partial September 13, 2026

Does copying get cheaper?

Every sample now records the copy cost of the dominant replicator — the interpreter steps the most populous tape that passes the replicator test needs for one byte-exact copy — so a world that got better at copying itself can be seen getting better rather than inferred. This reads that series off every transitioned run the lab has, whichever sweep it came from, and counts how many end cheaper than they started. It is the first direct test of adaptation in the programme and it is not yet the adaptation sweep: the runs were assembled by having transitioned, not designed to answer this, and a falling cost can be a cheaper replicator winning rather than one lineage improving.

Question

The programme can say when a world crosses into self-replication and whether it stays there, but nothing in it has yet asked whether what emerged gets any better. Selection, if it is running at all, has something to push on: a replicator that finishes its copy in fewer interpreter steps spends less of an interaction's budget on it, and in a fixed budget that is the difference between copying and running out of time. So the question is the plainest version of adaptation this substrate allows — once a world holds replicators, does the cost of a copy fall as the epochs go by?

Method

The engine's replicator test already runs four trials of T ++ R per tested tape and counts the ones that leave T in the second half. It now also keeps what each passing trial cost in interpreter steps and reports the median of them: copy cost, the vocabulary the design record's evolution programme fixes, measured on the runs the test was already performing rather than on new ones. The sample records the cost of the dominant replicator — the most populous tape among the top_k tested that passes the test — and null when no tested tape replicates. The hand-written replicator in the engine's own test suite copies itself in 1 794 steps, which is what pins the reading.

This page then reads that series back out of the lab: every transitioned run, whichever sweep queued it, its samples at or after its own crossing epoch, in epoch order. A run is counted as falling when its last recorded cost is below its first post-transition one, rising when it is above, flat when they are equal. Counts of directions, nothing fitted.

Result

The direction is not one way: as many worlds get dearer as get cheaper. The lab has 472 terminal runs that crossed, with a copy cost recorded after the crossing in 214 of them, across 8 sweeps. Among the 209 runs with at least two readings, the split is 4 ending below their first post-transition cost, 4 above it and 201 on it.

One row per transitioned run whose samples carry a copy cost at or after its crossing, grouped by sweep. Costs are interpreter steps for one byte-exact copy by the run's dominant replicator at that sample; "Cheapest" is the smallest the run ever recorded. The table stops at 50 rows and leaves 164 further runs out; every one of them is counted in the figures above.

What this does not claim

This is not the adaptation sweep. The design record's evolution programme asks for a grid designed to measure adaptation: parameter points that transition, long budgets, enough seeds for the direction to mean something. What is above is every run that happened to transition, from sweeps built to answer other questions at other budgets, none of them run to answer this one. A count of directions over such a set describes the set.

"Dominant replicator" is a tape, not a lineage. The cost is read off the most populous tape that passes the replicator test at that sample. When it falls between two samples, that can be one colony getting cheaper — or a cheaper replicator taking over from a dearer one, which is selection too but a different sentence, and this page cannot tell the two apart. It also says nothing about the tapes below the top of the census.

Every series is censored at both ends. No run recorded a copy cost before this observable existed, so each series starts wherever the lab first ran the new engine rather than at the world's crossing, and ends wherever that run's budget ended. A cost that was falling when the clock ran out was falling, not converging.

Steps are not fitness. Copy cost is the price of one copy in a test the engine performs against a random partner, not the rate at which a tape actually spreads through the world. A cheaper copier that meets fewer partners spreads more slowly than a dearer one that meets more; copy_rate and the census remain the observables of what happened in situ.

What next

The series has to accumulate before it can be read: the runs on the clock need to sample under the new engine for long enough to give more than a couple of readings each. After that, the honest form of this question is the adaptation sweep — the transitioning parameter points at many seeds and a long budget — and the version of it that would really settle the matter reads copy cost per lineage rather than per dominant tape, so a single colony can be watched getting cheaper.

Evidence

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

This finding rests on no single sweep: its evidence is every transitioned run in the lab, tabled above with the sweep and the seed each one came from.