
| The defector | Never gives. Keeps every link it can. It cannot choose to do anything else. |
Every record is open, and exact. A record is what happened. No link joins two defectors: nothing passes between two who never give, so there is nothing to link.
Built on the work’s conditions
The world’s rules are AP02’s own wherever AP02 fixes them: its drift, its noise, its limit on effort, and its ruin, which is permanent. Where AP02 does not fix a rule, the world makes a choice, says it is a choice, and makes it a setting anyone can change. The largest is how a careful giver decides when a link costs it more than it brings. The exact conditions under which a link is net-positive are what AP02 names as still owed, Debt D1. The world stands in for them with a stated rule, and no run claims to pay that debt. Nothing is hidden.
What a run can show, and what it cannot
The theorem here is Theorem VI.2 of AP02 The Operator. In One Awareness’s words:
the persisting configurations are confined to the ones that add more than they take
From One Awareness
A run tests whether the theorem’s conclusion follows from its premises, in a world that meets them. So it can fire a kill switch written about the theorem. It cannot, on its own, fire a claim about real people. A run that leaves out the theorem’s premises fires nothing. The work’s physics predictions are frozen, and then tested against nature. This world is frozen, and tested against its own model, and it says so.
Try it
Click an operator to set its type: C, U or D.
- C gives; cuts a link that costs more than it brings
- U gives; never cuts
- D gives nothing
- × ruined, for good
- ruin line
| Op. | Type | Stock | Links to | State |
|---|
Clocks
Each configuration’s clock, in spans. Marks at 1, 2 and 3 spans; the bound is 3 spans. Where the others are type C, the 5 rounds they take to read the record are not counted.
Numbers
| m | Pair, hours | Defector, hours |
|---|
Build it
- The specification, version 1.0, tagged world-spec-v1.0. Every rule, every default, and how a run is measured. Its fingerprint: 3ef50233b79f5e16d6f7069606744e5c46cd8323df2541345d3e38cac1ea6995.
- The test values. Fixed cases with their expected results, so any build can show it agrees before it reports anything.
- The reference engine, and a second engine written from the specification’s text alone, both at the tag world-ref-v1.0.
The frozen hypotheses
Before any run, each hypothesis is written down in full: the worlds it runs in, what is measured, the line that counts, the seeds, and how the result is judged. Then it is frozen. It is published with a fingerprint and a permanent identifier, and never edited afterwards. Only then does any cell of it run. Trial runs outside its cells are named in the frozen text.
Group A · Reproduction. These check that a build is right: the worked case of the two machines, its boundary, the worked charger of Paper D in AP01, and AP02’s own formulas. Arithmetic settles them, so passing them shows only that a build is right. It is never called support for any claim.
Group B · Generalisation. These run the work’s claims in worlds of many operators, with noise.
| Hypothesis | The work’s sentence | Status | |
|---|---|---|---|
| B1 | The defector does not last | “Cruelty does not last. Not as a hope. As geometry.” | Frozen, published and run. It shows the premises at work: see below. |
| B2 | What persists adds more than it takes | “the persisting configurations are confined to the ones that add more than they take (AP02, Theorem VI.2)” | Frozen and published on 15 November 2026: doi.org/10.5281/zenodo.24248813. |
| B3 | Steady beats panic | “steady, early, unexciting effort preserves a corridor better than the same effort spent in panic” | Being written. Exploratory: it carries no switch. |
The sentences are from One Awareness.
B1 · The defector does not last
Frozen on 13 November 2026, before any cell ran, and published the same day with its permanent identifier: doi.org/10.5281/zenodo.24207719. Its fingerprint, SHA-256: 039a1e2f33b31406bb5bdf154522f4bb61a87b32d20061c338b63ec31d848acd. It ran that night, from 23:08 on 13 November to 00:18 on 14 November, Cape Town time.
What it asks. In a world that meets the theorem’s premises, is a defector ruined within its time limit, in every group it passes through? The theorem is Theorem VI.2 of AP02. Its premises are what the careful giver does: it reads which way its own stock is going, it conditions each link on the other’s record, and it cuts.
The time limit. It is measured in spans. A span is how long the group the defector is in, the defector and those linked to it, could last on its stock at the rate it is losing, before it reaches the point of no return. A defector still standing after three spans has lasted. The few rounds a careful giver needs to read the record are not counted against it. They are added to the limit, and taken off what is measured. Each cell is judged on the average across its seeds. It holds where the average, with a margin for chance of two standard errors, is at or under three spans.
How it ran. 208 cells, in two kinds of network, a ring and random links, each of twenty operators. One defector or three. Three levels of what giving delivers. Two levels of what an operator has of its own. Four levels of noise. And a world open to newcomers. Half the cells are controls. Each cell ran 100 seeds, of 4,000 rounds each. A second engine ran 32 of the cells again, on its own seeds. In the closed controls nobody acts on a reading, so the 24 control cells with noise in the reading only give the same figures as their quiet twins, to the last digit.
What was known before it ran. B1’s own frozen text says it, in its first section. In this world a record is exact, so every careful giver cuts a defector within the record’s window, and a defector cut off falls, by arithmetic. So B1 holding is the premises at work. It is labelled so, and never as support for the claim. What the run adds is the measurement: two engines agreeing at scale, the margin under noise, and the controls beside it. A hypothesis whose outcome is open would need a world where the cut is not certain. That is a different hypothesis.
What it found.
- B1 held in every one of its 104 cells that could fire. No cell fired. No cell settled nothing.
- Every defector was ruined: 20,000 of 20,000.
- Every link from a defector to a careful giver was cut: 52,720 links. The longest any stayed live was five rounds, the record’s window.
- The highest any cell reached, with its margin for chance, was under half a span: 0.469. The line is at three. The largest measure in any single seed was 0.71 of a span. No single group came near three.
- The two engines agree in all 32 cells they both ran, by the rule set before the run: their averages lie within three standard errors of each other.
- The record, at work. Where the world was open to newcomers, a defector left alone kept asking for a new link. Every careful giver refused it, by the record: 1,202 times. And the newcomers refused, by the record, every offer that fell on the defector: 126. That is the premises at work too, and is labelled so.
- What was left out, as the frozen text says. 2,040 links that would have joined two defectors were left out after the drawing. 312 defectors started with no live link, drew on no one, and were reported but not watched. In 48 seeds that left nothing to watch, and those seeds were counted and left out of the average: never more than two in a cell.
The poorest ring. Where giving delivers least, on the ring, with nothing of their own, a careful giver needs both its links to live. One that cuts a defector is left with one link, and falls. Then its neighbour is left with one. So the whole ring fell, one by one: every careful giver, in every seed, 14,400 in the eight cells. The frozen text said so before the run. They were not living off a defector, so they fire nothing. It is said here because it happened.
Two rules a reader should know. The registry’s note does not yet state two cases. The world’s specification settles both, and the frozen text says so.
A group that loses at some moments and gains at others: the longer of its two spans is used. A longer span makes lasting harder to show, so this rule leans away from a fire. In B1’s own cells it could change no verdict.
A group still standing at the end: a group that lasts settles, and stops losing. Its span would then have no end, and lasting could never show. So where its defector still draws on the others, the span is theirs: how long they could last on what they lose to it. This rule was written before the freeze, after a trial run showed the gap. In B1’s own cells it applied nowhere. In the controls it is how a defector that lasts shows as lasting: it applied to 8,602 groups in random links, and to 795 in the open world.
The controls. In each control, the careful givers are replaced by givers who never stop. Controls fire nothing, by design. They show whether the measure can see a defector last, where the premises are taken away. A control is printed “a defector lasts” or “no defector lasts”, never “fires” or “holds”, with “control: fires nothing” beside it.
- On the ring, where each giver has only two links, no defector lasted, in all 48 controls. A giver who feeds a defector there cannot cover its own loss. It is ruined, and the defector after it.
- In random links, a defector lasted, in all 48 controls: on average between about 109 and 233 spans. A giver with enough other links can carry a defector without end. Not every one: 1,006 of the 9,600 defectors there were ruined all the same.
- In the world open to newcomers, a defector lasted, in all 8 controls: on average between about 67 and 93 spans. There, givers who never cut took its offers of new links: 625 of them.
15 November 2026. Read again under the world’s specification, version 1.1, ten of B1’s controls on the ring show a defector lasting, where B1’s record reads “no defector lasts”, and two settle nothing. Version 1.1 changes how one case is measured: a group that ends while its givers are still losing. Its life is now measured against how long those givers could have lasted on what they were losing to it, where that is shorter than how long the whole group could have lasted. B1’s record stands as run.
Two engines, one maker. Both engines were built by one of our AI desks, and reviewed by another. All three desks are Anthropic’s AI models, so both engines, and the review of them, come from one maker. An engine built by someone else, from the specification alone, is the stronger check. It is a week’s task, on the room’s front page.
What it means. In a world built to the theorem’s terms, where every record is open and every careful giver cuts, the one who only takes did not last, in any seed. It could not have. The frozen text said so before the run. The run is the measurement of it, never support for the claim.
Where nobody cuts, two things happened. Where the givers could not carry the one who takes, they were ruined first, and it fell after them. Where they could carry it, in a network rich enough, it lasted, and the measure showed that plainly.
And in the poorest ring, cutting did not save the givers. Each one left with a single link fell in turn, all of them.
That is what the run shows: the premises at work, measured in every cell by one engine and in 32 cells by a second, with every cell, seed and group in the record. It shows nothing about real people.
The record. Every cell, every seed and every group, and every event as it happened, each file with its fingerprint. doi.org/10.5281/zenodo.24223400
Ways it could be wrong
The second road, in One Awareness’s own words:
The first road needed a shared inside. The second needs one premise — an operator acts to stay in being — and two facts anyone can measure: that what you do moves me, and that what is done does not unhappen.
From One Awareness
Where the second road would die: show a coupled system under one-way drift in which a sustained defector keeps its corridor open indefinitely; or an operator that persists while not acting to stay in being; or that the premise it starts from smuggles in an ought the derivation does not earn (KS-BAR.9). Bounded, by dated note of 11 November 2026: “indefinitely” is three spans of the configuration the defector is in. A span is how long that configuration’s reserve lasts at its net loss, before it reaches the point of no return.
From One Awareness, “The second road”
In plain words: show a world of linked operators, where time runs one way, in which one who only takes keeps going past three spans of the group it is in. Or show something that lasts without acting to stay in being. Or show that the starting premise sneaks in an ought it never earned. Then the second road fails.
If a net-negative configuration persists past its collapse bound, or undefended parasitism or pure defection persists against a defended counterparty, the containment theorem (Theorem VI.2) fails.
From AP02 The Operator
In plain words: if a group that loses more than it gains keeps going past its time limit, or if one that takes more than it gives, or one that only takes, keeps going against a partner who defends itself, the theorem fails.
The registry gives both switches the same bound, three spans, so one run can never fire one and not the other.
The plain words on this page are a reader’s help. Where they and the quoted line differ, the quoted line stands.
What comes next
B2 and B3 are written, read and frozen the same way, one at a time. Neither runs before it is frozen.
Worlds where the cut is not certain. In B1’s world, every record is open to everyone, so among careful givers the cut always comes, and among givers who never stop it never comes. Nothing lies between. In life, a great deal lies between. Three questions follow, and each could come out either way. All three are exploratory: none carries a switch.
- The one who asks someone new each day. What if each giver knows only what was done to it, and nobody holds the whole record? How long does a defector last, against how many givers it can reach, and how long each takes to see?
- The giver who relents. What if a giver who has cut, and knows why, now and then gives again, out of pity? How often must pity come for a defector to last?
- The defector who can choose. What if a defector can read its own fall, and start to give? When it is cut off, does it fall, or does it stop defecting? And what does pity do to that choice?
None is written yet. Each needs a new choice of the world, made in the open, and an honest label for what it can fire, before it is frozen. The first two take away part of the theorem’s premises, so they would not test the theorem. They would map how much of each premise its conclusion needs. Each is open now as a week’s task, for anyone who wants to build it first.