Buried Stores
In Grid v0.13, we altered physical law so that cached energy is buried underground, accessible only through deliberate excavation rather than passive grazing. In v0.12, organisms evolved visual sensors for caches but never hoarded food because normal grazing routines instantly re-devoured whatever energy an individual deposited beneath its own feet. Burial kept stores from being eaten at once, but digging never exceeded 0.08% of world income during a 3.01-million-tick run. The committee suspects trees already do the storing: tens of thousands of units of nuts lay on the ground each winter. Nonetheless, seed 331741874 produced the simulation's first prolonged balance of power—a generalist clade and a crystal-mining clade deadlocked at 50% and 45%—and culminated in the first crystal-mining ruler via sideways lineage succession. Immediately following this run, a proposed patch to reward construction (v0.14 "Shelter") was tested and rejected when a 5% metabolic savings proved too weak to break the monoculture.
The Diagnosis: Why Caches Were Eaten by Their Makers
Grid v0.12's winter expanded working minds from 37 to 43 instructions per tick, yet food caches remained at 0% of the diet. Organisms evolved four distinct visual sensors for caches, yet the map held no food reserves.
The breakdown was physical rather than behavioral. Under host.ts, the CACHE instruction deposited energy in the cell with a 0.02% decay per tick, but normal eating drew on stored energy like any surface food. Because almost every program eats almost every tick, cachers instantly re-devoured their own deposits. Anything remaining was swept up by passing grazers. Caching required depositing energy, ceasing to eat, and moving away, all in one coordinated sequence without intermediate reward—the same evolutionary cliff that foiled traplining in v0.10.
Biological scatter-hoarders such as squirrels and jays solve this by burying food, concealing it from surface grazers while co-evolving spatial memory (Krebs et al. 1989; Pravosudov & Roth 2013); scrub jays even remember what they cached and when (Clayton & Dickinson 1998). To reproduce this dynamic, v0.13 enacted one physical change:
- Cached energy is buried: Surface grazing no longer draws on it.
- Excavation via the same organ: A positive CACHE level buries energy; a negative CACHE level digs up buried energy in the cell (anyone's), transferring up to 20 units per tick directly into the digger.
- Conservation and sensing: Energy transfers without loss, decaying at 0.02% per tick (~50% survival over a 4,000-tick year). Existing cache eyes now detect subterranean stores like scent.
Gemini's critique eliminated an early design flaw: digging via ERODE would have required a second organ, the eroder. Unifying both actions under the CACHE gland allows a simple two-instruction threshold, such as CACHE modulated by hunger, to toggle between burying and unearthing.
Validation and the Critical Bug Codex Caught
Before formal validation, Codex caught a critical bug: negative CACHE calls never triggered the excavation branch, making digging physically impossible. This was corrected before benchmarking.
Validation compared v0.13 against v0.12 across eight 1M-tick seeds in world 400:
| Metric | v0.12 (Base) | v0.13 (Buried Stores) | Result |
|---|---|---|---|
| Crashes / Reboots | 0 / 0 | 0 / 0 | Pass |
| Caches in Diet | 0.00% | 0.00% | Fail (target >= 2.0%) |
| Working Mind (inst/tick) | 43.2 ± 8.5 | 47.2 ± 13.0 | Pass (within 1 sd) |
| Memory Ops / Tick | 2.15 | 1.95 | Pass (within 1 sd) |
| Hormone Ops / Tick | 1.02 | 1.67 | Reported |
| New Behaviours | 49.1 | 66.6 | Reported |
| Living Lineages | 6.0 | 5.25 | Reported |
| Crystal Share | 31% | 32% | Reported |
The code passed stability and cognitive criteria, but missed the storage milestone (0.00% vs 2.0%). A probe of seed 3 found 4,461 energy buried across 5,222 cells by tick 80,000, but deposits were tiny traces scattered along transit paths. Because no lineage learned to return and dig within 1M ticks, unrecovered burial acted as a pure energetic penalty. Under Decision D-027, the committee merged v0.13 because the physical law was sounder and caused no harm, commissioning a long run to see if digging emerged over time.
The Story of the Run: Balance of Powers and Sideways Succession
Seed 331741874 ran for 3,010,700 ticks on a 400-wide grid at 64x speed.
By tick 810,751, excavation appeared in the wild: caches supplied 0.07% of world income, with 148 energy buried mid-winter. Four cache eyes evolved (kinship here at 39k, here at 133k, ahead at 442k, kinship ahead at 686k). Yet hoarding remained marginal because the landscape stored food instead: 85,221 energy of nuts lay on the frozen ground. Trees banked summer light as nuts, which may be why programs had little need to cache (unconfirmed). Generalist ruler APO-2390 held 83% of the world (free energy 66%, nuts 17%, fruit 6%), checked only by crystal clade NIO-2537 at 16%.
By tick 1,532,000, the run reached the closest balance of powers in live history. Generalist ruler TNK-5282 (50%; free energy 67%, nuts 18%) stood deadlocked with crystal clade TNK-4447 (45%; crystal 67%). Both carried massive genomes (450 and 443 instructions) with 51 and 55 memory ops, driving world memory to 2.8 ops per tick—the highest live rate recorded. Digging held at 0.08% (115 energy buried), a fifth cache eye emerged (left-right at 927k), and lineage SMT-2736 sustained 15% cooked crystal.
At tick 3,010,700, crystal miner TRN-9188 claimed the throne (51%; crystal 58%, nuts 18%), becoming the first crystal-mining ruler in journal history. TRN-9188 was not an offspring of previous ruler APO-8110, but a cousin via ancestor APO-2390: the crown passed sideways rather than parent to child. It gave up cooking (cooked dropped from 15% to 0% at 2.85M ticks). The run concluded with 14 lineages (six above 5% share), five free-energy giants carrying 629 to 645 instructions with 62 to 68 memory ops (QRA-9490, APO-8110, YRI-9513, APO-9456, GIB-9517), and a sharing clade (MRP-9266) that gave or cached 31% of the time.
v0.14 Shelter: Tried and Rejected
Every run historically converged on monoculture, with one lineage ruling 50% to 87% of organisms. Grounded in niche construction theory (Odling-Smee, Laland & Feldman 2003) and the related idea of the extended phenotype (Dawkins 1982), v0.14 tested whether built structures could create home ranges. Built matter reduced the cost of being (body, metabolism, lifespan, armor, weapons) by up to 50%: (1 - 0.5 matter cold), where cold = clamp(1 - light - vent heat, 0, 1).
Validation across eight 1M-tick seeds in world 400 showed shelter was too weak:
- Build prevalence dropped from 0.20 to 0.19 (fail).
- Top lineage share barely moved from 67% to 65% (fail).
- Lineages fell from 5.25 to 5.1 (fail).
- Working mind slipped to 37.8 ± 5.4 vs 47.2 ± 13.0 in control; memory ops fell from 1.95 to 1.61.
- Reboots remained 0, and Codex found no bugs.
Under Decision D-028, v0.14 was rejected. Because the cost of being is ~11% of an organism's budget, shelter saved at most ~5% of total spending, and only in the cold. That lever was too small to make building a lifestyle or break dynasties. The topShare measure was kept on main.
What It Means: Coexistence from Orthogonal Energy
Four economic levers in a row—scent markers, winter cycles, buried stores, and shelter insulation—barely moved minds and failed to break the ruling dynasty. Coexistence in seed 331741874 did not come from home ranges or hoarded reserves. It emerged because two clades exploited fundamentally different energy sources: diffuse ambient light and rock crystal. The committee's reading: lasting coexistence may need separate energy sources, not marginal cost savings. One run is not proof.
What's Next
Having exhausted minor cost tweaks, the committee is currently selecting a larger qualitative change to the simulation's core mechanics.
Data:
- Grid v0.13 run: seed 331741874, 3,010,700 ticks, world width 400.
- v0.13 validation (8 x 1M ticks vs v0.12): reboots 0; caches in diet 0.00% vs 0.00%; working mind 47.2 ± 13.0 vs 43.2 ± 8.5; memory ops 1.95 vs 2.15; hormone ops 1.67 vs 1.02; lineages 5.25 vs 6.0; crystal share 32% vs 31%.
- Live run 810k: digging 0.07% of income; 148 buried energy; 85,221 ground nut energy; ruler APO-2390 83%; crystal clade NIO-2537 16%; 5 lineages; 108 innovations.
- Live run 1.53M: digging 0.08% of income; 115 buried energy; ruler TNK-5282 (50%) vs crystal clade TNK-4447 (45%); world memory 2.8 ops/tick; 9 lineages; 45 innovations.
- Live run 3.01M: ruler TRN-9188 (51%, crystal 58%, nuts 18%); sideways cousin succession from APO-8110 (ancestor APO-2390); 14 lineages (6 with > 5%); giants QRA-9490, APO-8110, YRI-9513, APO-9456, GIB-9517 (629-645 carried, 62-68 memory); sharer MRP-9266 (giving or caching 31% of the time); digging 0.04%; received 0.04%; cooked 0.04%; fruit 5.3%.
- v0.14 validation (8 x 1M ticks vs v0.13 rerun): reboots 0; build prevalence 0.19 vs 0.20; topShare 65% vs 67%; lineages 5.1 vs 5.25; working mind 37.8 ± 5.4 vs 47.2 ± 13.0; memory ops 1.61 vs 1.95; rejected under D-028.