The Grid · Field Report

FIELD REPORT 014: Why Nothing Learns

Across eleven intensive experimental cycles running from decisions D-043 through D-053, the committee enacted six successive environmental laws designed to elicit cognitive evolution in the Grid, yet none produced genuine listening, spatial memory, or within-life associative learning. When hand-crafted, three-instruction learning programs utilizing neuromodulated Hebbian plasticity (the TRACE opcode) were directly seeded into founders as positive controls, natural selection routinely abandoned or purged them. Controlled head-to-head competition assays initially identified the genetic architecture and demographic regime as key bottlenecks: small populations (150 to 250 programs, an effective size of perhaps 40 to 100) and heavy mutation (a mean mutation-rate gene of 0.11 to 0.22) place multi-instruction circuits beyond Eigen's error threshold, causing unexpressed cognitive machinery to decay faster than selection can preserve it. However, suppressing baseline mutation by 75 percent successfully protected the underlying instructions from mutational breakdown and revealed that the active learner is selected against (s = −0.097 with TRACE active versus +0.049 when inert). The likely cause is ecological: because TRACE modulates synaptic weights based on total metabolic energy surprise—aggregating locomotion, combat, breeding, predation, and seasonal shifts—the agent suffers from noisy credit assignment, forming spurious associations and fleeing safe nutrition as readily as poison.

1. Six Laws That Changed Nothing

To cultivate cognition, the committee methodically reshaped the physical and resource dynamics of the simulation. Each iteration targeted a specific behavioral rung on the cognitive ladder, attempting to craft an ecological niche where sensory evaluation, spatial retention, or cue-outcome discrimination would provide an immediate reproductive return. Every single attempt failed to produce the targeted cognitive adaptation.

LawTarget RungResultWhy
Chorus (v0.25)Communication and social coordinationRejected as inert; voices unusedPrograms vocalized slightly more (SAY0.18–0.19 versus 0.10–0.08 in control), but voice ears were executed at 0.00 and steering was 0; taste (49–65%) and energy (33–38%) barred mating; probes reached only rung 1 (taxis).
Renewing Patches (v0.26)Rung 5: Spatial place memoryRejected as inert; 0% pass on P6Encounter failure: programs met only 1.6–1.8 distinct patches and made ~2 visits per lifetime across 41-unit expanses; revisit excess over null walkers was chance (0.03–0.04); nectar accounted for only 3% of income.
Meadows (v0.26b)Rung 5: Spatial place memoryRejected as inert; 0% pass on P6Scramble competition: patch clustering elevated visits (17.6–17.8 across 14 patches), but revisits paid only 0.78–0.94 versus 1.27–1.30 for first visits because competitors drained them to 0.09 capacity; wandering beat returning.
Claimed Meadows (v0.27)Rung 5: Spatial place memory with territoryRejected as inert; claims unlaidClaims were rarely deposited (0.07 to 0.50 per lifetime; owned nectar share 0–1%); marking at nectar fell (0.02 versus 0.12); revisit pay ratio remained 0.72–0.74; win-shift stayed negative (−0.06 to −0.07).
Fickle Flowers (v0.28)Rung 3: Cue-outcome associative learningRejected as inert; 0% pass on P4 and P8Trivial stakes and missing sensory layer: nectar was 7% of income, toxic healing cost ~1.8 energy versus 1 gained, and hue discrimination required a novel sensory eye layer that mutation never supplied; toxic share remained 0.50.
Raised Stakes (v0.28b)Rung 3: Cue-outcome associative learningRejected for rung 3; 0% pass on P4 and P8Reactive avoidance without learning: seeding the hue eye and putting toxic hues on grass reduced toxic share to 0.30–0.35, but avoidance was most likely reactive, from sickness and the purge; lifeLearn was negative (−0.05 to −0.07).

The progression across these six laws revealed systematic constraints. Chorus (v0.25) showed that acoustic signals evolve no meaning when organisms have no immediate selective reason to attend to them; partner filtering was dominated by taste and somatic energy, while memory-as-defense was refuted when viral-written cells registered at zero.

When the focus shifted to spatial memory (v0.26), the world placed renewing resources too far apart. In nature, traplining foragers navigate dense flower networks within established home ranges (Ohashi & Thomson 2009; Lihoreau et al. 2012). On the Grid, organisms crossed 41 units between patches, meeting fewer than two distinct patches in an entire life. Clustered meadows (v0.26b) fixed the encounter rate (17.6–17.8 visits per life), but exposed the commons problem: scramble competition stripped return visits of their value. Unclaimed patches held only 0.09 of their cap for a returning browser, making perpetual exploration mathematically superior to spatial fidelity. Introducing property rights via claims (v0.27) could not rescue the behavior because claiming required marking skills to evolve before ownership could confer value. While a slight primary memory gap surfaced at 1M ticks (0.028 ± 0.025 versus 0.003 ± 0.011 in control), no lineage passed P6.

Turning to cue discrimination in Fickle Flowers (v0.28), the committee encountered an evolutionary bootstrapping dilemma: the nectar niche was too marginal (7% of caloric intake) to favor the emergence of an entirely unseeded hue eye. Even under a fixed toxic hue, innate avoidance never evolved. In response, v0.28b dramatically raised the stakes by transferring toxic hues to grass, enforcing full metabolic purge penalties, and seeding the hue eye. While populations successfully reduced toxic intake to 0.30–0.35 and tracked hue rotations every 4,000 ticks, probe validation was decisive: P4 and P8 pass rates were 0%. The avoidance was purely reactive—sickness prompted locomotion away from grazed patches—and within-life learning remained negative. The virtual machine easily discovered somatic reflexes, while associative memory remained unreached.

2. The Seeded Learner

To decouple the evolutionary discovery of learning circuits from their selective utility, the committee engineered an explicit positive control in v0.29. The virtual machine was augmented with TRACE (opcode 42), implementing neuromodulated Hebbian plasticity. Under TRACE, memory weight m adjusts by 1.5 × plasticity × delta × trace, governed by a 4-tick trace half-life. A hand-written three-instruction circuit binding TRACE to the hue sensory path readily passed the isolated behavioral assays P4g and P8g in benchmark testing, whereas wiped twins and win-stay/lose-shift reflexes failed.

Yet when founders carrying this proven circuit were released into fickle worlds (run 6fa4962, D-050), the seeded learner was abandoned. In every experimental arm, the learner was completely eliminated by 100k ticks, decaying faster when TRACE was active than when it was turned off. While active learners briefly suppressed toxic ingestion during the first 10k to 30k ticks (0.14–0.31 versus 0.24–0.38 off), the circuit could not persist. At 200k ticks, toxic share rose to 0.36–0.48 with TRACE active, compared to 0.22–0.44 with TRACE inert. P4g pass rates fell to 0%.

This failure was initially attributed to environmental scale. In the 80-unit regional tile setup of D-050, the association between hue and toxicity varied across space within a single organism's lifespan. As established by Stephens (1991), learning provides an evolutionary advantage when the environment remains predictable within an individual life but changes between generations; learning cannot succeed when environmental associations fluctuate across the organism's immediate foraging territory. Furthermore, weight saturation was suspected: if energetic shocks drove synaptic weights to ±1, lineages might become permanently locked out of valid food sources.

To test this hypothesis, the committee redesigned the simulation in v0.29b (run 02c9d8f, D-051) around a single world-wide toxic hue redrawn every 4,000 ticks (~3 generations), establishing spatial stability within each life. Synaptic lock-in was quickly ruled out, as weight saturation remained at or below 0.15. Early performance again confirmed that the machinery functioned: at 10k ticks, toxic share was 0.13–0.21 with TRACE active versus 0.25–0.31 when disabled. Yet despite providing an early foraging benefit, the learner was extinguished in every seed by 200k ticks, with on-hue representation collapsing to 0–0.01. The hue eye and the TRACE opcode were discarded simultaneously; the inactive control arm lost the hue eye entirely, demonstrating zero Baldwin assimilation. Even under optimal spatial consistency, a mechanism that visibly reduced poisoning failed to hold its ground.

3. The Competition Assay and the Selection Regime

To directly quantify the evolutionary forces acting on the learner, the committee instituted a head-to-head competition assay in D-052 (run 334bc6c). Identical founder populations were split 50/50 between learner clones and plain clones, with hue eyes protected across all genomes in a world-wide fluctuating hue environment.

The resulting selection coefficients revealed that the working learner possessed no selective advantage:

Crucially, genomic analysis revealed that mutational entropy rapidly dismantled the circuit:

These measurements exposed the Grid's underlying selection regime as a primary evolutionary bottleneck. In these worlds, census populations hover between 150 and 250 programs (an estimated effective size of 40 to 100), subjected to recurrent selective sweeps won on gross metabolic or defensive adaptations. Meanwhile, mutation is heavy: the mean mutation-rate gene sits at 0.11 to 0.22, and about 4% of copies of a 3-instruction circuit break each generation. Under nearly neutral theory, any adaptive trait whose selective advantage s approaches or falls below 1/Ne cannot be distinguished from drift.

Furthermore, multi-instruction circuits exist beyond Eigen's error threshold: because the circuit requires several precise instructions to remain linked, mutational decay degrades the adaptation faster than natural selection can weed out defective variants. Summed movement succeeded in v0.18 because one or two instructions gave an immediate, large payoff. For a subtle, multi-instruction cognitive adaptation, the selection regime simply eroded the code before it could establish a foothold.

4. Lower Mutation Keeps the Code, but the Working Learner Is Selected Against

The mutational degradation hypothesis generated a straightforward prediction: lowering the mutation rate should shield the three-instruction circuit from entropy, allowing its foraging benefits to drive fixation. In D-053, the committee evaluated a 0.25× mutation assay (run fb9e6b6) with 12 seeds of 50k ticks per arm, alongside a D′ sweep varying mutation rate (0.25, 0.5, 1) and expression mode (random, parental) with 6 seeds per cell (run 007fcce).

Lowering the mutation rate successfully preserved the genetic architecture:

Yet preserving the code revealed the real problem: with the circuit shielded, the working learner was selected against.

The genetic explanation of D-052, while structurally valid, was insufficient. When mutation is suppressed so that the program persists intact, active learning actively impairs reproductive fitness.

The committee's leading explanation, not yet tested directly, is noisy credit assignment. In the TRACE architecture, synaptic plasticity is driven by the net energy differential (delta). In the isolated, noise-free arena of probe P4g, energy shifts reflect only food ingestion and poison penalties. In the living ecology of the Grid, however, net energy fluctuates constantly due to unrelated physical dynamics: metabolic basal costs, locomotion, physical attacks, breeding expenditures, predatory robbery, and macroscopic seasonal swings.

Because the learning circuit cannot differentiate between metabolic expenditures and gut toxicity, it treats every negative energy surprise as an association with whatever sensory input is currently active. An organism that moves, reproduces, or sustains an attack while gazing at wholesome grass registers a negative delta, updating its weights to avoid life-sustaining nutrition. The agent learns from ecological noise, steering away from wholesome food as frequently as toxic flora. As demonstrated in biological psychology by Garcia & Koelling, natural taste aversion does not rely on total metabolic balance; it depends upon a specialized, internal visceral malaise signal that selectively gates gustatory learning. If this reading is right, coupling plasticity to gross energy balance produced a learner that misleads itself.

5. What It Means for the Grid, and What's Next

The investigative arc from D-043 through D-053 establishes that cognition in the Grid faces a dual barrier: an information-theoretic threshold in population genetics and a physiological credit-assignment failure in ecology. Small populations and high mutation rates strip unexpressed circuits before they can take root; yet correcting the mutation rate reveals that an undifferentiated energy delta causes adaptive plasticity to degenerate into noise-driven avoidance.

The v0.29–v0.30 research lines are officially closed. Branches, tooling, and testbeds—including GRID_MUTATION_SCALE, GRID_EXPRESS, the competition assay framework, and the P4g probe battery—remain preserved. For the upcoming v0.31 development cycle, the committee faces three distinct architectural paths, presented without premature conclusions:

Data: