Chasing the signs of life.

Milestone 3 (M3) asked for founders that can mend themselves. Its screen found genetically distinct seeds that regrow after a wound removes about 30% of a body, and die in a no-light control. A stricter retest narrowed the founder set to 12 genomes, which seeded the long-run ensemble.

Passing this gate does not establish open-ended evolution.

All 170 registered runs of 10⁶ steps finished with exact conservation. With mutation, new adaptive activity exceeded both controls in one of the two worlds. The second test failed in both: neutral runs, where mutations do not change behaviour, also counted as “growing”. A 10⁷-step extension showed the same pattern.

Living things kept recycling the waste, and different ways of living kept coexisting. A further hypothesis, held back to be checked separately, was not supported by any of its 4 measures.

Genomes are copied faithfully: grown separately, the genomes of tracked parents and offspring behave alike (slopes of 0.85–0.96, where 1 would mean offspring traits rise one-for-one with their parents’, on average), though most pairs carried identical genomes. Mutation rarely changes what a genome does in the ecosystem: the mutants of only 1 of the 12 founders reliably took up a new role. A broader sample of genomes did so far more often (11 of 24), so the way founders were chosen was a bottleneck.

Evolved lineages also lost the founders’ ability to regenerate (0.17 against 0.98).

The tracker credits each birth to a nearby parent. A direct check found that in 135 of 200 sampled births the credited parent contributed no copied material at the time of the birth. A tracked birth is therefore not, by itself, evidence that an individual reproduced.

This is a limit of the measurement, not a negative result. It qualifies every result built on tracked births. New exploratory worlds can still be tried while better measurements are developed.

None of the four activity, novelty and role measures met every validation requirement: three flagged too many neutral runs, and the role measure barely separated the comparison it was meant to.

This limits what those scores can tell us about evolution. A visually interesting world remains a lead to investigate.

What the founders were mattered more than how varied they were. Clouds of drifted founder genomes did little better than single founders (2 of 12 against 1 of 12), while a set made only of non-producing genomes met the new-role criterion in 4 of 5 runs, though part of that may be sorting of roles already present. Whether new roles keep arriving after niches first fill is mixed: they do not in the main comparison, and an excess in the 10⁷-step extension becomes unclear once ambiguous returns are set aside.

A bug that zeroed a background genome’s weights in one screen was found and fixed; the affected result is corrected in the plan.

Following experiments that impose a group life cycle on microbes (Black, Bourrat & Rainey, 2020), each tile became a pond. At every cycle, ponds are cleared and reseeded with a small fragment from a successful pond. In new histories, with each test fixed before its data, heritable differences between ponds (genetic or structural) appeared in 6 of 6. Fragments from cycled ponds also passed a regrowth test more often than those from random-donor ponds, uncycled ponds and the ancestor, both at once and after 200,000 steps without the cycle, again in 6 of 6.

Exploratory, not a registered milestone. Random-donor ponds show heritable differences too, so the cycle’s causal role is not shown. The genome-swap test passed only at its threshold, and a later matched control showed that most of that apparent genome gain came from mutants in the control. The life cycle is imposed rather than discovered.

The pond life cycle then went through a registered test: hypotheses and decision rules frozen on 2 October, 24 histories per arm. Both primary hypotheses were confirmed. Worlds evolved under the cycle with selected donors founded new ponds from fragments more often than random-donor worlds, uncycled worlds and the ancestor, both when the cycle stopped and 200,000 steps later (median success 0.97 against 0.61 and 0.66 for the last two at the later point). Separately, the genome taken when the cycle stopped, placed on ancestral material, founded more ponds than the ancestor’s genome in 24 of 24 histories. That test does not explain the later persistence.

The genome’s effect is small: a median of 19 more successes in 512 fragments, and its gain did not meet the registered threshold of half the worlds’ advantage (5 of 24), a comparison that does not measure the genome’s causal share. The whole-world comparisons used fragments with different starting masses. Heritable differences between ponds were shown in 17 of 24 histories, and in as many random-donor histories, so they are not shown to be specific to the cycle. Many random-donor worlds collapsed, so the comparisons that carry the claim are those against uncycled worlds and the ancestor. The life cycle is still imposed; ponds reproducing on their own is not tested. By a decision made before the run, this counts toward no milestone.

The next question removes our hand from the ranking. If a pond’s number of offspring follows its own export rather than our choice, do ponds evolve to export better? A frozen, exploratory hunt passed its entry checks: worlds stayed viable under the new regime, and the export measure separated a known large difference in 6 of 6 cases.

A small diagnostic with mutation switched off (four pairs, 30 cycles) showed no consistent advantage in the export measure; it cannot rule out non-genetic effects. The evolutionary stage is prepared and has not run. A hit there would be a candidate that still needs a registered replication.

A new workbench asks which worlds let evolution accumulate, using exact cases that replay identically on a second machine. No scientific case has run on it yet. It starts from the audited records of a separate renewal experiment, which looked for habitats that keep a source of life going and also maintain a site outside it by local synthesis rather than imported biomass. None of the 24 main pilot cases qualified, so the confirmation stage was not run. All 40 histories conserved matter exactly, and the designated CPU and GPU replays matched.

A negative for this panel only, with one pilot seed per habitat. In calibration the detector raised no false alarm on 12 of 12 null controls and caught 3 of 4 known positives, so a steady renewal could be missed. Nothing here is evidence of evolvability, for or against.

Sandbox looks, one seed each. The full storm world was extinct by step 29,000. In a separate ladder of sun speeds on the calmer gradient world, the fastest sun was lethal to the regenerating founders and each slower one left more living cells; hand-built founders survived the fastest. Sixty cycles of breeding ponds for movement alone left 23 of 64 ponds occupied; adding the size of the founding packet to the score kept 63 occupied, moving at about half the speed. When a vision model picked the most unusual-looking ponds, 9 ponds were occupied after 16 cycles against 14 under random picking, and the median movement score was zero from cycle 8 on. One seed and one model draw, so no general difference between pickers is shown; novelty and movement are not the same thing.

Separately, eight screens of rules meant to make evolution keep bodies (lossy takeover, injury, inherited shape rules, declared cells) produced no arm that met its pre-stated bar for promising; each reads unclear or dead end. Feeding a world by hand is now a lab tool, and it is logged as the one deliberate break in a closed world.

A game-kernel line gave genomes one gainable gene: decomposition as a paid-for capability, off in every founder, flipped on or off by mutation. Under a criterion written before the first run, carriers reached a quarter of the living cells by step 75,000 in all three seeds, and held about 99% at every census from step 155,000 through step 200,000. In a null arm with the same cost and the same flips but no benefit, carriers hovered at 2 to 14% and never approached the bar.

The gene paid for the lineage, not for the world: the carriers’ worlds ended as a decomposing film — more living cells gathered into fewer individuals, no growth flux at the end, and less free energy. Three seeds and 200,000 steps, not a registered test, and the sweeping clade carried other mutations too, so the null argues the gene drove the sweep but does not prove it. The gene exists only in worlds that ask for it; every standard world behaves exactly as before.

The gene set then grew to two: polymer building as a second paid-for capability over the same founders, against a null arm whose build gene costs the same and does nothing. By the rule written before the first run, the result is confounded: the do-nothing gene reached about half the living cells in two of three null seeds, so the one treatment seed where the build gene’s share ended high says nothing about polymer. Carriers who could build barely did — build flux peaked at 16 quanta a step in one seed and read zero over the final stretch in the other two, and polymer never passed 2% of bound mass. Decomposition, the first gene, swept again in all six runs.

Three seeds and 200,000 steps, not a registered test. No run here tests a polymer gene that is expressed, so nothing follows about polymer itself. The proposed explanations — founders whose controllers barely drive building, a near-neutral bit riding clade turnover in a decomposer film — are readings, not measurements. The null design reads only on a quiet background: the first experiment’s world held its do-nothing gene to a few percent, while on this background with clade turnover a costly, unused gene reached the bar twice, so a zero-cost arm is the named next control. One treatment film collapsed after step 150,000, losing nine tenths of its cells; the null on the same seed did not, and no cause is assigned. Both genes exist only in worlds that ask for them; every standard world behaves exactly as before.

The first piece of an AI naturalist is a bell written in code, not a model. It scores how far a run has strayed from the neutral runs of the same world, where mutations do not change behaviour, and its rules were written down before it read any data. Over 85 finished runs, at the setting fixed beforehand, 30 of the 70 neutral runs rang where the bar allowed at most 2. By its own rule, then, it rings on drift. Every treatment and no-mutation run rang too, all but one of them in the first stretch the bell counts, from step 10,000. Those worlds are already smaller than neutral ones at the first census the bell counts, so it tells the kinds of run apart; it does not show it hearing something arise inside one.

A second reading, fixed beforehand as descriptive only, took each world’s threshold from 20 of its neutral runs and tried it on 30 other neutral runs in all: one rang, against about 1.4 expected if the two sets of runs are alike, and all 10 treatment and all 5 no-mutation runs rang. The false alarms at the first setting came almost entirely from two signals with long tails under drift, how many generations deep one line gets (20 runs) and how much of the jar one lineage holds (16), with one run on splits; a run can ring on more than one. No model was called and no new simulation ran. An exploratory instrument, not a registered test: this is a result about the instrument, not about the worlds. A field notebook built on the bell rings only above those thresholds, and its dated example is the one held-out neutral run that rang: a false alarm.

The bell’s second version drops the two signals behind nearly all of the first version’s false alarms: how many generations deep one line gets, and how much of the jar one lineage holds. Its scoring rules were written down before it ran, and it was judged on 70 new neutral runs of the gradient world that it had never seen. None of them rang, where the bar allowed at most 2; the first version’s rule rang on 19 of the same 70. All 10 treatment runs still rang.

As before, it tells the kinds of run apart; this test says nothing about hearing something arise inside a run. One world and 70 runs of 300,000 steps: the bar can tell a bell that false-alarms near 2% of the time from one near 10%, but cannot certify 5%. Checks that each run’s files were whole were tightened while the runs were going, before any output was read. Exploratory, not a registered test. All 120 runs of this version ran locally, and no model was called.

The next test gave the bell twins. Ten treatment histories were each run again five times, identical to the original through step 200,000, then given one deliberate act there: a tiny wound of five sites (twice, in two different places), a wound of radius 64, a wound of radius 96 that removed over a third of the living cells, or a feed. The threshold came from the first set of tiny wounds (5.70). By the rule fixed before the runs, the bell had to ring on at least 8 of 10 twins for the radius-96 wound and for the feed. It rang on 4 of 10 for each, and on 0 of 10 for the radius-64 wound. The held-out tiny wounds rang on 0 of 10, so the threshold itself held up.

Scored afterwards on the eleven signals of the quieter bell, with the threshold taken again from the tiny wounds (2.58), the radius-96 wound would ring on 10 of 10, the feed on 8 of 10, the radius-64 wound on 5 of 10 and the held-out tiny wound on 1 of 10. That is a hypothesis for a next version to fix in advance and test on new twins, not a pass. One reason a sudden wound scores low: the jar regrows part of what it lost within the 1,000 steps a departure has to hold. All the twins come from the same ten histories of one world. The third part, a model reading the same twins, has not run; its calls are paid and wait for a yes.

Three measures read change from who descends from whom, rather than from lineage labels that every mutation renews: how recently most of the jar shares one ancestor, how often a mutant family sweeps to half the jar, and how much of the jar passes from one family to another. Over 130 finished runs (30 treatment, 70 neutral, 30 with mutation off), none of the three passed every check, written down before computing, that a measure must pass before it can be used to look for selection. In the gradient world every measure pointed the wrong way, ranking neutral runs as more “selective” than treatment runs. Each treatment run was compared with each neutral run, counting 1 when the treatment run looked more selective, a half for a tie and 0 otherwise; the average had to reach 0.8, and the three scored 0.34, 0.18 and 0.00. The two measures that flag single runs flagged 16 and 40 of the 70 neutral runs, where the bar allowed 2.

In the spots world the expected direction held, but against neutral runs showing almost no change by these tree measures, so that pass says little. Neutral and treatment worlds also differ, by world, in size, in how many lineages they hold and in how many mutations they mint, and those differences alone separate them; a neutral comparison matched on size and lineages is needed before a tree measure can mean selection. Nothing was dropped to rescue a measure. Exploratory, computed from runs already on disk with no new simulation: a result about the measures, not about the worlds.

The shared garden plans to let visitors drop stones: wounds of a chosen size. This test, its rules written down before it ran, asked how long one stays findable in the big storm, founded as the garden’s commons would be. Twin worlds were forked from three early checkpoints of each of three seeds, identical but for one stone. At each census, every 1,024 steps, the twins are compared by the share of cells, living in either, whose lineage label differs. A stone’s horizon is the first census from which that share in the patch around it (a disc twice its radius) is no larger than in the rest of the jar, four censuses running. The prediction, written on 9 October, was that a radius-8 stone fades within about 13,000 steps, the time in which the storm itself wounds each cell about once. By the rule it fails: the median is 97,279 steps over 9 pairs, and 4 of them had not met the rule when the window closed, one sun wander after the stone. Stones of radius 32 and 64 read by the rule as marks with no lasting effect, with medians of 26,623 and 35,839 steps. Both are under one wander, so the garden’s provisional day of one wander stays.

The rest of the jar drifts from its twin fast: it is at least nine tenths different within 5 to 20 censuses in 23 of 27 pairs, and after census 16 both regions read 0.99 or more in 759 of 1,089 censuses. The measure saturates, and a small patch reaches its ceiling more easily, which is why radius 8 runs longer than 32 and 64. So these are the rule’s outcomes, and what they say about how long a stone stays findable is unclear. Three seeds, one of them a carpet (97 to 99% of the jar living); one of the first three seeds had died out before the first checkpoint and was replaced. Nine pairs per radius. “The sun is the first letter” is the rule’s branch, not something the measure showed. The rules were not committed before the run, so git alone cannot show that they came first. A catalogue test, not registered; it ran locally in 3 h 48 min and changed no garden setting.

Two controls followed, forked from the same nine checkpoints (three from each of the three seeds), their rules fixed before they ran. A null stone was the same wound dropped on a disc chosen so that it would change no cell. All 7 such twins stayed identical to the undisturbed world at all 16 censuses checked: radius 8 at 5 of the 9 checkpoints, and radius 32 at 2 of the three seeds’ first checkpoints, the only ones with such a disc that large; none had one of radius 64. A one-cell nudge, which removes no biomass, stood in as the smallest change. Read in patches the size of each stone’s own, its median horizon was 1,023 steps against the radius-8 stone and 9,215 against radius 64, under half the stones’, so by the rule both radii read as stone effects. At radius 32 it was 30,719 against the stone’s 26,623, so by the rule that radius reads as a divergence clock: the horizon times how any difference spreads, not the stone. None of the three readings is firm: checkpoint by checkpoint, the nudge lasted at least half as long as the stone at 4, 5 and 5 of the 9 at radius 8, 32 and 64. The split is by day: at the first checkpoint the nudge was under half the stone’s in all three seeds at every radius, and at the two later ones it lasted at least half as long at 5 of 6 checkpoints at radius 32 and 64, and 4 of 6 at radius 8. Censused every 64 steps through the first 8,192 steps after the first checkpoint, each stone’s patch ran ahead of the rest of the jar, summed over that window, by more than twice the nudge’s at radius 8 and 32 in all three seeds, and at radius 64 in 2 of 3.

The nudge clears one cell that holds no biomass but that the physics still counts as living, and its patch is centred there, not where the stone fell, so “stone effect” could be the size of the change alone. A stone removes the bodies in its own patch, so a larger early difference there is close to what a stone is. The null stone checks the plumbing of a scheduled wound, not “the act of intervening”: a wound that writes nothing intervenes in nothing. In one seed the nudged twin’s lineage labels first differed from the undisturbed world’s at the 26th of the 128 fine readings, at all three patch sizes, though the two worlds’ states already differed at the first census. The fine census covers the first checkpoint only; at the two later ones, a stone is not shown here to be more findable than any other change. The rules were not committed before the run, so that they came first rests on hashes of the note taken at the time, not on a commit, and the nudge cells and null discs were chosen, from the checkpoints alone, before the rules were fixed. The same three seeds and nine checkpoints as the first test; a catalogue test, not registered. It ran locally in 1 h 35 min; the garden’s day stays one wander and no setting changes.