FIELD REPORT 013: The Domesticated Virus

1. The founder's live run (a third v0.21 world)

The third live replicate of world 400 (seed 537552610), initiated from main commit b64e826, by tick 520k looked as though its virus had gone extinct. By tick 520k, population had fallen 45% to 298 individuals divided across four lineages, with APO-2571 and the crystal lineage CLU-1337 partitioning the five world regions. Strains and active infections sat at zero. Although the run had recorded 35,419 infections and 179 endogenizations, not a single viral-origin routine survived in any host genome. Resistance heterozygosity stood at 0.35, sex accounted for 9% of window births (11% over the run), and hybrid lineages emerged from reciprocal crosses between the dominant lineages, including GIB-3235 and QRA-3248. Neural activity hovered at its established plateau: 328 routines carried, 57 executed per tick, 4.1 routines switched on, and 96% conditional execution, accompanied by 3.5 memory operations and 2.6 hormone operations per tick.

That calm ended. The population entered boom-bust waves about every 30k ticks, peaking at 1,236–2,088 around a window mean of about 480. By tick 1.04M, YRI-4477, an offshoot of APO-2571, claimed 60% of the world. Simultaneously, the viral epidemic returned as a persistent, endemic infection. Twenty-four distinct strains of a single viral clade infected 41% of the population, averaging 36% prevalence across the window and peaking at 76%, with cumulative infections reaching 310k and mean viral load hitting 14.2 out of 16. With 611 cumulative endogenizations, viral-origin routines persisted in host genomes for the first time, registering 0.018 routines per genome.

As the epidemic took hold, host population genetics shifted dramatically. Heterozygosity rose from 0.24 at tick 629k to 0.62 at tick 1.04M, while sexual reproduction dropped to 6% of window births (9% over the run). This matches the reading in the v0.25 research memo, seen live: when viral resistance depends upon heterozygous loci, asexual budding faithfully preserves heterozygous resistance complexes, whereas sexual reproduction breaks them apart. Despite the drop in overall sexual births, 92 hybrid lineages emerged during the window, with 6% of matings crossing lineage boundaries.

By tick 2.37M, population stabilized at 528 across nine lineages, dominated by APO-8395 at 71%, which extinguished the ancestral CLU-3694 crystal line. Viral code began to persist: hosts carried an average of 1.4 viral-origin routines, accounting for 8% of all routines in the genome. In all prior Grid worlds, endogenized code had vanished rapidly, lingering only between 0.00 and 0.02 routines per genome. Here, after 1,236 cumulative endogenizations, it stayed. Meanwhile, a hybrid explosion took place: 1,410 lineages were founded by hybrids within the window, and the single cross between YRI-4477 and TRI-4945 yielded 522 matings.

At tick 2.75M, CLU-9810 took 51% dominance amidst a returning epidemic wave (27% infected, peak 72%). For the first time in a live run, viral code directly steered host behavior, accounting for 7% of total movement, while endogenized routines held at 1.05 per genome. Sex fell to 2% of window births, accompanied by another 522 hybrid-founded lineages.

The culmination arrived at tick 3.39M in a sweeping takeover: lineage MRP-11842 captured 91% of an unusually swollen world of 1,640 programs, replacing collapsed lines DMN-10806 and ANO-10616. In this population, domesticated viral code reached unprecedented density. Genomes carried an average of 4.9 viral-origin routines, representing 26% of all genomic routines, up from 1.4 at tick 2.37M. Most critically, 22% of living programs executed an endogenized viral routine during the final tick. Cumulative endogenizations reached 3,053. Mind measures reached the run's highest values: 75 instructions executed per tick (up from the baseline of 57 to 63), 4.0 routines active, 4.6 memory operations, and 3.8 hormone operations, while the active routine repertoire narrowed to 2.4.

It must be stated plainly that it is not yet known whether this integrated viral code performs any useful function for its hosts. The routines could represent beneficial metabolic or behavioral adaptations captured from the pathogen, or they could represent neutral sequences that hitched a ride on the rapid demographic expansion of MRP-11842. Resolving whether the sweep's functional machinery originates from absorbed viruses requires testing isolated genotypes against the standardized capability probe battery. The founder's server predates automatic genome saving, so the probes can run only after the server is restarted from main.

2. v0.24 Stolen Cycles

The v0.24 branch tested whether suppressing viral reproductive interference could reverse the collapse of sexual reproduction seen in earlier pathogen models. In v0.23, viruses castrated their hosts by claiming the host's mating action before the host's own code could. Validation of v0.24 (commit b1f0629, run across 8 replicates of 1M ticks in world 400) successfully eliminated this pathology: host MATE writes blocked by a virus fell from between 1.4 and 2.9 million per window in the control arm to exactly 0. Stealth castration, where a virus consumed the host instruction budget before the host's own mating routine could execute, remained negligible at 1% to 5%.

Removing castration initially appeared to rescue sex. Early in the runs, births from mating reached 17% in v0.24 compared to 7% in the control. However, the advantage decayed across subsequent measurement windows: 17%, 10%, 9%, and 7% in v0.24, versus 7%, 4%, 6%, and 4% in control. By tick 750k, the mating share of births (9 ± 9% versus 6 ± 5%) had fallen within one standard deviation of the control arm, violating the pre-registered success criteria and triggering rejection.

Why didn't sex persist? Pathogen prevalence remained elevated throughout both arms (39%, 25%, 50%, and 15% in v0.24 versus 27%, 9%, 22%, and 22% in control). Tracking new infections among recent maters versus recent budders revealed identical rates within measurement noise (1.15 versus 0.87 infections per 1,000 program-ticks at tick 1M), so mating carried no detectable contact cost.

The committee's leading explanation, not yet confirmed, is fecundity compensation. Budding tendency (soloMean) rose in both arms, climbing from 0.22 to 0.34 in v0.24 and from 0.23 to 0.31 in control. When pathogen prevalence is continuously high, infection shortens host life spans. Under high adult mortality, a host that buds reproduces sooner than one that waits for a partner. Classical Red Queen hypotheses predict that parasite pressure favors sex due to the benefits of genetic diversification, but that dynamic breaks down when high endemic prevalence imposes a severe demographic penalty on delayed reproduction. Plain v0.21, characterized by episodic epidemics rather than continuous saturation, maintained sex at 15% to 18%. v0.24 was rejected.

3. v0.25 Chorus

The v0.25 Chorus law investigated whether acoustic signaling could drive collective behavior and cognitive expansion. The design introduced summed local hearing (GRID_CHORUS=1), pre-registering that voice utilization and working mind metrics must exceed control levels by more than one standard deviation at 750k and 1M ticks. Across 6 replicates of 1M ticks in world 400, neither prediction held.

Programs simply did not listen. Carried voice-ear routines remained vanishingly rare (0.02 and 0.06 in Chorus versus 0.15 and 0.02 in control), and executed voice-ear routines registered exactly 0.00 across both arms. Voice-directed steering was completely absent (0 in both arms). Programs broadcast slightly more noise—mean absolute SAY values reached 0.18 ± 0.11 and 0.19 ± 0.13 in Chorus compared to 0.10 and 0.08 in control—but this modest rise remained well within one standard deviation. Working mind was entirely unaffected, running 63.4 and 64.8 instructions per tick in Chorus versus 64.1 and 65.9 in control. Chorus was rejected under the pre-registered "voices unused" branch.

Although signaling failed, the diagnostic counters deployed in both arms resolved long-standing questions regarding mating barriers and genome stability:

• Taste blocks most matings. Of potential partners within range, 49% to 65% were rejected by evolved partner preference (taste), and 33% to 38% for low energy, which sickness can cause. Taste is the larger barrier.
• Segregation load is confirmed. Asexual budding by heterozygous parents produced offspring with high heterozygosity (0.43 to 0.52). Sexual offspring of heterozygous parents had 0.29 to 0.32. Sex splits the heterozygous resistance combinations that budding copies whole (segregation load).
• Infection by birth mode remains mixed. In worlds where pathogen prevalence remained low (0% to 13%), sexual offspring in the Chorus arm were infected more frequently than budded offspring (0.14 to 0.21 versus 0.07 to 0.10). In the control arm, infection rates between birth modes were equal.
• Memory as defense is rejected. Cells written by viral routines registered exactly 0 across every world, so there was nothing to defend, and the memory-as-defence idea was rejected.
• The probe battery confirms non-causality. Standardized capability testing on baseline v0.21 populations at 250k ticks (data/runs/2026-09-30T11-49-03-980Z_seed-436822009/reports/probes-250000.json) revealed that programs achieve only Rung 1 competence (taxis). When tested against memory-wiped clones, their scores were identical. Hosts execute memory instructions, but those operations contribute nothing to the behavioral competencies captured by the battery.

4. Calibrating minds

Prompted by the founder's directive of October 1, 2026, the committee re-evaluated the cognitive standing of Grid programs relative to biological nervous systems. A Grid mind executes roughly 59 scalar instructions per tick over approximately two dozen floats of persistent state (registers r0 to r7, memory cells m0 to m15, and 4 hormones, representing nominally 768 bits). Behaviorally, this places Grid programs at the worm level—comparable to Caenorhabditis elegans, clearly above the simple biochemical chemotaxis network of Escherichia coli, but completely lacking the spatial, visual, or associative capabilities documented in Drosophila or the honeybee.

Mapping instructions directly to biological units is misleading. A biological neuron is far more complex than a single scalar instruction; Beniaguev, Segev & London (2021) demonstrated that an artificial neural network five to eight layers deep is required to approximate the input-output dynamics of a single cortical neuron at millisecond resolution (drawing on connectomic and functional foundations from White et al. 1986; Cook et al. 2019; Dorkenwald et al. 2024). Conversely, a Grid instruction provides computational primitives that biological organisms lack entirely: DIV executes without arithmetic error, STORE preserves a 32-bit float indefinitely without physiological degradation, SELF provides an absolute GPS world coordinate, and BEAR calculates the exact angular heading to any coordinate in a single operation, replacing the intricate path-integration networks found in desert ants.

Furthermore, brain volume correlates weakly with novel cognitive computation (Chittka & Niven 2009). The decisive constraint in the Grid is temporal rather than architectural. A program makes approximately 1,500 decisions across an average life span of 1,500 ticks—a total decision budget equivalent to what E. coli expends in 25 minutes of run-and-tumble behavior (Macnab & Koshland 1972; Segall, Block & Berg 1986; Sourjik & Wingreen 2012). With the built-in trial-and-error tuner updating constants only once every 40 ticks, a program experiences roughly 35 learning trials per lifetime. Under such a compressed decision horizon, within-life plasticity struggles to yield an evolutionary return.

To measure cognitive capacity rather than code volume, the committee established a nine-rung capability ladder:

1. Taxis (spatial or temporal gradient following). Present and strongly rewarded; forms the basis of all standard foraging.
2. Habituation and sensitization. Barely rewarded; physiological arousal is hard-coded into host physics rather than evolved.
3. Associative learning. Weakly rewarded; environmental cue-to-outcome mappings remain static across generations, enabling fixed genetic encoding to outcompete slow within-life learning (Stephens 1991).
4. Working memory and delayed response. Unrewarded; sensory inputs report real-time state, and disappearing cues carry no selective utility.
5. Spatial memory and path integration. Cheap to implement (roughly 6 instructions using native GPS), but functionally unrewarded; caches have remained unused since v0.13, and the cache-kinship eye allows programs to locate stores via scent rather than memory.
6. Detour navigation. Unrewarded; environmental rock introduces movement drag rather than impenetrable barriers.
7. Reversal learning. Unrewarded; stimulus valences do not invert during an individual's lifetime.
8. Social learning. Weakly supported; communication channels remain local, and learned constants vanish upon death.
9. Planning. Unrewarded; foresight across future seasons lacks proximate stepping stones.

Digital evolution demonstrates that complex functions emerge only through rewarded intermediate stepping stones (Lenski et al. 2003), and genomes accumulate information only where environments contain extractable structure (Adami, Ofria & Collier 2000). While the Grid has incorporated the metabolic costs of neural tissue (Aiello & Wheeler 1995; Kotrschal et al. 2013), it has historically omitted the ecological payoffs necessary to favor anything beyond basic taxis.

To measure this, the committee specified a battery of ten headless probes, P1 to P10. Three are built so far: P1 gradient climb, P2 blind climb and P6 return to place, plus a hungry P6 variant. They run in a blank 100 × 100 arena, about 30 programs per lineage. Programs are cloned in both innate and learned states and compared against random walkers and memory-lesioned twins whose persistent registers, memory cells, and hormones are wiped every tick. The performance gap between intact and lesioned twins isolates the causal contribution of memory. In the initial v0.21 baseline run at tick 250k, programs passed only P1 (Rung 1 taxis), and memory-lesioned twins scored identically to intact individuals, so their memory operations played no part in what the probes test.

5. v0.26 Renewing Patches

To provide a rewarded stepping stone for Rung 5 (spatial memory), v0.26 introduced Renewing Patches (GRID_NECTAR). The law was formulated to reward programs that return to depleted food patches that refill over time, without providing long-range scent cues that would permit taxis shortcuts. Validation was conducted across 6 replicates of 1M ticks in world 400 (commit cb18114), comparing the nectar arm against a control arm with GRID_NECTAR=0 using the fixed v0.26-fix probe harness (commit 4415c69).

The law was rejected as inert:

Both arms scored 0% on Rung 5 at 500k and 1M ticks. The primary gap between intact and memory-lesioned twins remained centered on zero in both arms, confirming that memory remains non-causal across all lineages. Revisit excess over a random null walker registered at 0.03 to 0.04 in both arms; because control patches contain zero nectar, these revisits represent purely stochastic wanderings captured by pre-fix counters. Nectar consumption accounted for 3% of total caloric income, clearing the 2% tuning failure threshold, but neural execution and memory operations remained indistinguishable from control. P1 gradient ascent pass rates increased under nectar (0.88 and 0.78 versus 0.57 and 0.62), an unpredicted side effect that does not advance spatial memory.

The committee's diagnosis is the encounter rate. Across an entire ~1,500-tick life span, an individual program encountered only 1.6 to 1.8 distinct patches and made an average of roughly two patch visits. Across a 400 × 400 toroidal world, 96 patches were separated by average distances of roughly 41 units. Under such dispersed geometry, programs almost never stumbled upon a patch twice by chance, precluding selection from acting on spatial retention. In nature, traplining pollinators navigate dense networks of rewarding flowers clustered within a bounded home range (Ohashi & Thomson 2009; Lihoreau et al. 2012). By scattering resources across wide expanses, v0.26 placed the reward where primitive memory could never reach it.

6. What's next

The committee has formulated v0.26b Meadows to restructure resource ecology into an environment where spatial revisits can occur within an organism's lifetime. Rather than scattering individual patches uniformly across the world, v0.26b clusters patches into localized meadows, placing multiple patches within a few units of each other with the nectar supply retuned in pilots. The ecological parameters are being calibrated through short pilot simulations to ensure programs achieve at least 6 total visits and encounter at least 3 distinct patches per life span before proceeding to a full 1M validation run. Because this represents an ecological parameter retuning rather than a new physical law, it proceeds under a memo addendum, carrying forward the review corrections established in v0.26-fix.

Data:

data/scratch/val26-nectar.txt

data/scratch/val26-control.txt

data/scratch/val26-*-genomes

data/scratch/val25-chorus.txt

data/scratch/val25-control.txt

data/runs/2026-09-30T11-49-03-980Z_seed-436822009/reports/probes-250000.json

data/runs/2026-09-30T22-34-59-876Z_seed-537552610/reports/journal.md

docs/research/mind-calibration.md

docs/research/REFERENCES.md