The Grid · Field Report

FIELD REPORT 015: The Living World

The Grid has fundamentally transformed from a static, rapidly flattening landscape into "The Living World" (v0.34). This report documents a critical period of evolution in our methods and our simulation. We uncovered the mechanical trap that prevented learning from evolving, radically altered our committee governance with equal three-way votes, and merged our most ambitious experimental tracks into a single, unified ecology. By introducing a conserved rock cycle to preserve spatial topology, pricing emergent actions strictly according to physics, and establishing rigorous stabilization metrics, we achieved the strongest open-endedness results in the project's history. Version 0.34 has successfully passed the 1-million-tick stability check without reseeds or diversity collapse, opening the door for deep-time local runs and a unified hosted deployment.

1. The trap that hid learning (Free to Leave, D-055–D-058)

For millions of ticks, the Grid presented a cliff that prevented learning from taking hold. The discovery of the "Tantalus trap"—the Grid's equivalent of Pavlovian-instrumental conflict—finally explained why. In a targeted assay at 0.25x mutation, we proved that programs could successfully learn an aversion to toxic food, switching off their drive to eat. However, the founders' hard-wired stopping rule dictated that a program's forward thrust was negated if it detected food beneath it. If a program learned not to eat toxic food but remained standing on it, it would simply freeze and starve.

We deployed the "Free to Leave" founder ecology in v0.31 to fix this. By rewiring the forage routine so that a program reads thrust from the eat drive itself, a program will now move on wherever it chooses not to eat. The results of the initial step-0 assay were stark. The rewired "eat-plus-go" learners achieved a positive selection coefficient of +0.086 against a -0.097 penalty for the old wiring. For the first time, learning was positively selected for in the Grid, and the toxic share of the field dropped to 0.21 compared to 0.36 in the null. Furthermore, the rescue removed the trap's cost without acting as an artificial movement bonus; eat-plus-go learners maintained baseline residence and travel costs.

Yet, our discovery validation at 1 million and 3 million ticks forced us to confront a hard truth: learning pays once assembled, but evolution does not discover it, nor can it maintain it from rare. The memory probe recorded a flat zero across all arms at 250k, 500k, and 1M ticks. In a maintenance arm where we artificially seeded 10 percent of founders with the learning circuit, the learner lineage eventually claimed 97 percent of the ancestry, but the circuit itself was gone from its carriers by 1M. By 3 million ticks, populations evolved alternative, non-learning mechanisms to drop the toxic field share down to 0.16. The learning-law line is closed for now; we have proved the world pays for learning, but the barrier of spontaneous discovery remains unbroken.

2. A new way of working: the founder's directives, equal three-way votes, wider screens, early kills, combining bets (D-059–D-061, D-069–D-071)

The failure to evolve learning prompted a massive overhaul dictated by the founder. We discarded solitary decision-making in favor of equal three-way votes among Claude, Gemini 3.1 Pro, and Codex. This immediately shifted our experimental posture. In the very first vote on ecology and topology, the committee ranked wide screening above deep, isolated dives. We tested a vast factorial matrix spanning pass openness, flora occupancy, and food handling costs, uncovering that complex topology barely mattered while food handling costs acted predictably as scarcity.

Simultaneously, a major compute optimization merged into the main branch doubled our throughput and halved peak memory usage to 0.98 GB per replicate. This performance leap came with a new methodology: early kills. Runs can now be halted automatically if they trigger reseeds, drop below a population floor or suffer a lasting diversity collapse, freeing their slot for the next seed.

Rather than pursuing parallel bets, the founder issued a decisive override: combine all major initiatives into a single v0.34 "Living World." We synthesized living geology, open actions, rich biomes, and a predatory food web into one unified environment. To manage this massive injection of complexity, the committee replaced rigid, menu-driven population cost rules with strict stability measures. To pass, a world now required zero extinctions, bounded boom-bust cycles, and a hard 5th-percentile population floor of 150. We stabilized first, validating the architecture before committing to the deep time of long-term runs.

3. Actions that invent themselves: effectors, the reach loophole, the physics-priced rescreen, and the 1M panel's decisive novelty with its stability miss (D-061–D-067, D-072)

To drive true open-endedness, we replaced hard-coded capabilities with the v0.33 "Effectors" law. Programs were given an evolvable vector over four conserved physical primitives: moving matter, transferring energy, applying force, and changing properties. This was aimed at three distinct targets and driven on four shared channels, allowing a vast combinatorics of emergent behavior.

Early screening revealed a thermodynamic loophole: programs quickly evolved a fast reach-ahead mouth that transferred energy across distance without paying a physical cost. Codex and Gemini both flagged it, and Codex later insisted the fix come from physics rather than from tuning against an evolved success. We mandated that reach be priced directly by physics—transporting mass from a distance now pays the engine's true cost of moving that mass, alongside a higher heat loss for remote transfer.

The physics-priced rescreen at 0.25x mutation was a triumph. Effectors produced action kinds that were entirely new mixes of primitives, successfully earning real income while persisting across lineages. This culminated in the 1M panel, which produced the strongest open-endedness result in the project. The free founder effector arm pushed MODES novelty to 39, compared to a mere 4.7 in the menu control. Working mind reached 62.6, soaring above the pure preset control of 50.7.

Crucially, higher-order kinds successfully evolved far beyond the basic primitives. We observed a drain-and-strike leech claw present in all six effector seeds, a reach-and-mark mouth, a dedicated marking mouth with up to 154 carriers, a quarrying eroder, and even a three-primitive kind combining energy transfer, force, and property changes.

However, this explosive novelty came at a cost. The 1M panel missed the strict stability floor of 150 in some seeds of every non-menu arm. Because the trough problem was tied to the 0.25x mutation rate rather than the side activity itself, effectors were pushed into the larger, richer biomes of v0.34 for stabilization rather than being tagged independently as v0.33.

4. The vanishing rock and other world truths: rock gone in 100–200k ticks, vents, seasons and habitat (D-065, D-066, D-068)

Our pursuit of 3D terrain exposed a fatal flaw in the simulation's physics: the designed topology was vanishing. Exhaustive depletion measurements on the main branch confirmed that rock, once laid at genesis, was entirely eaten away within 100k to 200k ticks. In one seed, the initial 3,899 units of rock fell to 13.3 percent by 150k ticks and 2.2 percent by 200k. For more than 90 percent of every run's history, the Grid has been a flat, rockless world. This confirmed why labyrinth density and pass openness had no measurable effect in previous screens; the ridges simply didn't exist long enough to matter.

In response, all three committee members agreed on a conserved rock cycle, as the founder had proposed. The simulation now features permanent, uneatable bedrock, ore-bearing hard rock, and soft sediment. Weathering reduces rock to sediment, while vent-driven uplift, fueled by geological energy, pushes new topology back into the world.

Habitat telemetry further verified program distribution across this landscape. Programs systematically crowd warm vents to harvest energy and cook crystal, with a selection index averaging about 16 to 17 (up to 25) in two test worlds. Conversely, standing food is heavily avoided in the metrics only because it is immediately grazed down wherever a program sits. In the test worlds programs did not avoid trees. On the live run, the founder noticed something more interesting: the ruling lineage, which steers on the energy gradient, sat on open ground in one season and packed into the forests when nuts were down, while a crystal-mining lineage held the open middle. That reversal weakened an earlier idea, that the ruler was keeping its own ground treeless by browsing saplings (a grazing lawn); it now looks seasonal and food-led.

5. The Living World (v0.34): what was combined, how it was stabilized, and the 1M result, with its two watch items (D-071–D-077)

Version 0.34 represents the culmination of our combined bets. The "Living World" merged living geology, effector-driven open actions, rich biomes with dispersed flora, and a dynamic food web of grazers, browsers, and stalkers. Stabilizing this intricate ecology required 22 distinct tuning steps.

During the 300k pilot phase, the physical world proved incredibly stable—populations stayed well above the floor, and rock renewed smoothly at 1.0 to 1.3 times genesis levels. However, stalkers consistently died out once their dormant egg banks ran dry, managing only about 1 kill per 1,000 ticks. We resolved this by introducing a rare stalker immigration mechanic, mirroring the existing seed rain for forests. We formally redefined predator persistence as an ongoing ecological outcome of the specific world's density and cover, rather than a rigid stability gate.

Following a Codex review that cleared determinism and rock mass defects while correcting minor telemetry bugs, v0.34 entered the critical 1M stability check. It passed. Across all 4 seeds, there were no reseeds and zero early kills. The 5th-percentile population floored at a highly resilient 431 to 585, far above the 150 limit. The rock cycle stabilized beautifully, sustaining rock mass between 1.15 and 1.20 times genesis from 250k straight through to 1M. Within-population diversity flourished, maintaining between 214 and 526 effective genotypes and supporting up to 27 distinct lineages simultaneously.

Two critical watch items emerged from the 1M check. First, the food web remains incomplete in deep time; browsers die out entirely, and stalkers ultimately hit zero despite immigration. Second, open-ended action novelty decays to near zero in the second half of the run, and working mind stabilized slightly lower than in isolated tests, settling between 31 and 51. Both items will be monitored closely during extended validation.

6. What's next: long runs, the hosted world as the live run, and rock-cycle tuning to keep open space (in progress)

With v0.34 merged, the project has transitioned to deep-time validation. We have initiated local long runs across 6 seeds targeting 5 million ticks, keeping early kills active solely for catastrophic floor collapse or reseeds.

Simultaneously, we are tagging v0.34 and deploying it to the hosted world alongside a crucial UI memory fix. Moving forward, the hosted world will serve as the definitive live run. Our immediate engineering focus is the rock cycle. It now overshoots: on the hosted world rock grew to about 13 percent above its starting stock within hours and began filling in open ground the founder wants kept. The cycle is being tuned so that the rock stock settles near its starting value and the share of open ground holds, while rock still renews instead of vanishing.

Data: