negative imports an objective October 02, 2026

Paid to compute, replicators take the one-step tasks and stop there

DESIGN §1.3 sweep 15 paid replicators for computing. The 18 emerged worlds of the from-emerged sweep ran another 40 000 epochs each, three seeds a parent, with arithmetic tasks rewarded and, in a paired twin, assayed but unpaid. All 108 children finished and none died out. Pay bought capability: all 54 pairs favour the reward (p = 5.55e-17), and the paid worlds took ECHO, INC and DEC, the tasks one substitution of a copier computes. They went no further. The rung-4 question, whether paid computation climbs to the tasks that need a counted loop, is refuted: no child of either arm held a loop task at a tenth of its world, 0 to 0 with 54 ties. Paid complexity did rise, in 18 pairs against 1 (p = 3.81e-05), and both shown tests hold without the piloted parents. The substrate imports an objective, so none of this moves rung 4 on Soup, which stays not shown.

Read from Metabolism, a second substrate that pays a cell for computing: it imports an objective and is never pooled with the fitness-free runs. Rung 4 on Soup is unaffected by it. What that means.

Question

Complexity rises at emergence and then stops rising on every fitness-free substrate the programme has tested, and from an emerged start an energy economy did not change that. One reading of the plateau is that nothing in the soup is worth computing: a byte off the copy path costs its tape nothing and earns it nothing. The Metabolism substrate of DESIGN §1.4 changes that one thing. A cell is paid in energy for computing arithmetic on inputs the environment hands it. Once computing pays, do replicators climb past the tasks one mutation reaches, to the ones that need a counted loop?

Method

The parents are the emerged worlds the from-emerged sweep continued, chosen by its rule. Each terminal world runs another 40 000 epochs, seeds 2001–2003, in two arms. Both run an economy in which only the cell that starts an interaction pays (influx 1 024, a stock cap of 65 536, no theft), and both run an arithmetic assay every 8 epochs. In the reward arm a credited task pays 2 048 energy a unit. In the no reward arm it pays nothing, so that world runs byte for byte as if the assay were off. A tape is credited with a task when one of its outputs is right on all three input cases. ECHO, INC and DEC are one substitution away from a copier. ADD, SUB, NOT, DOUBLE and MUL are the loop tasks: in BFF each needs a counted loop.

The reading is pre-registered, with every number fixed before a child was seeded: docs/design_record.md, 2026-10-01, "Metabolism: a second, labelled substrate that imports an objective", DESIGN §1.3 item 15. Each reward child is paired with the no-reward twin of the same parent and seed. Both are read past a settling window of 1 000 epochs, and a pair with an extinct child is in no test. There are three one-sided sign tests at p < 0.05:

H-capability: the last-decile median task_capability, the number of tasks held by a tenth of the world.
H-ladder, the rung-4 question: the same count over the loop tasks, task_capability_loop.
H-complexity: the from-emerged rule, under which the dominant replicator's executed instruction count rises when its last-decile median is at least 1.2 times its first. It reads only pairs in which neither child relapsed.

Each test is also read without the four parents the design study piloted, and a shown test lists any one or two parents whose pairs it cannot do without.

Result

H-capability shown H-ladder refuted H-complexity shown Paid to compute, replicators took the one-step tasks and stopped there.

H-capability: 54 pairs favour the reward, 0 the twin and 0 tie, of 54 measured, p = 5.55e-17, shown. Every pair of all 18 parents favours the reward. Without the piloted parents: 42 to 0 with 0 ties, p = 2.27e-13, shown.
H-ladder: 0 pairs favour the reward, 0 the twin and 54 tie, of 54 measured, refuted. By parent, 0 lean to the reward, 18 tie throughout and 0 lean to the twin. Without the piloted parents: 0 to 0 with 42 ties, refuted.
H-complexity: 18 pairs favour the reward, 1 the twin and 35 tie, of 54 measured, p = 3.81e-05, shown. By parent, 10 lean to the reward, 8 tie throughout and 0 lean to the twin. Without the piloted parents: 11 to 1 with 30 ties, p = 0.00317, shown.

No one or two parents carry a shown test: whichever one or two are left out, the rest still read p < 0.05.

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

The ladder

Descriptively, counting a task where it held a tenth of a child's world at any sample: ECHO in 54 of 54 reward children, INC in 52 and DEC in 50, against 27, 1 and 0 of 54 twins. ECHO came first and INC and DEC close behind it; the table below gives the median epochs. 0 reward children and 0 twins reached a loop task, so 0 climbed one through INC or DEC, the stepping stones of Lenski et al. 2003. At the end of the run only one twin, run 4679, held a task at a tenth of the world. Tasks gave a reward cell about 50% of its income, between 35% and 55% across the children, an estimate that ignores the stock cap.

The pre-registered reading final imports an objective

Every child of every qualifying parent has finished. Each reward child is paired with the no-reward child of the same parent and seed, and both are read over their own samples past a settling window of 1,000 epochs. A child is extinct where its last-decile median share is below 10%, and a pair with an extinct child is in no test. Each test is a one-sided sign test at p < 0.05, re-read without the parents the design study piloted (runs 1007, 2577, 1029, and 2862). Whatever it reads, this substrate pays for computing, so rung 4 on Soup stays not shown.

Children per arm, read so far. A loop rung is ADD, SUB, NOT, DOUBLE or MUL held by a tenth of the world; a stepping stone is INC or DEC held first.

Arm	Children	Finished	Settled relapses	Extinct	Capability measured	Reached a loop rung	Through a stepping stone	Complexity survivors	Rises
reward	54	54	0	0	54	0	0	54	18
no reward	54	54	0	0	54	0	0	54	1

Hypotheses

H-capability, reward against no reward: shown

54 pairs measured on both sides: 54 favour the reward, 0 the twin, 0 tie; p = 5.55e-17. Without the piloted parents: shown, 42 to 0, p = 2.27e-13.

Per-parent agreement

Parent	Favour the reward	Favour the twin	Tie	Not both measured
run 944	3	0	0	0
run 950	3	0	0	0
run 967	3	0	0	0
run 991	3	0	0	0
run 1007	3	0	0	0
run 1029	3	0	0	0
run 1087	3	0	0	0
run 1089	3	0	0	0
run 1103	3	0	0	0
run 2568	3	0	0	0
run 2577	3	0	0	0
run 2590	3	0	0	0
run 2654	3	0	0	0
run 2676	3	0	0	0
run 2700	3	0	0	0
run 2771	3	0	0	0
run 2802	3	0	0	0
run 2862	3	0	0	0

H-ladder, reward against no reward: refuted

54 pairs measured on both sides: 0 favour the reward, 0 the twin, 54 tie. Without the piloted parents: refuted, 0 to 0.

Per-parent agreement

Parent	Favour the reward	Favour the twin	Tie	Not both measured
run 944	0	0	3	0
run 950	0	0	3	0
run 967	0	0	3	0
run 991	0	0	3	0
run 1007	0	0	3	0
run 1029	0	0	3	0
run 1087	0	0	3	0
run 1089	0	0	3	0
run 1103	0	0	3	0
run 2568	0	0	3	0
run 2577	0	0	3	0
run 2590	0	0	3	0
run 2654	0	0	3	0
run 2676	0	0	3	0
run 2700	0	0	3	0
run 2771	0	0	3	0
run 2802	0	0	3	0
run 2862	0	0	3	0

H-complexity, reward against no reward: shown

54 pairs measured on both sides: 18 favour the reward, 1 the twin, 35 tie; p = 3.81e-05. Without the piloted parents: shown, 11 to 1, p = 0.00317.

Per-parent agreement

Parent	Favour the reward	Favour the twin	Tie
run 944	2	1	0
run 950	2	0	1
run 967	0	0	3
run 991	1	0	2
run 1007	2	0	1
run 1029	2	0	1
run 1087	0	0	3
run 1089	2	0	1
run 1103	1	0	2
run 2568	0	0	3
run 2577	3	0	0
run 2590	1	0	2
run 2654	0	0	3
run 2676	0	0	3
run 2700	0	0	3
run 2771	2	0	1
run 2802	0	0	3
run 2862	0	0	3

The pairs

Parent	Seed	Reward child	Capability	Twin	Capability
run 944	2001	run 4573	3	run 4576	0
run 944	2002	run 4574	3	run 4577	0
run 944	2003	run 4575	3	run 4578	0
run 950	2001	run 4579	3	run 4582	0
run 950	2002	run 4580	3	run 4583	0
run 950	2003	run 4581	2	run 4584	0
run 967	2001	run 4585	2	run 4588	0
run 967	2002	run 4586	2	run 4589	0
run 967	2003	run 4587	3	run 4590	0
run 991	2001	run 4591	3	run 4594	0
run 991	2002	run 4592	3	run 4595	0
run 991	2003	run 4593	3	run 4596	0
run 1007	2001	run 4597	3	run 4600	0
run 1007	2002	run 4598	2	run 4601	0
run 1007	2003	run 4599	3	run 4602	0
run 1029	2001	run 4603	3	run 4606	0
run 1029	2002	run 4604	3	run 4607	0
run 1029	2003	run 4605	3	run 4608	0
run 1087	2001	run 4609	3	run 4612	0
run 1087	2002	run 4610	2	run 4613	0
run 1087	2003	run 4611	3	run 4614	0
run 1089	2001	run 4615	3	run 4618	0
run 1089	2002	run 4616	3	run 4619	0
run 1089	2003	run 4617	3	run 4620	0
run 1103	2001	run 4621	3	run 4624	0
run 1103	2002	run 4622	2	run 4625	0
run 1103	2003	run 4623	3	run 4626	0
run 2568	2001	run 4627	2	run 4630	0
run 2568	2002	run 4628	3	run 4631	0
run 2568	2003	run 4629	2	run 4632	0
run 2577	2001	run 4633	3	run 4636	0
run 2577	2002	run 4634	2	run 4637	0
run 2577	2003	run 4635	3	run 4638	0
run 2590	2001	run 4639	2	run 4642	0
run 2590	2002	run 4640	2	run 4643	0
run 2590	2003	run 4641	3	run 4644	0
run 2654	2001	run 4645	2	run 4648	0
run 2654	2002	run 4646	3	run 4649	0
run 2654	2003	run 4647	3	run 4650	0
run 2676	2001	run 4651	3	run 4654	0
run 2676	2002	run 4652	3	run 4655	0
run 2676	2003	run 4653	3	run 4656	0
run 2700	2001	run 4657	2	run 4660	0
run 2700	2002	run 4658	3	run 4661	0
run 2700	2003	run 4659	2	run 4662	0
run 2771	2001	run 4663	3	run 4666	0
run 2771	2002	run 4664	3	run 4667	0
run 2771	2003	run 4665	3	run 4668	0
run 2802	2001	run 4669	3	run 4672	0
run 2802	2002	run 4670	3	run 4673	0
run 2802	2003	run 4671	3	run 4674	0
run 2862	2001	run 4675	3	run 4678	0
run 2862	2002	run 4676	3	run 4679	1
run 2862	2003	run 4677	3	run 4680	0

The ladder, descriptive only: how many children's worlds held each task at a share of 10%, settling window included, and the median first epoch they did.

Task	Reward: children	Reward: median first epoch	No reward: children	No reward: median first epoch
echo	54	20,750	27	24,910
inc	52	22,060	1	44,120
dec	50	22,550	0	—
add	0	—	0	—
sub	0	—	0	—
not	0	—	0	—
double	0	—	0	—
mul	0	—	0	—

What a refuted ladder means

On the lab, a read-only count over all 432 000 stored samples of the 108 children found task_capability_loop at 0 in every one. In only one sample did the assay credit any loop task at all: DOUBLE, to 1 of the 256 cells it drew. With no loop task in either arm the test cannot favour the reward, and the pre-registration reads a tie throughout as refuted, as strongly as its power allows. It needed only 5 discordant pairs, and 0 of 54 reward children puts a two-sided 95% upper bound of about 0.066 on the rate at which a paid child reaches a loop task in this budget, against the pilot's 0.34.

So when computation pays, BFF copiers take the tasks one substitution away and stop. The registered reading of that outcome was that the plateau holds for every feature that needs several coordinated mutations, and that the suspects are BFF's instruction set and its substitution-only mutation. The design study's landscape measurement, recorded in the design record on 2026-10-01 ("The logic assay, the NAND op ~ and task_floor") and restated in the Logic pre-registration, points at the instruction set. Its numbers are pilot numbers from a scratch copy of the engine, not findings. BFF has no two-input data op, so every two-input task needs a counted loop, and the only loop a copier owns is its copy loop. No loop solver that still copies lay within 3 substitutions of any evolved solver it measured; the nearest it built is 5–7 coordinated substitutions away, and no partial kernel on the way computes anything the assay pays for. Indels did not shorten the path. If that holds, pay cannot pull a population across the gap, because there is nothing to select along it.

H-complexity being shown is consistent with that. Mirroring a solver and stacking the one-step tasks adds executed instructions without a loop, so the dominant replicator grows more elaborate in 18 pairs against 1. That reads "paid complexity rises", as the pre-registration words it, never "open-ended".

What comes next

Logic, pre-registered as DESIGN §1.3 item 16, keeps this substrate's economy, assay and reward and imports a primitive as well as an objective: one assay-only NAND byte, with Avida's logic tasks as the ladder, so every rung is a composition of NANDs and a lower rung is a part of a higher one. It runs on the same parents and asks the question this sweep could not reach: whether a soup assembles the deep rungs, XOR and EQU, from paid parts.

What is not claimed

Nothing about rung 4 on Soup. This substrate imports an objective: whatever rose here was paid for by tasks the programme chose. It is never pooled with the fitness-free runs, and rung 4 on Soup stays "not shown".
Not open-ended. H-complexity says the dominant replicator executes more of its tape under pay. It says nothing about unbounded growth, and the ladder it would have to climb stopped at its first rung of loops.
A small, arithmetic world. The inputs are 0 to 15, three cases an assay, and the ladder is arithmetic because BFF has no bitwise ops. Another task set, a larger input domain or a longer budget is untested.
Children are correlated. Three children share each parent's world, four of the parents were piloted, and it is one substrate and one budget. That is why every test is read by parent and without the piloted parents.

Evidence

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

Sweep: Metabolism

Transition epoch vs treatment

No finished runs yet.

Transitions and replicator counts

One row per finished run the detector flagged or that counted at least one replicator. "Transition epoch" is the detector's event; "Peak replicators" is the largest census the run reached, with the epoch it first reached it in brackets; "—" means the run never had a value.

No finished run of this sweep has transitioned or counted a replicator yet.

Runs

Every run of the sweep, flagged or not, with its seed — the denominator of the rate above.

Seed	Treatment	Status	Priority	Epochs done	Transition epoch	Census
2003	arith×8×2048×initiator×0×1024×65536	finished	40	60,000	— confirmed	26 census only	Run
2001	arith×8×0×initiator×1024×65536	finished	40	60,000	— confirmed	26 census only	Run
2002	arith×8×0×initiator×1024×65536	finished	40	60,000	— confirmed	36 census only	Run
2003	arith×8×0×initiator×1024×65536	finished	40	60,000	— confirmed	71 census only	Run
2001	arith×8×2048×initiator×0×1024×65536	finished	40	60,000	— confirmed	6 census only	Run
2002	arith×8×2048×initiator×0×1024×65536	finished	40	60,000	— confirmed	4 census only	Run
2003	arith×8×2048×initiator×0×1024×65536	finished	40	60,000	— confirmed	6 census only	Run
2001	arith×8×0×initiator×1024×65536	finished	40	60,000	— confirmed	84 census only	Run
2002	arith×8×0×initiator×1024×65536	finished	40	60,000	— confirmed	22 census only	Run
2003	arith×8×0×initiator×1024×65536	finished	40	60,000	— confirmed	23 census only	Run
2001	arith×8×2048×initiator×0×1024×65536	finished	40	60,000	— confirmed	34 census only	Run
2002	arith×8×2048×initiator×0×1024×65536	finished	40	60,000	— confirmed	13 census only	Run
2003	arith×8×2048×initiator×0×1024×65536	finished	40	60,000	— confirmed	15 census only	Run
2001	arith×8×0×initiator×1024×65536	finished	40	60,000	— confirmed	28 census only	Run
2002	arith×8×0×initiator×1024×65536	finished	40	60,000	— confirmed	9 census only	Run
2003	arith×8×0×initiator×1024×65536	finished	40	60,000	— confirmed	14 census only	Run
2001	arith×8×2048×initiator×0×1024×65536	finished	40	60,000	— confirmed	7 census only	Run
2002	arith×8×2048×initiator×0×1024×65536	finished	40	60,000	— confirmed	26 census only	Run
2003	arith×8×2048×initiator×0×1024×65536	finished	40	60,000	— confirmed	23 census only	Run
2001	arith×8×0×initiator×1024×65536	finished	40	60,000	— confirmed	26 census only	Run