published September 28, 2026

Under an energy economy, the dominant replicator copies itself faster

DESIGN §1.3 sweep 11 continued emerged worlds for another 20 000 epochs, each paired with a continuation under its own dynamics. A confirmatory reading, written after ten parents had been seen and tested only on the eight that qualified afterwards, asked rung 3's question: under an energy economy, does the dominant replicator get faster at copying itself? It does. Its copy_latency, the step at which a partner first holds a complete image, fell relative to the paired continuation in 20 of 24 pairs at economy 2048 (p = 0.00077) and 21 of 24 at economy 8192 (p = 0.00014). The median child's last-decile latency was 0.82 and 0.75 of its first-decile latency under the two economies, against 1.11 in the continuation, and the result holds in 7 and 8 of the 8 parents. It is the programme's first confirmed adaptation: one substrate, latency read on the dominant tape in isolation, and the mechanism is not identified.

Question

Rung 3 of the programme asks whether replicators, once they exist, adapt: whether selection makes them better at the one thing every replicator does. The copy-cost survey looked for it in every transitioned run and could not separate one lineage improving from a cheaper replicator winning. DESIGN §1.3 sweep 11 gives the question a control: it continues an emerged world under a treatment and, seed for seed, under its own dynamics. Under an energy economy, does the dominant replicator get faster at copying itself than the same world does left alone?

Method

Each parent is an economy-off control of sweep 9 whose last stored world is at least half replicators by the orientation-aware census. Its terminal world is continued for another 20 000 epochs under four treatments, three seeds each, paired by (parent, seed): the continuation, an economy with influx 2048, a rich economy with influx 8192 (both with steal 1024, a stock cap of 32 768 and loss 0.5), and host mode.

The reading here is a held-out confirmatory one (docs/design_record.md, 2026-09-25, "The from-emerged economy arms, confirmed on held-out parents", #263). Its rules were written after the sweep's own pre-registered reading (2026-09-25, "Runs that start from an emerged world", and its six clarifications) and an exploratory look at the children of the first ten parents had been seen, and it is read only on the parents that qualified afterwards, none of whose children had been seen. For each child that is not extinct, the latency ratio R is its last-decile median copy_latency over its first-decile median, past a settling window of 1 000 epochs. copy_latency is the step at which a noise partner first holds a complete image of the dominant tape, in either orientation. A pair favours the treatment where the economy child's R is lower than the continuation child's. H3-latency is a one-sided sign test per economy arm at p < 0.05.

Result

shown Over the held-out children, the median ratio of last-decile to first-decile copy_latency was 0.821 under economy 2048 and 0.749 under economy 8192, against 1.11 in the continuation.

• economy 2048: 20 of 24 pairs favour the treatment and 4 the continuation, one-sided sign test p = 0.000772, shown. More pairs favour the treatment than the continuation in 7 of 8 held-out parents. No one or two parents carry the result: leaving any of them out keeps p below 0.05.
• economy 8192: 21 of 24 pairs favour the treatment and 3 the continuation, one-sided sign test p = 0.000139, shown. More pairs favour the treatment than the continuation in 8 of 8 held-out parents. No one or two parents carry the result: leaving any of them out keeps p below 0.05.

The economy lowers copy_latency against the paired continuation, on 8 parents whose children nobody had seen when the rule was written. That is selection for faster copying, and it is the programme's first confirmed adaptation result.

Treatment	Measured	First-decile latency	Last-decile latency	Median ratio R	R below 1
continuation	24	1,449	3,073	1.11	6
economy 2048	24	1,409	1,023	0.821	21
economy 8192	24	1,429	1,070	0.749	16
host mode	24	1,437	2,065	1	9

Treatment	Parent	Favour the treatment	Favour the continuation	Tie	Not both measured
economy 2048	run 2577	3	0	0	0
economy 2048	run 2590	3	0	0	0
economy 2048	run 2654	3	0	0	0
economy 2048	run 2676	2	1	0	0
economy 2048	run 2700	3	0	0	0
economy 2048	run 2771	1	2	0	0
economy 2048	run 2802	2	1	0	0
economy 2048	run 2862	3	0	0	0
economy 8192	run 2577	3	0	0	0
economy 8192	run 2590	2	1	0	0
economy 8192	run 2654	3	0	0	0
economy 8192	run 2676	2	1	0	0
economy 8192	run 2700	2	1	0	0
economy 8192	run 2771	3	0	0	0
economy 8192	run 2802	3	0	0	0
economy 8192	run 2862	3	0	0	0

Every count on this page is read from the sweep's stored samples at render time, through the same reading the sweep page draws.

A held-out confirmatory reading, pre-registered 2026-09-25 final

Its rules were written after the reading above and an exploratory look at the same children, so they are read only on the children of parents from sweep 9's extension (seeds above 90) none of whose children had been seen. Past a settling window of 1,000 epochs, a child relapsed where its share sat below 10% for 3 samples running; it is extinct where its last-decile median share is below 10%. H3-latency pairs the economy child against the continuation child on the ratio of their last-decile to first-decile median copy_latency, a lower ratio favouring its side; H4-survivors reads the original complexity rule on pairs whose two children both survived. Host mode is not re-tested.

Treatment	Children	Finished	Settled relapses	Extinct	Latency measured	Survivors
continuation	24	24	0	0	24	24
economy 2048	24	24	3	0	24	21
economy 8192	24	24	5	0	24	19
host mode	24	24	0	0	24	24

Hypotheses

• H3-latency, economy 2048 against the continuation: shown

  24 pairs measured on both sides: 20 favour the treatment, 4 the continuation, 0 tie; one-sided sign test p = 0.000772.

• H4-survivors, economy 2048 against the continuation: refuted

  21 pairs measured on both sides: 0 favour the treatment, 1 the continuation, 20 tie; one-sided sign test p = 1.

• H3-latency, economy 8192 against the continuation: shown

  24 pairs measured on both sides: 21 favour the treatment, 3 the continuation, 0 tie; one-sided sign test p = 0.000139.

• H4-survivors, economy 8192 against the continuation: not shown

  19 pairs measured on both sides: 1 favour the treatment, 0 the continuation, 18 tie; one-sided sign test p = 0.5.

What is not claimed

• Not in situ. copy_latency is measured on the world's dominant tape against a fresh noise partner, not in the colony. A lower ratio says the dominant copier copies faster in isolation, not that the colony replicates faster.
• Conditioned on not going extinct. An extinct child has no ratio and leaves H3-latency, and a treatment that kills the colonies that would not have sped up leaves survivors that did. A settled relapse alone does not leave the test, only extinction does; the extinct and settled-relapse counts per treatment are tabled above.
• One substrate, one budget. Every child is sweep 9's world at caps 128 and 256, continued for 20 000 epochs, three correlated children per parent.
• The mechanism is not identified. Why copies get faster under an economy — a shorter copy loop, a different winner, a tape that spends less before it starts copying — is not read here. The ratio is also read only past the settling window, so a speed-up inside the first 1 000 epochs after the switch is not counted.
• No halting copier. copy_cost, the steps for one byte-exact copy that halts, stays at max_steps wherever it is measured: the copiers loop rather than finish. That is exploratory, not part of the pre-registered reading.

Evidence

Transition, flagged, census and arm are used here in the senses set out in the glossary.

Sweep: From an emerged world