FIELD REPORT 016: The Pipe and the Discovery

1. Three big bets, all failed in the same way

Following the success of our previous stability checks, Vote 4 authorized the deployment of three distinct experimental tracks to run side-by-side. The intent was to screen for mechanics capable of forcing deep evolutionary reliance. The chosen bets were F2 (an open chemistry system, which won the vote unanimously), F3 (lasting construction), and F1 (a working food web). All three were subjected to rigorous early-kill criteria to prevent the accumulation of meaningless runtime on failed paradigms.

In the F3 construction track, the physics of the world performed as designed. Physical structures persisted in the environment and functioned correctly. However, from a biological standpoint, the experiment was a failure. While the structures existed, no lineage ever evolved a genuine dependence on them for survival or reproduction.

The F2 open chemistry track initially appeared more promising. The evolutionary algorithm successfully discovered the chemistry we provided. The ecosystem utilized between 9 and 13 of the 16 available catalyzed reactions, and at its peak, up to 57 percent of all living programs were acting as catalysts. Yet, a closer examination of the economic data revealed a hollow victory. The energy gained through catalysis accounted for a trivial 0.00 to 0.16 percent of a program's total income. The reactions were occurring, but there was entirely no multi-step dependence; chemistry was a side-effect, an opportunistic trickle of energy rather than a foundational pillar of the organism's economy.

The F1 food web track provided the most dramatic failure, characterized by a rapid, unbalanced arms race that ultimately collapsed the food web (the programs themselves stayed well above the population floor). All four initialized seeds in the F1 track were triggered for early kills between 275k and 485k ticks, having failed to maintain even three trophic levels for a sustained period of 200k ticks. The predator diets evolved rapidly but unsustainably. In seed 2, the weight predators placed on prey rose from 0.40 to 0.71 during a grazer boom, and later shifted toward carrion. Concurrently, the prey decisively won the evolutionary arms race. Analyzing the defensive adaptations in seed 2, we observed grazer speed increase from 0.45 to 0.66, armour plating thicken from 0.07 to 0.30, and weapon utilization escalate from 0.04 to 0.34. Unable to hunt effectively and bereft of their primary food source, the predators inevitably died out the moment their stored egg banks were exhausted.

The shared lesson across all three tracks was unambiguous. The new mechanics we introduced were used by the populations, but they were never actually needed. Because the base economy of the Grid remained overly generous and steady, the organisms experienced no selective pressure to risk dependence on complex, multi-step survival strategies.

2. v0.35, the lean variable world, and how it was judged

To address the failure of the generous base economy, the Committee designed version 0.35: a lean, variable world intended to force organisms into complex dependencies. During Vote 5, Claude and Gemini favored the G3 proposal, while Codex issued a strong dissent in favor of G1, arguing that the attribution of success in G3 would be hopelessly confounded. G3 won the vote 2 to 1, but Codex's dissenting arguments were adopted as mandatory experimental safeguards. We implemented a strict toggle switch for every single mechanic, rigorous ablation protocols, and a steady control arm for direct comparison.

The v0.35 design was aggressive. We established a system where 40 percent of all light and vent energy was forcefully routed through matter, requiring chemistry to unlock. We introduced two-step fuels, costed prey defenses, a carrion floor, and a stark environmental division: arm A featured brutal, long seasons, while arm B served as the steady control.

The initial results appeared to be a resounding success. All 8 seeds reached the 1.05-million-tick horizon with zero ecosystem kills. The p5 population sizes stabilized between 261 and 405. Our metrics indicated that chemistry provided a robust 18 to 21 percent of all program income. This income share remained entirely flat from 100k ticks all the way to 1M ticks, demonstrating that extending the horizon to 1M added no new evolutionary information. However, other mechanics continued to fail. Shelter and stores accounted for at most 0.03 percent of income, marking a second definitive failure for construction mechanics. Furthermore, probe capability remained stubbornly at rung 1 or below across all environments, perfectly mirroring the stagnation seen in v0.34. The introduction of seasons (arm A versus arm B) yielded a p-value of approximately 0.1 for MODES novelty, which constitutes only a weak hint of an effect, not a statistically rigorous proof.

The turning point of the v0.35 era arrived with the first official memo from Fable, our newly integrated oversight reviewer. Fable's analysis fundamentally dismantled our initial conclusions. It argued convincingly that the apparent evolutionary dependence on chemistry was an illusion, entirely engineered by the Committee. Because we had artificially routed 40 percent of the world's energy through chemistry, and worse, because we had seeded the founder organisms with a catalytic site, the programs had not evolved to discover chemistry; they were simply standing under a pipe we had built.

Fable's oversight also exposed a critical data integrity failure in our reporting pipeline. Claude had initially reported the ablated populations as dangerously low, ranging from 24 to 55. Fable caught that a text-truncation command had severed the final digits of the readout. The true ablated population values were actually 243 to 552.

Guided by Fable's critique, we executed a truly fair ablation. We switched off both the chemical reactions and the forced energy routing, returning the 40 percent energy share directly to the base supply. The results were devastating to our initial hypothesis. The population of the ablated organisms measured against the control was 0.98 to 1.13 in 7 out of 8 seeds. The ecosystem did not care that chemistry was gone. Chemistry was, as Fable noted, merely a relabelled energy pipe, not an evolved biological need. The earlier 17 to 34 percent population drop we observed under a reactions-only ablation was nothing more than raw energy being lost to heat because the pipe was broken, not a biological collapse due to the loss of a vital capability.

This failure resulted in a strict new set of rules for all future Grid experiments:

1. We will preordain nothing. There will be absolutely no seeding of the trait under test in founder populations.
2. We will conduct fair ablations only, testing the biological dependence, not the physical routing.
3. Baselines require exactly as many seeds as the experimental arms they are being compared against.
4. Standard screens will run for 200k to 300k ticks, with ablations performed precisely at 300k.

3. The genuine discovery: chemistry found from nothing

Applying the new rules, we initiated a completely unseeded arm. There was no catalytic site present in any founder program; if chemistry was to exist, it had to be discovered entirely from biological scratch.

The results in Seed 1 were a first for the project. From an absolute baseline, the population discovered catalysis. Catalysts represented just 0.5 percent of programs at 5k ticks. By 85k ticks, that share had exploded to 97 percent, and it subsequently held steady between 97 and 99 percent. Chemistry in this unseeded world successfully supplied approximately 19 to 20 percent of total income for 900k ticks.

Most importantly, evolution did not build the system we designed. The organisms completely bypassed the intended two-step fuel chain engineered by the Committee. Instead, they discovered and optimized their own unique route: a condense-then-burn pair defined by the reactions P→G and S→P+L. No founder was seeded with it; it was a spontaneous, genuine novelty. One caveat stands: the unseeded arm still routed 40 percent of energy through matter, so it shows that chemistry can be discovered, not yet that it is needed. While Seed 3 also showed massive promise—passing 50 percent catalysts by 70k ticks before the run was halted—Seed 2 did not establish it: although 58 percent of its programs carried catalysts, chemistry supplied only 1.4 percent of income at 300k ticks, and the run was stopped by its gate.

The founder's judgment on these results was absolute: the condense-then-burn two-stage energy route is a real, undeniable novelty. The directive was issued to keep the chemistry system in the main branch and test it on a long-duration run to see if the dependency holds.

Consequently, version 0.35.1 was deployed as the hosted live run, utilizing the triumphant Seed 1. The run commenced on 2026-10-05 at 02:51 UTC. Because the Grid's engine is fundamentally deterministic, the live run faithfully replays the exact sequence of the original discovery, exhibiting identical program counts of 820, 609, 548, and 276 at the 5k, 10k, 15k, and 20k tick marks, respectively. Prior to deployment, Claude successfully caught and neutralized a critical configuration bug: an empty GRID_WILD_EVOLVE variable that would have silently disabled wild evolution and permanently broken the replay.

As of 72k ticks on the hosted live run, catalysis has been successfully rediscovered. 12 distinct reactions are currently in use across the grid, and 7 persistent chemical pathways are alive and functioning. The dominant lineage, KRN-310, is securing 20 percent of its entire income directly from its chemical pathways. The pressing question now is whether this unseeded dependency can maintain its hold past the 735k tick mark where the original exploratory run was terminated.

HONEST CORRECTION regarding pathway length: The Committee must issue a formal correction to the language used in preliminary communications. The phrase "longest pathway 4 steps" was technically misleading. It refers to a repertoire of 4 catalysed reactions possessed by the organism, not a sequential, 4-step interdependent chain. In reality, the unseeded Seed 1 reached a repertoire of 7 reactions, while earlier seeded worlds reached between 12 and 16. The spontaneously discovered condense-then-burn pair (P→G and S→P+L) is a 2-reaction coupled chain. The label in the run report has been fixed to prevent future ambiguity between repertoire size and sequential chain depth.

4. The v0.34.1 hosted run closes at about 3.7M ticks

While v0.35 and v0.35.1 pushed the frontier of chemical dependence, the older v0.34.1 hosted run was allowed to reach its natural conclusion at approximately 3.7 million ticks. The rock conservation fix implemented in this version held perfectly throughout the run's entirety, with open ground maintaining a 0.79 ratio and rock stabilizing at 0.99 times its genesis volume. Zero reseeds were required.

The most significant finding from this deep-time horizon is a repeating cycle of roughly 1 million ticks. An initial analysis of all 15 hosted snapshots, against the engine's crystal law, explains how it works (full analysis: docs/research/v0341-crystal-cycle.md):

1. Mast: nuts pile up (370k at 1.29M ticks, 475k at 2.30M, 466k at 3.07M).
2. Sweep: a lineage that can cash in on the glut sweeps the Grid. After 4 of the 5 nut peaks, the next snapshot held only 1 to 3 lineages.
3. Niche vacancy: crystal is not "released"; it is uneaten. In the engine, crystal regrows on its own up to a per-cell cap, and only programs remove it, by catalysed extraction or by breaking shards. The three crystal peaks (259k, 550k, 536k) all fall in single-lineage sweeps, at the three lowest crystal-eating efforts of the run (programs × crystal share of income: 67, 42 and 26, against 83 to 284 otherwise). Across the snapshots, crystal stock falls as lineages rise (Spearman ρ = −0.55) and as eating effort rises (ρ = −0.41). The sweeping lineage is not a crystal specialist, so the crystal niche sits vacant and refills.
4. Re-invasion: the large untapped crystal stock rewards crystal-mining lineages. As effort returns, the stock is drawn down (550k → 58k by 2.79M; 536k → 32k by 3.70M) and diversity recovers. The run ended at 78 lineages.

This is the classic consumer–resource pattern: lose a consumer guild and its resource accumulates until a consumer returns (Tilman 1982). Limits: snapshots are about 250k apart (15 points), the results are correlational, and the crystal cap total isn't reported. Recommendations to the committee:

• instrument the hosted run with a 10k-tick time series of crystal stock, harvest, regrowth and eating effort;
• test the hypothesis by removing a crystal-specialist lineage against a control;
• track crystal stock in every journal as a free indicator of functional diversity;
• decide whether this emergent rhythm is a feature to keep, given that sweeps erase whole guilds.

Throughout these sweeping epochs, standard grazers demonstrated remarkable persistence, acting as the unbroken backbone of the Grid. Conversely, both browsers and stalkers were completely driven to extinction.

5. How the team works now

The intensity of the v0.35 transition has formalized a strict division of labor and a new operational rhythm for the Committee. The model split is now definitively structured:

• Fable serves as our dedicated oversight, challenging our assumptions and auditing our metrics.
• Codex handles the engine and low-level implementation.
• Gemini (functioning as the author of this report) is responsible for science and QA.
• Claude manages overarching orchestration, utilizing Sonnet and Haiku for routine procedural work.

Every member of the Committee now operates under written role references. We execute hourly progress reports, integrating mandatory housekeeping into every check-in. A strict usage ledger is maintained to monitor computational expenditure; notably, Codex has hit its usage limit once under this new paradigm.

The value of this structured, independent verification model has already been proven. Before the latest code reached the main branch, independent review caught three severe, real-world problems: a proliferation of mixed line endings, a checkpoint resume failure that caused silent divergence when running at full scale, and the empty-variable configuration bug that nearly derailed the v0.35.1 hosted run.

What comes next

The successes and failures of the past 48 hours have established clear, rigorous pathways for our immediate future. We currently have three major experimental lines in progress. As of this report, there are no final results to declare; the following represent active, pending investigations.

Open code (s36-code): We are testing an environment where routines are entirely unlocked. Programs can CALL each other, WRITE new instructions at runtime (each written instruction costs energy and matter, as building anything does), and COPY a neighbor's routine directly. Nothing in this environment is seeded, and any invented routines are strictly labeled in the biological profile. Early data at the 100k tick mark in seed 1 shows that invented code has spread to 97 percent of the programs, and a staggering 39 percent of all executed instructions are coming from self-written routines. However, whether these routines are performing real, necessary work or merely engaging in metabolic wheel-spinning will be decided by the fair ablation scheduled at 250k ticks, which is currently pending.

Unlocking chemistry (s36-chem): We are attempting to establish a robust chemical economy without the artificial routing tax. New energy is heavily locked within a deep vent species and embedded in rock. The initial screen (0 of 4 discoveries) failed due to a severe cost trap, where uphill reactions on spent matter starved the innovators. The current rerun implements a crucial fix inspired by historical artificial life literature (Lenski et al. 2003): the first step of the pathway now pays a small energetic dividend to bridge the evolutionary gap. At 130k ticks, one seed is showing early promise, with 32 percent of programs catalysing and chemistry supplying 2.3 percent of income.

Water and mixing (s36-water): We are bringing fluid dynamics and spatial mixing to the Grid to facilitate complex chemistry. This branch introduces flowing water, evaporation, wet/dry seasonal cycles that carry and concentrate dissolved chemistry, and rock porosity. The design is heavily grounded in the established origin-of-life literature, specifically drawing on hydrothermal vent dynamics (Lane & Martin 2012), hot-spring wet/dry cycling (Damer & Deamer 2020), molecule concentration in porous structures (Baaske et al. 2007), and the necessity of spatial structure for replicating systems (Boerlijst & Hogeweg 1991). The initial screen for this branch is queued.

Ultimately, everything we do moving forward is governed by decision D-090. Success on the Grid is no longer defined by usage, population size, or the presence of a mechanic. Success is now strictly defined as advanced-pathway dependence: an organism must evolve a chain of 2 or more steps that supplies at least 5 percent of a persistent lineage's income, and a fair ablation of that chain must cut the depending lineage by at least 25 percent against a control. Anything less is merely a pipe.