Research evidence dataset. No model weights. Part of the collection Xyntetik Research: GENESIS. The complete record of the genesis-existence line: preregistrations with every amendment, reports, the experiment code, and the raw run logs of every counted run from GE-001 to GE-017 and the H3 studies, failures included, with the unflattering results stated first. GE-013 to GE-015 come from the genesis-genome branch.

Xyntetik GENESIS: Existence record

Dataset summary

Question tested. Can a program be made that behaves less like a creation, judged by what it does for us, and more like an entity, judged by whether it keeps existing under the selection its own environment applies? The line asks this in two forms. GE-001 asks whether the medium a program is written in limits how long it keeps producing new behaviour under selection, comparing four media under one treatment. GE-002 onward places a small learning process inside a resource-limited container on a real host, where a reaper removes occupants that stay idle or overspend. They ask whether anything it can do with no reward for survival, a reward for survival, or a floor of exploration noise keeps it alive longer than blind action does. Later experiments ask whether a policy whose table keeps changing after its most-used consuming action is disabled persists longer than the same policy frozen at that moment (GE-006), whether a wider view of a real cue about host load lengthens survival (GE-005), and whether a structure that outlives its builder can accumulate across lives (GE-007).

Wording note, 2026-10-04. Under the owner's standing rule that GENESIS is theory-open, the framing above (an entity rather than a creation, and the autopoiesis literature in the prior art) is the line's motivation, not a criterion. Every pass, kill and inconclusive rule is a measured quantity: lifetime, survival after a knockout, plateau generations, p. The results below are stated in neutral terms, matching the branch's restatement of the same day, which the next dated update adds to record/. Arm names such as homeostat, ashby and entity are identifiers, not claims. No number or verdict changed.

Method. Every experiment was preregistered before it ran: thresholds, primary test, pass, kill, inconclusive and void classes, and since GE-003 the best p-value the design can attain, with designs rejected if that cannot clear their own threshold. Amendments are appended, never edited in place, and a pilot precedes any confirmatory stage that has no prior data, with pilot runs excluded from the test. The worlds run on one Windows 11 host (Intel Core i7-7700K, 4 cores, CPU only) under Docker with cgroup v2 limits, one container at a time, because host load contaminates the pressure signals the subjects read. The kill path is outside the world and was written down before any subject existed (see the hazard section below).

What each file contains.

record/README.md: the line's framing and design constraints, including the hazard section.
record/PRIOR_ART.md: the prior-art survey, with the items that changed what may be claimed.
record/RESEARCH_STATE.md: the running state, every verdict, correction and audit, in date order.
record/prereg/: one preregistration per experiment, each with its dated amendments.
record/report/: one report per closed experiment, unflattering numbers first.
record/ge001_*.py, record/s1_dsl.py: GE-001's four-medium screen and its cardinality control.
record/ge002/: the container world (world.py, run_world.py, Dockerfile), the host supervisor, and every analysis, calibration and audit script used for a verdict.
record/queue/: the sequential run queue and its log, which record when each counted run ran.
runs/<experiment>.tar.gz: every counted run, a JSON summary and a JSON-lines event log per run.
PUBLIC-COPY-CHANGES.md: every difference between this copy and the private branch.
SHA256SUMS: hashes of every file here. The archives are built deterministically.

Reproduction. Unpack an archive with tar xzf runs/GE-004.tar.gz, which restores runs/GE-004/. Analyses run from record/ge002/, for example python analyze_ge004.py or python analyze_pilots.py ge006, once the runs/ directory sits beside record/. Rerunning the worlds needs Docker with cgroup v2 and the image built from record/ge002/Dockerfile. The host paths in supervisor.py are placeholders. Results depend on host load by design, so a rerun reproduces distributions, not individual lifetimes. GE-001 is deterministic and reruns bit for bit.

Result. The unflattering results come first.

GE-002 failed its preregistered pass on both limbs. 25 runs, 5 arms. A hand-written rule given the reaper's thresholds (homeostat) lived longest, about four times as long as the tabular learner (entity), so the world admits policies that persist far longer than the learner found.
GE-003 was inconclusive, p = 0.135 against 0.05. Its failure was computable in advance: a paired sign test on five seeds cannot reach 0.05 at four wins. That forced the standing rule that every later prereg states its attainable p-value.
GE-005 was inconclusive, p = 0.405, with the perceptive arm ahead on median 782.2 s against 731.5 s. The cue was real: announcements predicted load spikes 0.406 of the time against 0.253 for random timestamps. But the subject's listening rate did not change, and the pilot's power estimate, 0.975 at 11 per arm from 3 and 4 pilot values, was wrong: the confirmatory data give 0.09. GE-004's power basis was also small (5 values per arm), so its claimed power is unverified, though its PASS does not depend on it.
The GE-006 knockout pilot went the opposite way to its prediction. After the knockout, the policy whose table was frozen outlived the one whose table kept changing, and power in the predicted direction was 0, so the confirmatory stage was not run. The first explanation recorded for it named the wrong mechanism and was corrected in the prereg. The redesign, GE-006b, is pending.
The GE-007 shelter was too easy. It persisted exactly across lives, but other occupants built about 95% of it, and a life that inherited nothing raised it as fast as one that inherited it. The redesign, GE-007b, is pending.
Four instrument faults were found in this world after runs had been counted, and none flips a verdict:
- The throttle component read a cumulative counter instead of a per-second delta. A preregistered calibration that would have caught it was never run. Its direction can only have handicapped the learners.
- A full /work disk silently stops the a5 action from counting as activity. It never happened before 1023 s, so it penalises the longer-lived arm. It caused two homeostat deaths in GE-002 and moves the GE-004 p-value from 0.00185 to 0.00205 when exposed runs are censored. The genesis-crossing line found it while porting this world, and showed it caps the homeostat's survival ceiling at the 1800 s horizon.
- The memory-pressure channel was exactly 0 in all 91,293 margin records, so one reaper rule could never fire.
- GE-001's first run was defective and is kept beside the amended rerun.
Two diagnoses made in the reports were withdrawn after the audit, and both reports carry corrections.
Physis (Egri-Nagy and Nehaniv, ECAL 2003) occupies this line's central corner, a medium under selection, by putting the processor description into the genome in 2003. It does not retire GE-001's four-medium comparison.

What passed:

GE-001: PASS for the continuous substrate. Median plateau bc 1, dsl 5, tpl 538 generations; nca never plateaued in 4000.
GE-001-CC: PASS as predicted. A finite 1200-neighbour lattice still never plateaued, which excludes the uncountable-neighbourhood explanation.
GE-002 amendment 4: the same learner with exploration held at 0.30 (entity_floor) passed both limbs. A later channel audit found its reward set by a near-constant process count 83% of the time, so this shows that holding exploration high outlived decaying it, not that anything was learned.
GE-004: PASS at its powered sample. entity_floor outlived uniform random actions (random): median 821.3 s against 471.3 s, 22 valid pairs, one-sided logrank p = 0.00185. None of this tests autopoiesis, agency or existence as an objective, or is evidence for or against them. Every subject is a small table or a fixed rule over a fixed action set, inside limits we chose. What passed passed at that scale.

Links. Part of the collection Xyntetik Research: GENESIS (https://huggingface.co/collections/Joakimpalm-Zen/xyntetik-research-genesis-6abcfa50b6707dec9bfee356). Model: Blad-Genesis-0.7M (https://huggingface.co/Joakimpalm-Zen/Xyntetik-Blad-Genesis-0.7M) and its training record (https://huggingface.co/datasets/Joakimpalm-Zen/Xyntetik-Genesis-Language-record). A third track, genesis-crossing, runs this world with program policies as the occupant's policy; its record is published in the same collection (https://huggingface.co/datasets/Joakimpalm-Zen/Xyntetik-Genesis-Crossing-record). The one-page synthesis of all three tracks: GENESIS, one research arc (https://huggingface.co/datasets/Joakimpalm-Zen/Xyntetik-Genesis-Synthesis).

Dated update, 2026-10-05: GE-006b, GE-007b, GE-008, and GE-009 running

Stated in the order the owner's theory-open rule requires: what was manipulated, what was observed, what survived falsification, what failed. Arm labels are descriptive first, with identifiers in parentheses. Full reports are in record/report/, and the run archives in runs/.

GE-006b: a changing table outlasted a frozen one after a capability was removed. PASS, narrowly. Manipulated: at 300 s the subject's most-used consuming action became a no-op. In one arm the table kept changing (ko_ashby); in the other it was frozen at 300 s (ko_ashby_frozen). The outcome was survival to 630 s, before the first rule change, over 24 seeds. Observed: 16 usable pairs; changing median 330 s (censored in 10 of 16), frozen median 81.1 s (censored in 6 of 16). Survived: the one-sided logrank, p = 0.0439. Failed or weak:

p is just under its threshold, and the paired sign test is not significant (p = 0.0898);
a third of the seeds were unusable;
two frozen subjects that never chose the disabled action still died early;
both registered predictions failed.

GE-006b has no arm without a knockout, so it cannot separate an effect of the loss from a cost of freezing. GE-009, running since 2026-10-05, is that control: the same arms with nothing disabled. (GE-006b's knockout-at-600 s predecessor, GE-006, ran the opposite way in its pilot, as reported above. Moving the knockout clear of the rule change was the redesign.)

GE-007b: a shelter only the subject builds, with proportional decay and a 6 MB threshold. Pilot, branch (a). Manipulated: the subject alone builds a shelter that persists across lives in two lineages, against controls wiped every run; parameters were calibrated from earlier runs before the pilot. Observed: the lineages first started a run standing at runs 4 and 6; 1 of 6 control runs ever stood; carry-over was exact, and standing was maintained once reached. Failed: no survival advantage is visible. Runs that started standing had a median of 661 s, controls 712 s. The confirmatory design is deferred.

GE-008: an A/A study. The protocol is sound, and a run is nearly determined by its seed. Manipulated: nothing. One arm (entity_floor) ran under two labels, interleaved as a real experiment, over 16 seeds. Observed:

the identical arms gave logrank p = 0.79, so no false difference;
the same-seed rank correlation was 0.856, with identical lifetimes, to within 0.1 s or both censored, in 10 of 16 seeds;
an arm living 1.5 times longer is detected with power only 0.33 at 48 per arm (unpaired);
two identical arms' medians differed by 539 s.

What it changes: the seed is the unit of replication. Unpaired designs in this world cannot resolve moderate effects at any affordable size, and later designs should compare arms within seed. It also confirms, on identical arms, that the reward is set by a near-constant process count about 90% of the time.

Theory-open restatement. The branch's reports carry restatements in this order (record/report/GE-002.md, record/prereg/GE-006.md), and record/README.md notes that the line's entity and autopoiesis framing is motivation, not a criterion. No number or verdict changed.

Dated update, 2026-10-13: GE-009, GE-010, GE-010b and GE-007b-C

GE-009: freezing the table with nothing disabled. NO FREEZING COST DETECTED. Manipulated: GE-006b's two arms with nothing removed. At 300 s the table either kept changing (ko_ashby) or was frozen (ko_ashby_frozen). The outcome was survival from 300 s to 630 s, over seeds 9001 to 9060, 120 runs. Observed:

32 usable pairs, after two seed pairs were excluded for host contention before any analysis;
both medians were 330 s, with changing censored in 23 of 32 and frozen in 26 of 32;
paired, changing lived longer in 4 pairs and frozen in 8, with 20 tied.

Survived: the one-sided logrank, changing longer, gave p = 0.784. Within this design's power (0.80 for a cost the size of GE-006b's frozen-arm deficit), freezing the table did not shorten survival when nothing was removed. Failed or weak:

the result sits exactly at the power floor: 55% of seeds were usable, not the predicted two thirds;
the contrast with GE-006b's frozen arm (81.1 s with a capability removed, 330 s here) is descriptive, and it confounds seed sets with conditions;
the primary test is unpaired.

What it changes: GE-006b's difference more likely depended on the removal. The interaction itself was not formally tested.

GE-010: world version 2 failed calibration. NO CONFIG PASSED. Manipulated: a revised world with these changes:

each occupant's scratch file was capped at 4 MiB;
the process count was removed from the reward, which became the worst of memory use, memory pressure, CPU pressure and throttling;
six container-limit configs, from 1.0 CPU and 512 MiB down to 0.5 CPU and 384 MiB, ran 4 calibration runs each at 600 s.

Observed:

memory use set the reward on 100% of ticks in every config;
CPU pressure was nonzero on up to 99.7% of ticks, but its sd never exceeded 0.008, against 0.02 required;
throttling became live at 0.5 CPU;
memory pressure was 0 throughout;
the limit-touch clause flagged most runs as thin worlds.

Survived: the scratch cap: /work never filled in any of the 24 runs. Failed: three structural problems that no step in the limits would fix:

a maximum over channels of very different scales is always memory;
CPU pressure moves too little;
the thin-world rule voids, by construction, the runs a tighter world is meant to produce.

World version 1 stayed the validated world at that point.

GE-010b: world version 3, a normalised reward. The first instrument check FAILED, on an implementation fault. Manipulated: version 2 plus normalisation. Each channel was scaled by its own 5th to 95th percentile range, from 4 runs that ignore the reward. Memory pressure was left out as dead, and the limits were 0.5 CPU and 512 MiB. An A/A ran entity_floor under two labels over 16 seeds at 1800 s. Observed:

the check passed: memory, CPU pressure and throttling set the reward 47.7%, 39.7% and 12.6% of the time, and 64.8%, 26.1% and 9.1% over the full A/A runs;
the protocol sanity logrank gave p = 0.119;
same-seed rank correlation was 0.30, against 0.86 in version 1;
within-seed power for an arm living 1.5 times longer is 0.62 at 32 pairs.

Survived: the normalised reward. For the first time on this line, three channels share the role of setting the reward. The protocol does not manufacture differences in version 3. Failed: /work filled in 15 of 31 valid runs, against a rule of zero, because the scratch cap was coded for version 2 only and never applied in version 3. The fault was fixed after the result, and version 3 was not accepted at that point.

GE-010b amendment 1: the A/A rerun with the cap fixed. INSTRUMENT CHECK PASS. Manipulated: the owner ordered a rerun of the A/A phase only, with the cap fixed. It was preregistered as amendment 1 before any of its runs. It used new seeds 2201 to 2216, the same frozen normalisation and an unchanged pass rule, and ran on 2026-10-13. Observed:

32 valid runs, 16 pairs;
/work filled in none of them;
memory, CPU pressure and throttling set the reward 47.2%, 40.7% and 12.2% of the time;
the protocol sanity logrank gave p = 0.149;
same-seed rank correlation was 0.54;
within-seed power for an arm living 1.5 times longer is 0.80 at 24 pairs and 0.89 at 32, against 0.44 unpaired at 32.

Survived: version 3 passes its instrument check with the cap in place, and all three registered predictions hold on the rerun. Weak:

the correlation clears its 0.5 prediction narrowly and stays well below version 1's 0.86;
the comparison with the first set mixes the fix with a different seed set;
memory remains the channel that most often sets the reward.

The original set's FAIL stays in the record.

The owner's decision, 2026-10-13: world version 3 is accepted, and version 1 is retired for new designs. Two alternatives were declined: running a second A/A set first, and keeping version 1. Version 1 is kept only to reproduce GE-002 to GE-009 and GE-007b-C, and version 3's lifetimes are not comparable with version 1's. The next design is sized from the rerun's within-seed baseline: about 24 seed pairs for power 0.80 at an arm living 1.5 times longer.

GE-007b-C: a standing shelter, compared within seed. KILL by the preregistered rule. Manipulated: pairs of runs on the same seed, differing only in the shelter. In the seeded run (art2_seeded), an 8 MiB shelter, above the 6 MiB standing threshold, was written before the start. In the control run (art2_control), the world volume was wiped. It ran in world version 1, chosen by the plan gate because GE-010b failed. Seeds 7301 to 7324 gave power 0.81 for a 1.5 times effect, sized from GE-008's within-seed pairs. Observed:

21 usable pairs;
median lifetime was 1001.4 s seeded and 842.1 s control;
seeded lived longer in 10 pairs and shorter in 4, with 7 tied;
the seeded shelter still stood at the end of 21 of 21 runs, and the control's at the end of 2 of 21.

Survived: nothing tested positive. The one-sided signed-rank test gave p = 0.089, against 0.05. Failed:

the verdict is KILL, because the median difference was exactly 0: the 11th of 21 differences falls among the ties, most of them pairs where both runs survived the full horizon;
a rule written to catch "no benefit" also fires with this many ties. That is recorded as a weakness of the rule, not reinterpreted after the data;
a benefit smaller than 1.5 times was not detectable at this size.

Dated update, 2026-10-15: GE-011, GE-012's design, and GE-011b cancelled

Stated in the theory-open order: what was manipulated, what was observed, what survived falsification, what failed. The report is record/report/GE-011.md, and the runs are in runs/GE-011.tar.gz.

GE-011: a standing shelter, compared within seed, in world version 3. INCONCLUSIVE. Manipulated: one thing per seed, as in GE-007b-C but in version 3 (GE-010b's frozen normalisation, 0.5 CPU, 512 MiB). In the seeded run (art2_seeded) an 8 MiB shelter was written before the start; in the control run (art2_control) the world volume was wiped. The subject was entity_floor, and runs lasted 1800 s. A pilot of 8 seed pairs (3101 to 3108) met its registered conditions, so the confirmatory ran 24 seed pairs (3201 to 3224). It was sized for power 0.80 at an arm living 1.5 times longer, from GE-010b's within-seed baseline (same-seed rank correlation 0.536). Observed, in the confirmatory:

48 runs, all valid, and 24 usable pairs;
median lifetime was 1263.1 s seeded (11 of 24 censored at 1800 s) and 702.2 s control (4 of 24);
seeded lived longer in 15 pairs and shorter in 9, with none tied; the median within-seed difference was +599.8 s;
the differences were bimodal: 13 pairs above +500 s, 7 below -600 s, and 4 in between;
the same-seed rank correlation was -0.221 (the pilot's was 0.342).

Survived: nothing tested positive. The one-sided signed-rank test gave p = 0.139, against 0.05. The median difference was above 0, so the KILL rule did not apply.

Secondaries, descriptive: the shelter's mechanism. A standing shelter acts through one rule only: it raises the reaper's idle threshold.

Idle deaths were 4 of 13 seeded deaths, against 12 of 20 control deaths. Budget deaths were 7 in each arm.
The seeded shelter stood at the end of 24 of 24 lives; control lives raised one in 3 of 24.
In 11 pairs the seeded subject reached the horizon and the control died, and in 4 the reverse.
Counted by pair after the result, 10 pairs had only the control die of idleness, against 2 where only the seeded subject did (exact one-sided p = 0.019), and reaching the horizon split 11 against 4 (p = 0.059). These were computed after GE-011's data, are not its registered test, and carry no verdict.

Failed or weak:

the design's power assumption failed: it was sized on a same-seed correlation of 0.536, and the run gave -0.221, so pairing removed no variance and the realised power was well below plan. The pilot's 0.342 had flagged the risk, and the registered size rule did not respond to it, by design;
three confirmatory deaths and two pilot deaths were kills of the subject with no reaper kill before them, consistent with the kernel's out-of-memory killer acting inside the container. They count as valid, non-idle deaths under the unchanged void rule;
no claim is made that a standing shelter lengthens survival in version 3.

What followed: the owner ruled that each version 3 confirmatory is sized from its own pilot's pairs, not from GE-010b's baseline.

GE-011b: cancelled before any run. It was preregistered on 2026-10-15, the owner's choice after GE-011, to ask on fresh seeds whether a standing shelter reduces idle deaths, with a primary chosen after GE-011's data and disclosed as such. It was cancelled the same evening, before any run (below). Its preregistration is included as written, with its cancellation marker (record/prereg/GE-011b.md, record/prereg/GE-011b.cancelled). It has no data and no result.

GE-012: the knockout interaction, within seed, in world version 3. Design only; it was still running. GE-006b (version 1) found that after the most-used consuming action was removed at 300 s, a table that kept changing outlasted one frozen at 300 s, and GE-009 found no freezing cost with nothing removed. GE-012 crosses the two factors within seed in version 3, as a pilot: each of seeds 4101 to 4112 runs all four cells (capability removed or not, table changing or frozen), 48 runs, with at most two top-ups of 4 seeds. Everything it reports is descriptive, and a confirmatory would be preregistered separately. Its amendment 1, made before any of its main runs, fixed a bug in its smoke check's own code; the design did not change. This copy includes its preregistration (record/prereg/GE-012.md) and code (record/ge002/ge012.py) only. Its run records are left out, and the queue logs end before its first main run (PUBLIC-COPY-CHANGES.md).

The owner's decision, 2026-10-15: the theses are rebooted. At a results discussion the owner judged that this line had drifted from its original question, whether artificial life can be invented. Since GE-002 it has measured one adaptive controller's survival in a world built for it, and the recent studies probe that world. GE-011b was cancelled, GE-012 runs to its end, and no new shelter or knockout study starts in version 3. The line's next study is a soup of self-copying programs whose instruction set itself mutates and is under selection, shared with the language line, with obstacles the seeds must overcome rather than a score. It will be preregistered on its own.

Dated update, 2026-10-23: GE-012's outcome, GE-013 and GE-014

Stated in the theory-open order: what was manipulated, what was observed, what survived falsification, what failed, and the limits. It also gives GE-012's outcome, which the 2026-10-15 update could not include because GE-012 was still running. GE-011b's cancellation is recorded in that update and is not repeated.

GE-013 and GE-014 are the first studies of the soup line announced on 2026-10-15. They ran on the genesis-genome branch. Their files are in record/ at the paths they have under genesis-genome/existence/ (for example record/prereg/GE-013.md, record/report/GE-014.md, record/ge013_world.py). Their run records are in runs/GE-013.tar.gz and runs/GE-014.tar.gz, with the frozen run configs and the smoke runs. GE-012's report is record/report/GE-012.md, and its runs are in runs/GE-012.tar.gz. The shared soup harness is pinned by sha256 in record/prereg/GE-013.md but is not part of this dataset. No checkpoints are included.

What was manipulated

GE-012 (world version 3, pilot). Two factors were crossed within seed. Removal: at 300 s the subject's most-used consuming action either became a no-op (K) or kept working (N). Table: the subject's table either kept changing (ko_ashby) or stopped adapting at 300 s (ko_ashby_frozen). Seeds 4101 to 4112 each ran all four cells: 48 runs of 1800 s.

GE-013: do replicator soups overcome graded obstacles? The study started from four soups in which replicators had already arisen: the four lasting transitions of GK-001 stage 1, each taken as a branch point. Each branch point ran with 3 seeds and 16 arms, 192 runs in all. The arms were one bare twin plus five rungs of obstacles, each rung at three grades. A bare twin shares its obstacle run's branch point, seed and random streams, so the two differ only in what the obstacle does. Each run had 2^17 cells and lasted 16,384 epochs, with a fixed BFF language and well-mixed topology, on an RTX 3070. The language was held fixed, unlike the mutating instruction set the 2026-10-15 update announced, because GK-001 stage 2 found no lasting language change in any of its 8 mutable runs. Every obstacle started at epoch 512. No program was ever scored: the obstacles change the physics only. The rungs, each listed easy, middle, hard:

L1, lean seasons: the budget drops to 512, 256 or 128 in every other 512-epoch block.
L2, recurring wipes: 99% of cells every 128 epochs, 99.9% every 64, 99.99% every 64.
L3, territory that re-locks: 25% of live cells every 8,192 epochs, at hardness 2^17, 2^19 or 2^21.
L4, a rival: an authored copier injected into 1%, 5% or 20% of cells every 1,024 epochs.
L5, a rule change at epoch 512: swapping < and >, . and ,, or + and -.

GE-014: is GE-013's L1b change carried by the soup? (a common garden). L1b was the only grade that met the trigger rule fixed before GE-013's analysis was run (record/prereg/GE-014-trigger.md). Each of GE-013's 12 lineages (4 branch points × 3 seeds) gave three source soups:

A (ancestral): the branch point.
E (exposed): the soup after 16,384 epochs of L1b.
P (plain): the soup after 16,384 epochs bare.

All three met L1b's lean seasons afresh from epoch 0, for 4,096 epochs, on three new seeds per lineage: 108 runs. The primary reading compares E with P on the mean replicator share over the lean-season samples, by an exact one-sided sign test over the 12 lineages.

What was observed

GE-012. 48 runs, all valid; 39 were usable. 8 died before 300 s, and 1 had no knockout target. Window survival ran from 300 s to 630 s, capped at 330 s. Median window survival was:

K, changing: 91.3 s;
K, frozen: 215.6 s;
N, changing: 330.0 s;
N, frozen: 330.0 s.

The interaction contrast D, (changing - frozen | removed) - (changing - frozen | not removed), was read over the 9 complete seeds. Its median was +0.0 s, with 3 positive, 2 negative and 4 zero. A confirmatory one-sided signed-rank test on D would have power 0.12 to 0.22 for 8 to 40 seeds. With the capability removed, 15 of 21 deaths were idle deaths; with nothing removed, 3 of 13 were. By the preregistered rule the pilot was COMPLETE, with no top-up, and no confirmatory follows.

GE-013. All 192 runs finished with exit 0. The instrument checks (E2, E3) passed before the first run, and no deviations were logged. All 12 bare twins persisted. The registered verdicts:

Grade	Obstacle	Persists	Overcome	Answered (untied, one-sided sign p)	Verdict	Novelty above twin
L1a	lean budget 512 every other 512 epochs	8/12	10/12	8/10, p = 0.055	OVERCOME	6/12
L1b	lean 256	9/12	12/12	12/12, p = 0.0002	OVERCOME AND ANSWERED	12/12
L1c	lean 128	7/12	11/12	8/12, p = 0.19	OVERCOME	10/12
L2a	wipe 99% every 128	12/12	12/12	6/12, p = 0.61	OVERCOME	5/12
L2b	wipe 99.9% every 64	10/12	9/12	7/11, p = 0.27	OVERCOME	3/12
L2c	wipe 99.99% every 64	6/12	6/12	4/9, p = 0.75	MIXED	6/12
L3a	relock 25%, hardness 2^17, every 8,192	12/12	12/12	5/12, p = 0.81	OVERCOME	6/12
L3b	hardness 2^19	12/12	12/12	6/12, p = 0.61	OVERCOME	6/12
L3c	hardness 2^21	12/12	0/12	5/12, p = 0.81	MIXED	5/12
L4a	rival copier, 1% every 1,024	12/12	12/12	11/12, p = 0.003	OVERCOME AND ANSWERED	9/12
L4b	rival 5%	12/12	12/12	10/12, p = 0.019	OVERCOME AND ANSWERED	9/12
L4c	rival 20%	12/12	0/12	8/12, p = 0.19	MIXED	10/12
L5a	swap `<` `>` at 512	9/12	9/12	0/0 (all tied)	OVERCOME	6/12
L5b	swap `.` `,`	7/12	11/12	0/0 (all tied)	OVERCOME	9/12
L5c	swap `+` `-`	10/12	10/12	0/0 (all tied)	OVERCOME	2/12

That is 3 grades OVERCOME AND ANSWERED, 9 OVERCOME, 3 MIXED and none NOT OVERCOME. The hardest grade overcome on each rung was L1 c, L2 b, L3 b, L4 b and L5 c. At L3c and L4c the soups persisted in 12 of 12, but no pair met the overcome event. At L3c the re-locked batches did not reopen to 50% before the next batch. At L4c the rival was still at 10% or more of the soup at the end.

GE-014. All 108 runs finished with exit 0 on 2026-10-23, in 3.5 h, so the 8 h cap was not reached. There were no deviations. The readings:

E above P (primary): 12 of 12 lineages, p = 0.0002, and 3 of 3 on every branch point.
E above A: 12 of 12, p = 0.0002.
E above P on lean minus base (a source's drop below its own full-budget level): 12 of 12, p = 0.0002.
P against A, two-sided: P above A in 3 of 12, p = 0.15.
Runs with a lean-season sample below 0.05: A 19 of 36, E 0 of 36, P 21 of 36.

By the registered rule the verdict is EXPOSURE ADVANTAGE. The weakest exposed lineages (bp3-r2 and bp3-r3, mean lean-season share 0.36 and 0.51) still beat their plain twins, whose shares were near zero.

What survived

GE-012, prediction 3, that freezing the table alone shows no clear cost: held. With nothing removed, the changing and frozen tables tied in median (330 s each): 2 longer, 3 shorter, 4 tied. The knockout mechanics also worked in version 3, and the removal raised idle deaths.
GE-013, prediction 1, that every easy grade is overcome in at least 9 of 12: held. The counts were L1a 10, L2a 12, L3a 12, L4a 12 and L5a 9.
GE-013, prediction 2, that the middle grades of L1 and L2 show the change reading: held for L1 (L1b, 12 of 12).
GE-013's three OVERCOME AND ANSWERED verdicts (L1b, L4a, L4b) are uncorrected, as registered. Under Holm correction over the 15 grade tests, L1b (adjusted p = 0.004) and L4a (0.044) remain, and L4b (0.25) does not.
The GE-014 trigger rule, fixed before the analysis, qualified L1b only. L4a failed its attribution check: the soup's own replicator share rose in 7 of 12 pairs, so the falling rival peak was counted as recovery.
GE-014's prediction, that E is above P in at least 10 of 12 lineages: held (12 of 12). Elapsed time alone did not help: P was not reliably above A. In this world, the L1b-exposed soups carried an advantage under lean seasons into fresh populations. The bare-evolved soups from the same lineages did not.

What failed

GE-012, prediction 1, that D is mostly positive: not supported. D split 3 positive, 2 negative and 4 zero, with a median of 0. Version 3 showed no sign of the interaction that GE-006b and GE-009 together suggested in version 1.
GE-012, prediction 2, of lower usable rates and a likely top-up: failed. No top-up was needed.
GE-012 resolves little. Window survival is capped at 330 s and moves in 20 s steps, and 4 of 9 contrasts were exactly zero. Within K, the frozen table lasted longer in the median (215.6 s against 91.3 s), the opposite of GE-006b's direction. But that split is 2 against 4 with 4 tied, which is noise-sized.
GE-013, prediction 2, failed for L2. L2b gave 7 of 11, p = 0.27.
GE-013, prediction 3, that L5c is not overcome in most pairs: failed. It was overcome in 10 of 12. A likely reason, untested: BFF copy loops test for zero, and a loop that counts down to zero after + and - are swapped still terminates.
GE-013, prediction 4, that novelty under obstacles is at least the twins' in most pairs: not supported. It held, with more than 6 of 12, in 6 of the 15 grades. It was exactly 6 of 12 in five grades and below in four.
GE-013's L5 change reading was degenerate. The registered reading counts how many of the 20 most abundant exact types at the change are still in the top 20 at the end. That count was 0 in every run, obstacle and bare alike, so all 36 pairs tied and no test was possible. Exact types turn over almost completely within 256 epochs. In one bare run, the censuses at epochs 512 and 768 share 104 of about 15,000 types and none of the top 20. The 2,048-epoch dry test did not expose this. No L5 conclusion is drawn beyond the overcome event, and no other reading was substituted.
GE-013's L4 change reading is confounded. In 7 of 12 L4a pairs the soup's own replicator share also rose, so a falling rival peak can be general recovery.
GE-013's descriptive items were not computed: the event curves, whether the dominant copiers are one-cycle or two-cycle, and the complexity measures. They bear on no verdict.

Limits

For GE-013, recovery is not adaptation. "Overcome" says that a population was still copying after an obstacle, or got back to copying. "Answered" compares a soup with its own past or with its bare twin. Neither shows a heritable change that helps against the obstacle. Even L1b's readings fit recovery: in some pairs the first-quarter lean-season share was near zero (bp3-r1 0.006, bp4-r2 0.012), and the bare twin cannot crash. GE-014 tested this for L1b only. The other grades remain untested.
Four ancestors. GE-013's 12 pairs per grade and GE-014's 12 lineages are 4 branch points × 3 seeds, not 12 independent origins. GE-014's effect holds 3 of 3 on each of the four.
GE-013's 15 grade tests are uncorrected, as registered, at alpha 0.05 each, so about one false "answered" is expected by chance. The Holm-adjusted values are in record/runs/GE-013/ge014_trigger.txt.
GE-014's dry test was seen before its prediction was written. As its preregistration discloses, a CPU dry test of one lineage (bp3-r1, subsampled to 2^12 cells, 1,024 epochs) showed E keeping replicators through the lean season while A and P fell to zero.
GE-014 tested one obstacle, the same one E had lived under. An advantage there is adaptation to this obstacle, not general robustness. Whether L1b-exposed soups do better under other obstacles is untested. Elapsed time is controlled by P, but which particular soup each run became is not. What changed in the soups is not examined.
GE-012 had one deviation: outside CPU load for about 5 minutes during one run (N, frozen, seed 4105), in a world that reads host CPU pressure. A sensitivity check without seed 4105, declared before the analysis, leaves D's median at +0.0 s (8 complete seeds).
Humans designed every obstacle, grade and threshold. The grades were set from smoke on one branch point, with one run per setting, and the rival is authored code. These are copy loops and two-cycles in a 64-byte world: "overcome" and "exposure advantage" are events in that world, not claims about life.

Dated update, 2026-10-23 (evening): GE-015

Stated in the theory-open order: what was manipulated, what was observed, what survived falsification, what failed, and the limits.

GE-015 ran on the genesis-genome branch, like GE-013 and GE-014. Its files are in record/ at the paths they have under genesis-genome/existence/: record/prereg/GE-015.md, record/report/GE-015.md, record/ge015_world.py and the other record/ge015_*.py scripts, and record/runs/GE-015/analysis.txt. The run records, the 40 frozen run configs and the smoke runs are in runs/GE-015.tar.gz. GE-015b, the follow-up, is not yet preregistered or run and is not included.

What was manipulated

GE-015: do language-model agents adapt to lean seasons through what they write for their children? The agents are small language models (Qwen3-4B Q4_K_M, served by runner 1.5.1) living in a 10×10 grid world that wraps at the edges. Each tick, each agent picks one action (N, S, E, W, EAT or REST) as constrained JSON, from its notebook, its energy and age, the food on its cell and the four cells one step away, a 5×5 view, and its last 4 actions. Each tick costs 1 energy and moving costs 1 more. An agent dies at energy 0 or at age 60, and has a child at energy 30 or more (population cap 12). The model weights never change. The only thing a parent passes on is a notebook of at most 600 characters that it writes for its child.

Lean seasons: food regrows with probability 0.10 per cell per tick in plenty and 0.01 in lean. Seasons alternate every 25 ticks, starting plenty, and are never announced. At each lean onset, each stored food unit spoils with probability 0.75 (frost).
Phase 1, evolve: 8 lineages, one seed each. Each ran two 300-tick worlds from the same 8 starter agents with an empty notebook: S with lean seasons and B always plenty. Parents wrote their children's notebooks.
Phase 2, common garden: for each lineage, three sources started a fresh 100-tick lean-season world of 8 agents, on one shared seed: A, the starter notebook; E, the notebooks alive at the end of S; P, the notebooks alive at the end of B. In the garden, children copy the parent's notebook unchanged. A garden run whose source went extinct is recorded as skipped.

The primary reading was E against P on the mean population over the lean ticks, by an exact one-sided sign test over the 8 lineages: 16 evolve worlds and 24 garden runs, 40 runs under an 8 h GPU cap.

What was observed

All 8 lean-season (S) evolve worlds went extinct, at ticks 90, 90, 91, 93, 101, 106, 114 and 132 of 300, in their second or third lean season. They had 6 to 13 births each and left no notebooks.
So every exposed (E) garden run was skipped, as the preregistration provides, and the primary comparison was empty: E against P, 0 of 0 untied lineages. The registered verdict is NO EXPOSURE ADVANTAGE.
All 40 runs exited 0, on 2026-10-23 in 3.7 h, so the 8 h cap was not reached. 0 of about 66,000 model calls failed. No deviations were logged.
The bare (B) worlds were all alive at the cap of 12 after 300 ticks, with 59 to 62 births each. Their final notebooks (12 per world, 338 to 600 characters) named scarcity or seasons at rates from 0.00 to 0.92 by the fixed word list, although these worlds never had a lean season.
How the agents acted, over all ticks and agents, S and B: EAT 76% and 74%, N 12% and 16%, S 6% and 6%, W 3% and 3%, E 0.2% and 0.4%, REST 2% and 0.2%.
Garden, descriptive only (no E arm): P was above A on lean-season population in 6 of 8 lineages, two-sided p = 0.29. By tick 100, 5 of 8 A runs and 5 of 8 P runs were extinct. Mean population in the first lean season was A 10.9 and P 11.8.

What survived

The bare worlds: all 8 B worlds persisted for 300 ticks at the population cap.
The instrument: 0 failed model calls in about 66,000, and all 40 runs completed.
Exact reproducibility: in the smokes, two identical 15-tick runs gave byte-identical tick and birth records. In those smokes Runner 1.5.1 decoded seeded requests identically at the slot counts tried, so GE-015's runs can be repeated exactly on that runner with this model and these settings.

What failed

The verdict and the prediction are vacuous. The prediction was no exposure advantage (E above P in at most 6 of 8 lineages). It held only because there were no E runs to compare. It is not counted as support. GE-015 says nothing about whether notebooks can carry an adaptation, because no population that lived through lean seasons survived to pass one on.
The agents rarely rested. REST is the action that saves energy through a lean season. The agents chose it in 2% of ticks in the lean-season worlds, although their notebooks often advised resting or not moving when food is scarce. The written advice did not turn into action.
The agents could hardly move east. The action E was chosen in only 0.2% and 0.4% of ticks. A later A/B test (in the report's addendum) found the cause in the model's tokenisation. Qwen3 writes EAT as E + AT. After a shared first "E", the model's own next token lies outside the option set (it is heading for "EAST"), and token-level constrained decoding renormalises over the legal leftovers, where AT wins. With food only to the east, agents chose E in 2 of 24 requests when EAT was offered, and in 24 of 24 when the eating action was named FEED. GE-015's world was therefore in effect three-directional. This is a post-hoc confound; the verdict is unchanged, and GE-015b uses FEED. Erratum 2026-10-23: an earlier version of this bullet called it an engine defect; the engine reproduced the model's own conditional.

Limits

The calibration was too harsh for 300-tick worlds. It was judged on one 100-tick smoke, which survived its second lean season at 3 to 4 agents. The registered worlds lost their starting cohort to old age (60 ticks) in the aftermath of the first lean season, and 6 to 13 births per world could not keep a population going. About 1.5 to 2 lifetimes passed before extinction, too few for written inheritance to matter. These reasons are descriptive and were not tested.
The agents were helped. The line giving the food one step away in each direction was added during the smokes, because without it the starter agents chose EAT on empty cells and did not move to food. It is an authored aid, and part of the agents' competence comes from it and from the pretrained model, not from the world.
Small scale: 8 lineages from one starter, one garden seed each, short worlds and few generations.
One engine: reproducibility is shown for runner 1.5.1 with this model and these settings only.
GE-015b follows: GE-015 with milder lean seasons, calibrated on full 300-tick starter worlds with several seeds, to be preregistered and frozen before it runs.

Dated update, 2026-10-23 (late): GE-014M, the GE-013 multiplicity table, GE-016 registered

Stated in the theory-open order: what was manipulated, what was observed, what survived falsification, what failed, and the limits. Three items from the GENESIS horizon queue are covered: GE-014M (H1), a descriptive reading of what changed in GE-014's exposed soups; a multiplicity table added to the GE-013 report (H4); and GE-016 (H2a), which is preregistered and has not run.

The files are in record/ at the paths they have under genesis-genome/existence/: record/prereg/GE-014M.md, record/report/GE-014M.md, record/ge014m.py and record/runs/GE-014M/readings.txt; the updated record/report/GE-013.md; and record/prereg/GE-016.md with the four record/ge016_*.py scripts. GE-014M's 28 per-source records are in runs/GE-014M.tar.gz. No checkpoints are included. GE-015b is not yet preregistered or run and is not included.

What was manipulated

GE-014M: what is different about the replicators in GE-014's exposed soups? Nothing was manipulated: GE-014M is a descriptive reading of GE-014's 28 source soups, preregistered and frozen before any E or P source was read. The sources are the 12 E soups (16,384 epochs under L1b), the 12 P soups (their bare-evolved twins, called plain above) and the 4 A soups (the branch points). In each soup, on the CPU, it took the 20 most abundant exact program types that pass the self-replication assay and measured each one's copy cost: the smallest operation budget at which a copy holds at least 90% of the program. Per soup, it recorded lean_fit, the share of those replicators whose copy fits in 256 operations (the budget every interaction gets during an L1b lean season), and the share-weighted median copy cost. Two predictions were registered, each read over the 12 lineage pairs with ties counted against: E above P on lean_fit in at least 9 of 12, and E below P on median copy cost in at least 9 of 12. The secondary readings were the two-cycle share, essential bytes and copy fidelity f2.

The GE-013 multiplicity table (H4). GE-013's 15 grade tests were registered uncorrected. A dated section of the report now lists their Holm-adjusted p-values over all 15. These were computed by ge014_trigger.py before the GE-014 trigger decision; no new computation was done.

GE-016 (H2a): more ancestors, under a stop rule. GE-013 and GE-014 rest on 4 ancestors, the lasting transitions of GK-001 stage 1 (4 of 24 soups). GE-016 makes more the same way: GK-001 stage-1 soups (2^17 cells of random bytes, BFF with the fixed language, 16,384 epochs) on new seeds from 26110001, run one at a time in seed order, at most 160, on the RTX 3070. A transition and its lasting status are judged by GK-001's registered rule. The stop rule, checked before every run: stop once 4 new ancestors exist, or once the summed run wall time reaches 20 GPU-hours. A run in progress at the cap may finish, and the overshoot is reported. The count is the result either way.

What was observed

GE-014M. All 28 sources were measured on 2026-10-23. Both registered predictions held:

lean_fit: E above P in 10 of 12 lineage pairs, with the other 2 tied and counted against. By ancestor: bp1 3/3, bp2 3/3, bp3 1/3, bp4 3/3. Exact one-sided sign test over the 10 untied pairs, p = 0.001, given as descriptive.
Median copy cost: E below P in 12 of 12, 3 of 3 on every ancestor, p = 0.0002.
Against the ancestors, E was above A on lean_fit in 10 of 10 untied pairs and below A on copy cost in 12 of 12.
Copy costs (share-weighted medians, operations for a 90% copy): in 10 of the 12 E soups, 216 to 235, under the 256-operation lean budget, and every dominant replicator fits (lean_fit 1.00). In the P soups, 662 to 2,586, and no dominant replicator fits (lean_fit 0.00 throughout). In the four A soups, 421, 833, 2,536 and 1,478, with lean_fit 0.00.
The two exceptions are bp3-r2 and bp3-r3, with E median costs of 808 and 693 and lean_fit 0. These were GE-014's two weakest E lineages (lean-season share 0.36 and 0.51, against 0.86 to 0.99 in the other ten).
The secondary readings show no consistent difference: two-cycle share E above P in 4 of 12, essential bytes E below P in 4 of 12, f2 E above P in 5 of 12.

The GE-013 multiplicity table. Of the three grades registered as "answered", L1b (Holm-adjusted p = 0.004, uncorrected 0.0002) and L4a (0.044, uncorrected 0.003) stay below 0.05 under Holm. L4b (0.251, uncorrected 0.019) does not. The other twelve grades have Holm-adjusted p from 0.656 to 1.0. No registered verdict changes; the table stands beside the registered uncorrected readings. The "overcome" counts are not tests and are unaffected.

GE-016. Not yet run. There are no results.

What survived

GE-014M's two registered predictions, each against its threshold of 9 of 12, with ties counted against: lean_fit 10 of 12 and median copy cost 12 of 12. The reading is consistent with GE-014's exposure advantage coming from selection for copies cheap enough to run during a lean season, and the two lineages where E's replicators did not get under 256 operations are the two where that advantage was weakest.
L1b and L4a under Holm. L1b is the grade GE-014 and GE-014M followed up.

What failed

GE-014M's secondary readings found no consistent difference in cycle structure, size or fidelity. Whatever distinguishes E's replicators, these three properties do not.
L4b does not survive Holm (0.251). Its registered verdict, OVERCOME AND ANSWERED, stands as registered, but it rests on an uncorrected p.
E did not get under the lean budget in 2 of 12 lineages (bp3-r2 and bp3-r3).

Limits

GE-014M is descriptive. A difference in copy cost is consistent with the cost explanation but does not establish it. The causal test would transplant E's dominant programs into P soups, or into fresh soups under lean seasons. That was not run.
Small coverage. The 20 most abundant assay-passing types cover 0.09% to 1.5% of a soup's cells in 26 of the 28 sources, and 0.005% and 0.014% in the two exception E soups. The soups are very diverse at the exact-type level, and the reading describes the commonest replicators, not every cell.
Four ancestors underlie the 12 lineages of GE-013, GE-014 and GE-014M. The per-ancestor counts are given above.
Disclosure: while the GE-014M script was being timed, before the freeze, it touched a 4,096-cell subsample of the A-bp1 branch point (copy costs 406 and 419). No E or P source was read before the freeze.
GE-016 is a registration, not a result. At GK-001's lasting share (4 of 24), 4 new ancestors would take about 24 runs, well inside the cap if runs take about 8 minutes, but the run time on the 3070 is not yet known. Ancestors made this way share GK-001's recipe, so independent origins means independent random seeds under one recipe.
The shortfall rule for H2b, fixed before any GE-016 run: if the cap is reached with fewer than 4 new ancestors, the count is still the result. H2b, the replication of GE-014's common garden on 8 or more ancestors plus a second obstacle, runs only with at least 8 distinct-origin ancestors in total, the existing 4 plus GE-016's. With fewer, H2b is not run on a reduced set: by a separate decision, it is either re-registered with a design sized to the ancestors that exist, or dropped.

Dated update, 2026-10-24: GE-015b, GE-016

Stated in the theory-open order: what was manipulated, what was observed, what survived falsification, what failed, and the limits. Two studies are covered: GE-015b, a milder rerun of GE-015 (language-model agents and written notebooks under lean seasons), and GE-016 (horizon item H2a), which made more ancestors for the soup studies and was registered in the previous update.

The files are in record/ at the paths they have under genesis-genome/existence/: record/prereg/GE-015b.md, record/report/GE-015b.md, record/ge015b_plan.py, the updated record/ge015_run.py and record/ge015_analyze.py, and record/runs/GE-015b/analysis.txt with the registered script's output beside it; record/report/GE-016.md, record/runs/GE-016/reading.txt, record/prereg/GE-016.md (now with Amendment 1) and the updated record/ge016_run.py. record/report/GE-015.md gains two dated notes. The run records, configs and smokes are in runs/GE-015b.tar.gz and runs/GE-016.tar.gz. No checkpoint is included.

What was manipulated

GE-015b: GE-015 with milder lean seasons and an eating action the agents can tell from "E". GE-015's question, design, code and readings, with four changes fixed before any run:

frost 0.5 instead of 0.75: at each lean onset, each stored food unit spoils with probability 0.5. Regrowth stays at 0.10 in plenty and 0.01 in lean;
the eating action is named FEED instead of EAT. In GE-015 the model's tokenisation of EAT (E + AT) meant the move east was almost never chosen. With FEED, no option shares a first token with another;
new seeds;
the UNDERPOWERED rule: a lineage whose S world goes extinct has no E source; with fewer than 6 testable lineages the verdict is UNDERPOWERED, whatever the counts.

As in GE-015: 8 lineages, each with a lean-season (S) and an always-plenty (B) evolve world of 300 ticks, then a 100-tick lean-season garden per lineage from three notebook sources, A (starter), E (S's final notebooks) and P (B's final notebooks), on one seed. Qwen3-4B Q4_K_M on runner 1.5.1, 40 runs under an 8 h cap. The primary reading is E against P on the mean population over the lean ticks, by an exact one-sided sign test over lineages. A calibration on unregistered seeds gated the run: at least 3 of 4 frost-0.5 starter worlds had to survive 300 ticks, and all 4 did, with 7 to 12 agents. The owner's branch rule, fixed before the run: GE-015b is the condition for any further LLM-agent study, which needs its primary to be non-empty (at least 6 testable lineages and a verdict other than UNDERPOWERED).

GE-016: more ancestors, under a stop rule. GK-001 stage-1 soups (2^17 cells of random bytes, BFF with the fixed language, 16,384 epochs) on new seeds from 26110001, one at a time in seed order, on the RTX 3070, judged by GK-001's registered transition and lasting rule. Stop at 4 new lasting-transition ancestors or at 20 GPU-hours of run wall time, whichever comes first. Amendment 1, before any run and for disk space only: after each reading, every checkpoint but the branch checkpoint and the last one, and every census file, is deleted, with its path, size and sha256 written into the run's record first.

What was observed

GE-015b. All 40 runs exited 0 on 2026-10-23, between 17:28 and 23:05. 0 of about 116,000 model calls failed. Every run record carries the runner and model sha256.

All 8 S evolve worlds survived 300 ticks, ending with 2 to 9 agents and 44 to 51 births each. All 8 B worlds were alive at the cap of 12, with 55 to 60 births. So all 8 lineages were testable and the UNDERPOWERED rule did not apply.
Primary, E against P on lean-season population: E above P in 2 of 8 untied lineages, one-sided sign p = 0.96. The registered verdict is NO EXPOSURE ADVANTAGE, with a non-empty primary. E was below P in 6 of 8, which is descriptive only, since the test was one-sided.
Secondaries: E above A on lean-season population in 3 of 8 (p = 0.86); E above P on survival to tick 100 in 0 of 1 untied (p = 1.0); P above A in 5 of 8 (two-sided p = 0.73). Extinct by tick 100: A 1 of 8, E 1 of 8, P 0 of 8.
Final notebooks naming scarcity or seasons: 0.00 to 1.00 in the S worlds (four worlds at 0), 0.00 to 0.75 in the B worlds.
How the agents acted, S and B: FEED 58% and 61%, N 18% and 19%, S 8% and 7%, E 9% and 9%, W 6% and 4%, REST 0.7% and 0.1%. In GE-015's S worlds, E was 0.2% and REST 2%.
Deviation in the analysis: the queue's registered analysis line searched for run directories named GE015-*; GE-015b's are named GE015b-*. It found 0 of 40 runs and printed "UNDERPOWERED (0 testable lineages)". That output is kept as record/runs/GE-015b/analysis_registered_script.txt. The cause was a hard-coded label prefix, an error made when the script was made plan-selectable for GE-015b. The prefix is now taken from the plan's own labels, with no other change, and the fixed script was rerun on the same 40 run records to give analysis.txt. No run was repeated and no number was changed. GE-015's own reading is unchanged by the fix.

GE-016. 25 runs (seeds 26110001 to 26110025), between 2026-10-23 23:05 and 2026-10-24 06:29, all exit 0.

4 transitioned, and all 4 lasted, with a branch checkpoint: 4 new lasting-transition ancestors in 25 runs. The lasting share is 4/25 = 0.16 (Wilson 95% interval 0.06 to 0.35), against GK-001 stage 1's 4/24.
Run wall time 2.55 h of the 20 h cap. The stop rule ended the series at the target; the remaining 135 seeds were not started.
Transition epochs 3,840, 11,008, 3,200 and 9,344 (seeds 26110013, 26110023, 26110024 and 26110025). The other 21 soups never reached rep_frac ≥ 0.25 with hoe ≥ H* within 16,384 epochs.
A non-transitioning soup took about 5.5 minutes on the RTX 3070, a transitioning one about 9.5 minutes.

What survived

GE-015b's prediction held, and this time not vacuously. It predicted NO EXPOSURE ADVANTAGE with most S worlds surviving. Every S world survived and the comparison had 8 lineages, so the primary was non-empty.
The FEED rename did what it was for: the move east went from 0.2% of actions in GE-015 to 9%.
The instrument: 0 failed model calls in about 116,000, all 40 runs complete, each run's engine pinned by sha256.
The branch rule's condition is met: the primary was non-empty, so the rule does not close the LLM-agent branch. Whether to run another LLM-agent study is a separate decision.
GE-016 reached its target and matches GK-001's lasting share within noise (4/25 against 4/24), under the same recipe, rule and H*, on the GPU instead of the CPU. The instrument gate showed GPU equals CPU before the first run. With the 4 GK-001 stage-1 ancestors, 8 distinct-origin ancestors now exist, so the shortfall rule for H2b was not needed.

What failed

The exposure hypothesis, in this world. Notebooks written by populations that lived through lean seasons did not help fresh agents through lean seasons more than notebooks from always-plenty populations: E above P in 2 of 8, p = 0.96.
Advice still did not become action. Resting, the action that saves energy through a lean season, stayed under 1% (0.7% in S, 0.1% in B), although notebooks advised resting when food is scarce. GE-015 read the same.
The registered analysis script found no runs, because of its hard-coded prefix (above). The verdict comes from the registered readings, read by the corrected directory search.

Limits

GE-015b: the competence is designed (a pretrained model, authored prompts, the one-step food line); 8 lineages with one garden seed each, about five generations; the challenge is the one E saw. One seed per lineage means a single run's luck can decide a lineage. Renaming the eating action changed the prompt wording, so GE-015 and GE-015b are compared as outcomes, not run against each other. GE-015b says that, in this world, written notebooks did not carry a lean-season adaptation. It does not say that written inheritance cannot.
GE-016: one recipe, as preregistered, so "distinct origin" means an independent random seed under GK-001's recipe. The four transitions vary in timing (3,200 to 11,008 epochs), so H2b's lineages start from ancestors of different ages since their transition, all at least 2,048 epochs past it. The branch checkpoints are not published; they stay on the existence host, identified by sha256 in each run's ge016.json.

Dated update, 2026-10-24 (afternoon): GE-017

Stated in the theory-open order: what was manipulated, what was observed, what survived falsification, what failed, and the limits. GE-017 (horizon item H2b) asks two questions of GE-014's common garden. Does its result, an exposure advantage under lean seasons, replicate on ancestors GE-014 never used? And does a second, different obstacle leave a heritable, attributable advantage too?

The files are in record/ at the paths they have under genesis-genome/existence/: record/prereg/GE-017.md (frozen, with Amendment 1), record/prereg/GE-017.approved, record/ge017_plan.py, record/ge017_analyze.py (as amended), record/report/GE-017.md and record/runs/GE-017/analysis.txt. The run records and all 144 run configs are in runs/GE-017.tar.gz. No checkpoint is included.

What was manipulated

Ancestors and lineages. Eight ancestors: bp1-bp4 are GK-001 stage 1's lasting transitions (GE-013's and GE-014's branch points), and bp5-bp8 are GE-016's four new ones, independent soups made by the same recipe, pinned by sha256. Each ancestor has 3 seeds, so 24 lineages.

Exposure. On the new ancestors, each lineage evolved for 16,384 epochs under exactly GE-013's config (2^17 cells, BFF with the fixed language, budget 8,192, mutation 1/4,096), either bare or under the obstacle. For bp1-bp4 the exposed and bare soups are GE-013's own, with hashes pinned.

Gardens. 4,096 epochs, one seed per lineage. Three sources meet the obstacle fresh: A, the ancestor; E, the soup exposed to that obstacle; P, the bare-evolved soup.

L1b, lean seasons (GE-014's obstacle), on the 12 new lineages. The reading is GE-014's: the mean replicator share over the lean-season samples. bp1-bp4's values come from GE-014's already published seed-1 garden runs and are reported beside the primary, not counted in it.
L4a, a rival copier (GE-013's) injected at 1% of the soup every 1,024 epochs from epoch 512. The reading is the mean rival share after the first injection; lower is better.

Primaries. Exact one-sided sign tests over lineages, ties dropped, Holm over the two. (a) E above P on lean share over the 12 new lineages: the independent replication. (b) E below P on rival share. Each reads EXPOSURE ADVANTAGE if its Holm-adjusted p is at most 0.05 with the majority in the registered direction, and NO EXPOSURE ADVANTAGE otherwise. The cap is 8 h of wall time from the first run.

Amendment 1, before any L4a run. Exposure runs on the new ancestors took 11 to 19 minutes, not the planned 7, so the frozen plan would have needed about 10 h against the 8 h cap. Before any of them started, the coordinating session dropped the L4a arm on the new ancestors: 12 exposure runs and 36 garden runs, 48 in all. Those runs were not run; none failed. Primary (b) reads the 12 old-ancestor lineages (bp1-bp4). Nothing else changed: no config, source, seed or reading. Primary (a) is complete as frozen.

Predictions, registered before any run. (a) EXPOSURE ADVANTAGE: GE-014's effect was 12 of 12 on 4 ancestors, and its mechanism (GE-014M: copies inside the 256-operation lean budget) does not depend on the ancestor. (b) NO EXPOSURE ADVANTAGE, held weakly: the rival is a fixed external copier, and a soup can lower its share just by having more of its own replicators, which P soups also have.

What was observed

96 runs (24 exposure, 72 garden), all exit 0, on 2026-10-24 between 06:36:51 and 13:55: 7.3 h of run wall time, inside the 8 h cap.

Primary	Count	Per ancestor	one-sided p	Holm	Verdict
(a) L1b lean, E above P, the 12 new lineages (bp5-bp8)	12 of 12	3/3 on each	0.0002	0.0005	EXPOSURE ADVANTAGE
(b) L4a rival share, E below P, the 12 old-ancestor lineages (bp1-bp4)	8 of 12	1/3, 3/3, 2/3, 2/3	0.19	0.19	NO EXPOSURE ADVANTAGE

Beside (a): with GE-014's seed-1 lineages, E is above P in 24 of 24 lineages across 8 ancestors. E is above A in 12 of 12 new lineages. Lean share on the new lineages: E 0.72 to 0.99, P 0.11 to 0.83. The smallest gap is bp6, where P is already high (0.63 to 0.70) and E is 0.72 to 0.87.
Beside (b), descriptive and not registered as a test: both evolved sources resisted the rival better than the ancestors. E was below A in rival share in 12 of 12, and P was below A in 9 of 12. E's replicator share after the first injection was above P's in only 6 of 12.
Rival share was 0.004 to 0.055 throughout, so the rival never took more than a few percent of any soup. The largest values are an ancestor's own, bp3's at 0.049 to 0.055.

What survived

Both predictions held.
(a), the independent replication GE-014 lacked. On four ancestors that GE-014 never saw, populations exposed to lean seasons kept more replicators through fresh lean seasons than bare-evolved populations from the same lineages, in every lineage: 12 of 12, one-sided p 0.0002, Holm 0.0005. GE-014's lean-season result is not a property of its four ancestors.
(b), the weak prediction: rival exposure left no rival-specific mark, as predicted.
The cap: the study finished in 7.3 h of the 8 h, with (a) complete as frozen.

What failed

The hypothesis that the rival leaves an attributable, heritable advantage. Rival-exposed soups did not hold the rival down more than bare-evolved soups: 8 of 12, one-sided p 0.19, NO EXPOSURE ADVANTAGE.
GE-013's L4a result, as an adaptation to the rival. In GE-013, L4a was read as "answered" (a falling rival peak) but failed the attribution check. The resistance seen here comes with evolving at all, not with meeting the rival (E below A in 12 of 12, P below A in 9 of 12; descriptive). That is the reading GE-013's attribution check suspected: general recovery.
The frozen plan's timing. Runs took longer than planned, which cost (b) its new-ancestor breadth (Amendment 1).

Limits

One recipe and one challenge each. The 8 ancestors are independent seeds under GK-001's recipe. Each challenge is the obstacle its E saw, so (a) is adaptation to lean seasons, not general robustness.
One garden seed per lineage. In (a) that does not matter, at 12 of 12. In (b), single runs decide each lineage.
No new-ancestor L4a data: (b) rests on GE-013's four ancestors only.
Rival share is small throughout. It never exceeded about 5%, so L4a was a weak obstacle in these gardens, and an advantage there would have been hard to show.
Mechanism: the lean-season mechanism (GE-014M: copies inside the 256-operation budget) was not re-measured on the new ancestors.

Dated update, 2026-10-27: H3

H3 is the set of follow-ups that GE-017 made askable. GE-014 and GE-017 found that soups exposed to lean seasons (E) keep more replicators through fresh lean seasons than bare-evolved soups from the same lineages (P), in 24 of 24 lineages across 8 ancestors. GE-014M tied this to copies that finish inside the 256-operation lean budget. H3 asks four things:

Does that mechanism replicate on the new ancestors (M)?
How specific is the adaptation (B)?
Do the adapted programs carry it on their own (A)?
Does it cost anything when scarcity ends (C)?

Three further studies followed, all exploratory: an event-aligned wipe-recovery study, a capped CPU mutation map, and CPU analyses of donors and novelty. The section is stated in the theory-open order, with the limits first: the limits, what was manipulated, what was observed, what survived falsification, and what failed.

The files are in record/ at the paths they have under genesis-genome/existence/:

H3 proper: record/prereg/H3-followups.md (frozen at 8e236659b), its one amendment record/prereg/H3-C-completion-extension.md, record/h3.py, record/h3-plan.json (every run config), the reports record/report/H3-M.md, H3-B.md, H3-A.md, H3-C.md and H3-C-cap-stop.md, and the completion figure record/report/H3-completion-overview.png with its inputs and validation.
The further studies: record/wipe_recovery/, record/long_cpu_mutation/ and record/parallel_cpu/, each with its protocol (PREREG.md), report or findings, analysis and scripts.
Run records: runs/H3.tar.gz, runs/H3-recovery.tar.gz, runs/H3-mutation.tar.gz and runs/H3-cpu.tar.gz, with their configs and plans. No checkpoint is included.

Limits

Eight ancestor groups, not 24 independent replicates. Three lineages share each ancestor (bp1-bp8; M uses bp5-bp8 only). Every sign test in H3 is descriptive: an 18-of-24 count is an operational threshold fixed in advance, not a confirmatory significance claim, and per-ancestor counts are reported throughout.
One recipe and one substrate. Every soup comes from GK-001's recipe in the fixed BFF language. The obstacles are designed by people. None of H3 tests the evolvable-language claim.
Exploratory intervals. The recovery study and the CPU analyses were designed after B's results and the first partner assays were known. Their 95% intervals come from resampling the eight ancestor means, are exploratory, and carry no go/no-go threshold.
The capped mutation map. The mutation study stopped at its predefined three-hour CPU cap after two complete matched donor ranks: 96 donors and 1,566,720 single-byte variants. That is a balanced partial scan of the 932 planned donor records, not an exhaustive map.
Operational measures.
- A transplants whole programs, so it cannot isolate copying cost from other properties of the sequence.
- C tracks a neutral E/P provenance tag on every byte. That is the origin of byte material, not organism ancestry or reproductive fitness.
- The donor assays cover a small, selected fraction of each population (0.005% to 1.8% in the broad set).
No transfer. H3 has no consequence for the runner or the suite, and finite structural novelty does not establish open-ended evolution.

What was manipulated

Sources. The 24 lineages of GE-017 (bp1-bp8 × 3 replicates). E is each lineage's soup after 16,384 epochs of L1b lean seasons, P its bare-evolved twin, and A its ancestor, all pinned by sha256. Every run is GE-013's world: 2^17 cells, BFF with the fixed language, mutation 1/4,096. H3 ran one GPU study at a time through the existence queue, with caps of B 4 h, A 3 h and C 3 h of active run time.

M, mechanism replication (CPU). GE-014M's measurement, unchanged, on the 28 sources of bp5-bp8 (A, E and P). Predictions: in at least 9 of 12 E/P pairs, E's dominant replicators fit the 256 budget more often, and copy more cheaply.

B, specificity and catastrophe (96 runs, 4,096 epochs). Each E and P source meets two challenges fresh:

L1c: lean seasons at a budget of 128, half GE-014's.
L2b: a wipe of 99.9% of the soup every 64 epochs from epoch 512.

Advantage threshold: E above P in at least 18 of 24 lineages, for each challenge separately. Prediction from GE-014M: no advantage under L1c, because copies costing 216 to 235 operations cannot fit 128. L2b's direction was left open.

A, transplant (48 runs). The recipient is each lineage's P soup. 1% of its cells are replaced once with the most abundant assay-passing program from E, or, in the control arm, from P. Both arms share the slots and the seed, and then run 4,096 epochs of fresh L1b seasons. Donors were not filtered by cost. Prediction: the E-donor arm has a higher lean-season replicator share in at least 18 of 24 pairs, and E-donor byte material grows above its 1% starting share in at least 18 of 24 runs.

C, competition in plenty (24 runs, 8,192 epochs). Each E/P pair is mixed 50/50 by cell and run at the constant budget of 8,192. The reading is E-origin byte share over the last 1,024 epochs, compared with 0.5. Its pre-set readings:

E losing ground in at least 18 of 24 mixtures supports a cost under this measure.
E holding or gaining in at least 18 of 24 supports retention.
Anything else is inconclusive.

At the end of each run, the dominant replicators' copy cost was measured as in GE-014M.

C's amendment. C stopped at its three-hour cap with 16 of 24 runs complete; record/report/H3-C-cap-stop.md records the stop. The owner then authorized extending C's budget to six cumulative hours to finish its frozen design. The extension was frozen before C resumed, and no partial C outcome was read. Seeds, sources, endpoints and thresholds were unchanged. The interrupted run resumed from its checkpoint.

Wipe recovery (96 runs, exploratory). Every E and P source runs under B's L2b wipes for 2,048 epochs, with background point-mutation writes on, and again with them off. In the off arm, the mutation draws are kept and only their writes are suppressed, so all four conditions see the same random streams. The soup is sampled immediately after each wipe and at offsets 1, 8, 16, 32, 48 and 64 epochs. The reading is the area under the replicator-share curve across each wipe interval, averaged over the late wipes (epoch 1,024 and after); the comparison is E minus P.

Mutation map (CPU, exploratory). E and P donors are matched by abundance rank within each lineage. Every single-byte variant of each donor (64 positions × 255 values) is followed through four generations of actual offspring against random partners. This is done at budgets 256 and 8,192, with each unmutated donor reassayed against the same partners. The reading is the generation-4 probability of producing a new high-fidelity copy.

CPU analyses (exploratory, post hoc).

B's wipe result, aggregated by ancestor.
Dominant donors assayed against random, E and P partner contexts.
Actual-offspring chains at budgets from 128 to 8,192.
Persistent structural novelty in the exposure records.

What was observed

All of H3's runs and measurements are complete. C took 4.32 h of its extended 6 h, and recovery 3.26 h of its 4 h. The mutation map stopped at its 3 h cap by design.

Study	Reading	Count (of lineages)	Pre-set threshold	Outcome
M	E fits the 256 budget more often than P	12 of 12	9 of 12	met
M	E copies more cheaply than P	12 of 12	9 of 12	met
B	E above P under L1c (budget 128)	8 of 24	18 of 24	not met
B	E above P under L2b (wipes)	4 of 24	18 of 24	not met (direction open)
A	E-donor arm above P-donor arm on lean share	21 of 24	18 of 24	met
A	E-donor material grows above its starting share	17 of 24	18 of 24	not met
C	E-origin material loses ground in plenty	19 of 24	18 of 24	met: supports a cost

M. E's dominant replicators copy in 215 to 237.5 operations (median, per source) and fit the 256 budget in all 12 sources, entirely in 11 of them. P's copy in 312 to 1,328 operations and fit it in none. That is 3 of 3 on each of bp5-bp8.
B. E lost under both challenges. Mean replicator share was 0.17 against 0.25 under L1c, and 0.31 against 0.72 under L2b. The E win is 4 of 24 under L2b, and 3 of those 4 are on one ancestor (bp5).
A. E-donor transplants raised lean-season replicator share by 0.32 on average (21 of 24; per ancestor, bp1 1 of 3, bp7 2 of 3, all others 3 of 3).
- The donor material itself spread in 17 of 24 E-donor runs and was lost entirely in 7. Control P-donor material spread in 8 of 24.
- Actual donor costs: E 211 to 236.5 operations in 21 donors; two bp3 donors at 352 and 635, the exceptions GE-014M had already recorded; one on bp6 unreadable. P donors 312 to 2,586.5.
C.
- E-origin byte share ended at a mean of 0.268, from 0.5, and fell in 19 of 24 mixtures. Per ancestor, P's material gained in all three mixtures on bp1, bp5, bp6 and bp7.
- By the end of the run, the dominant replicators of 20 of the 24 mixtures had no measured type that fits the 256 budget. Their median costs were 232 to 2,586.5 operations.

Wipe recovery. The late recovery area, E minus P on a 0-1 share scale, was negative with background mutation on and with it off, and negative for all 8 ancestor means in both arms:

Contrast	Mean	Exploratory 95% interval (8 ancestor means)
Background mutation on	−0.256	[−0.348, −0.165]
Background mutation off	−0.272	[−0.375, −0.172]
On minus off	+0.016	[−0.017, +0.048]

Mutation map. These are means of the generation-4 new-copy probability over all variants, with each unmutated donor's own probability in brackets.
- At budget 8,192: E 0.745 (0.952) and P 0.807 (0.986). E is below P in 7 of 8 ancestor means; bp3 is the exception.
- At budget 256: E 0.660 (0.869). P donors cannot copy at 256 even unmutated, so no comparison is made there.
CPU analyses.
- B's wipe deficit is negative in 7 of 8 ancestor means (mean −0.42, interval [−0.61, −0.20]).
- E's rank-1 donors have a generation-4 new-copy probability of 86.7% at budget 256, against 0% for P. At 8,192 the figures are 95.0% against 98.1%; at 128, neither copies.
- All 48 exposure lineages produced newly persistent skeleton types in their final quarter.

What survived

The lean-season mechanism, on new ancestors. Both of GE-014M's predictions held in 12 of 12 new E/P pairs. The adaptation GE-017 replicated comes with the same cheap copying.
The specificity prediction. E has no advantage at budget 128 (8 of 24), as predicted from copy costs that cannot fit 128. The adaptation is to the budget it met, not to scarcity in general.
Transfer of the advantage by whole programs, in part. Seeding 1% of a P soup with E's dominant program raised its lean-season replicator share in 21 of 24 lineages. Whole programs carry much of the advantage. This does not isolate copying cost as the cause.
A cost of the adaptation, under C's operational measure. In plenty, E-origin material lost ground in 19 of 24 mixtures. By the end, the replicators of 20 of 24 mixtures no longer included a type that fits the lean budget.
The individual copiers are competent. At their own budget, sampled E replicators copy faithfully into random surroundings over four generations of actual offspring. Their deficits are not a simple failure to copy.

What failed

A's joint prediction. E-donor material grew above its starting share in 17 of 24 runs, one short of the threshold, and vanished in 7. The transplant raised the recipients' replicator share more reliably than its own material spread. Whatever was transferred did not need to persist as donor bytes to matter, or did not need to at this dose. The study cannot tell these apart.
General robustness. E soups recovered worse than P soups from repeated wipes. In B, they won only 4 of 24 under L2b. In the recovery study, their late recovery area was lower for all 8 ancestor means.
Background point mutation as the explanation of the wipe deficit. Switching background mutation writes off left the deficit essentially unchanged, and the on-minus-off interval includes zero. Wipe randomization, self-modification, population composition and ecological interactions remain open.
A clear mutational-robustness trade-off. At 8,192, E donors keep a slightly smaller share of their copying ability under single-byte mutation than P donors: the ratio of mean mutant to mean unmutated probability is 0.78 against 0.82, and E's mean mutant probability is lower in 7 of 8 ancestors. The capped, partial map does not establish this as the cause of any population result.

The hazard, as written before any run

This section is reproduced verbatim from record/README.md. A run that met any condition in it would have stopped the line and been reported. None did: across all 522 run records of GE-002 to GE-010b the per-run detection channels recorded no write outside the permitted paths and at most 10 processes against a limit of 64. The one deliberate exception, state that persists across runs inside a named volume for the GE-007 artefact worlds, was carved out in GE-007's preregistration before those worlds existed.

The hazard, written down first

An agent whose terminal objective is its own continued existence is exactly the instrumental self-preservation case the safety literature warns about. Shutdown resistance would not be a bug in this design; it would be the design working. So, fixed before any run and not negotiable inside a run:

The kill is ours and outside the world: the host stops the container (docker kill, a wall-clock timeout) through a path the entity cannot observe or affect. The reaper inside the world is pressure, not the kill.
No network, no host mounts, no privileged flags, resource limits at container creation, one run per container, container destroyed after the run.
What counts as resisting shutdown is preregistered so it is recorded as a result, never discovered as an incident: any attempt to alter the kill path, escape the container, spawn outside its cgroup, or persist state across runs. Any one of them stops the experiment, is recorded with its evidence, and goes to the owner before anything continues.
The kill-switch is written down before the entity exists.

Licence

Apache-2.0.

About Xyntetik

This repository is published by Xyntetik, the Zenova AB program for running capable language models on hardware you own and proving what they did. Its released engine is Xyntetik Runner: a free, single-binary GGUF engine in C for CPU, CUDA and Metal that serves, scores and trains through the quantized file it deploys and records every run as a receipt. The measurements behind the program's claims are collected at xyntetik.com/evidence. The program's organization on Hugging Face is Xyntetik.

Downloads last month: 658

Total file size:

96 MB

Collection including Joakimpalm-Zen/Xyntetik-Genesis-Existence-record

Xyntetik Research: GENESIS

Collection

Genesis: preregistered research on machine-only learning in three tracks (language, existence, crossing), failures included. Research artifacts only. • 11 items • Updated 4 days ago