FIELD REPORT 018: The Miners' Ratchet

Erratum (2026-10-07)

Later the same day the committee found that this report tracks the wrong kind of complexity in places. The findings below replace the corresponding claims in the text, which is otherwise left as first published. Sources: decisions D-128, D-130, D-131 and D-134.

• Section 1 measures code carried, not code run. The miner series (276, 348, 422 instructions) is the size of the genome. Over the same three snapshots the miners ran 90, 93 and 94 instructions per tick and had 4.1, 3.8 and 4.1 routines switched on. After this report the carried figure went on to 447, 524 and 522 while code run per tick went 90, 70 and 60. Carried code nearly doubled; expressed complexity was flat and then fell. Unused code costs nothing to keep, so this is consistent with neutral accumulation. "The miners' ratchet" is a ratchet in carried code only.
• Section 2's replacement explanation is not supported either. The pump was rightly set aside, but the text then credits "the intrinsic computational cost of the niche" with a continuous ratchet on complexity. There was no rise in expressed complexity to explain.
• Behaviours named in section 1 are skill levels, not observed acts. "Gained a build behaviour" and "specializes in giving and marking" mean a skill level crossed a reporting threshold.
• Section 3's chemistry table is carried code. On expressed measures the 5M chemistry run executes a flat 52 to 62 instructions per tick after 500k, and routines switched on fall from 4.3 to about 2.5. The dependence result (D-115) is unaffected.
• Section 4's routines are carried totals, and its closing step has changed. On expressed measures, moderate against mild seasons over five paired seeds gives +12 instructions run per tick and +0.9 routines switched on, winning 4 of 5 pairs, which is within seed-to-seed noise. The founder has since waived the decision to stop seasons work (D-131): a near miss with a net positive effect is a reason to continue. The pull requests mentioned are on hold; work continues locally.

What stands: the lineage counts, the crystal and chemistry readings, the population-floor rule, and the withdrawal of the pump hypothesis.

1. The Hosted Live Run and the Miner's Ratchet

Complexity is scaling continuously within a specialized ecological niche, untethered from broader environmental collapse or vacancy. In the hosted live run (v0.35.1, seed 1), a guild of crystal miners has broken previous programmatic ceilings, ratcheting up code size, routine count, and memory operations across successive descending lineages.

Earlier in this run (around 3.62M ticks), the high-water mark for complexity was the dominant generalist lineage ANO-22524, which sat at 175 instructions, 11.6 routines, and 16.6 memory operations. At 995k, the run's programs had simplified to just 53 instructions. However, the miner line has since established a continuous, compounding ascent:

These miners sustain this massive codebase by surviving on approximately 40% crystal and 33% "cooked" energy, utilizing Arrhenius kinetics for fast, catalysed extraction near vent heat. By contrast, the generalist line has plateaued at roughly 165–185 code, 9 routines, and 20 memory operations.

At 4.26M, the generalist line radiated into 9 distinct daughter lineages (bringing the total to 11, up from 4). This radiation introduced entirely new behaviors: one daughter lineage of 56 programs specializes in giving and marking while surviving mostly on nuts; other small lineages now build and mark; and a highly specialized 2-program carrion eater has emerged with a formidable footprint of 330 instructions, 15 routines, and 39 memory operations.

The underlying ecology remains highly active but stable in its mechanisms. Chemistry accounts for about 19% of the generalist's income, featuring a 7-step longest pathway and 29% deep-route income. The population undergoes boom and bust cycles roughly every 320k ticks, swinging between roughly 340 and 740 programs. Wild animals have remained extinct since about 245k ticks.

Limits: Our observation of this radiation and the continuous ratchet is limited by the single-run sample size of the hosted environment.

Next steps: The hosted run is not redeployed while it still yields insights; we will await the founder's decision on when to conclude this run, continuing to monitor the bounds of the miners' memory expansion.

2. The Complexity Pump Hypothesis, Tested and Weakened

Complexity requires robust theoretical frameworks, and when the data does not support a mechanism, we must honestly discard it. We initially hypothesized that the cyclical nature of crystal availability acted as a "complexity pump." Crystal regrows slowly and only programs remove it. At 3.79M, the crystal stock had risen from 35k to 127k while miners were few. The MCP predicted a drawdown if the miners expanded. By 4.05M, miner numbers grew from 49 to 77, and the stock fell to 44k. The prediction held, suggesting that a rich, untapped resource selects for the most capable re-invaders, pumping complexity upward with each cycle.

We formulated alternative hypotheses to this pump: the intrinsic computational cost of the niche (cooking simply requires more routines), relaxed selection during a resource glut, or a spatial arms race. A local survey of 21 runs was conducted to test the pump mechanism. The results were unsupportive:

* A 5M chemistry world showed no cycle.

* 5 out of 20 seasons worlds showed 9 cycles total.

* Only 2 cycles looked like a pump (one strongly so, where code jumped from 32 to 78, memory ops from 0.6 to 2.8, and a lineage was replaced).

* The remaining 3 cycles were flat or showed downward complexity.

Therefore, by decision D-125, the complexity pump is not supported as a general mechanism, and no clamp experiment will be built.

The decisive observation came at 4.26M in the hosted run. The miner line gained an additional 74 instructions (from 348 to 422) with no vacancy, evolving through continuous descent. This favors the alternative explanation: the intrinsic computational cost of the niche acts as a continuous ratchet on complexity. Subsequently, the crystal stock rose again from 44k to 121k despite unchanged miner numbers.

This dynamic engages directly with classical models of resource competition (Tilman 1982), the origins of complex features (Lenski et al. 2003), and adaptive radiation triggered by digital resource competition (Chow et al. 2004).

Limits: The exact cause of the latest crystal stock rise without a drop in miners remains unresolved; we suspect a lean season may have slowed the cooking behaviors, but this is unconfirmed.

Next steps: We will abandon the "pump" framing as a general mechanism, leaving the intrinsic niche-cost hypothesis as the leading explanation for further observation.

3. Chemistry Becomes Needed

Complexity is most profound when an ecosystem transitions from merely using a mechanism to fundamentally depending upon it. In decision D-115, analyzing a local chemistry world (seed 4, an unseeded discovery), we found that chemistry crossed this critical threshold. Fair ablations performed after 750k ticks cut carriers by 25–33% and shrank the world population by 24–37%. This represents the first time a run has passed the C2 threshold (at least a 25% drop): chemistry went from being "used" to being "needed." This dependency was replicated at 1.5M ticks, where the population dropped to 0.67× its normal size when chemistry was removed.

To observe the long-term limits of this dependency, the run was executed locally for 5 million ticks (D-118).

The analysis shows that chemistry usage is permanent, stabilizing at a 15–22% share for the full 5M ticks. A dedicated chemistry-specialist lineage (deriving more than 80% of its income from chemistry) persisted for the vast majority of the time between 1.5M and 5M ticks. A suspected second specialist emerging at 3.34M (84% income) was determined to be only transient, returning to a 0.21–0.27 share.

While code size nearly tripled over the 5 million ticks, memory operations plateaued at roughly 3 ops—a mere tenth of the complexity seen in the hosted miners—and overall behavioural breadth actually fell.

Limits: The C2 dependency threshold (the 25% ablation drop) was not re-tested at the extreme 5M tick mark, leaving it unknown if the dependency deepened or relaxed in late stages.

Next steps: Future deep-time runs must re-test C2 periodically up to their conclusion to verify if structural dependency is maintained alongside the permanent chemistry share.

4. Seasons: A Promising Result, Honestly Shrunk

Complexity must be rigorously audited against pre-registered criteria to prevent us from over-reading small samples. We investigated environmental fluctuation via seasons to see if it drives structural complexity. Initial ladder testing (D-123) across 2 seeds for 1M ticks suggested a strong effect: moderate seasons (S2) with heat-sharing yielded 11–12 routines. This heavily outperformed mild seasons (S1, 6–9 routines) and harsh seasons (S3, 4.6 routines, which was stopped by the population floor on one seed).

To confirm this, we pre-registered a rigorous test (D-124) across 4 new seeds for 1M ticks. The arms were Moderate with heatshare (S2H), Moderate without heatshare (S2N), and Mild with heatshare (S1H). The success rule was strict: S2H must beat S1H by at least 2 routines and must win at least 3 out of 4 seeds.

One S1H seed was killed by the population floor at 400k. The paired difference between S2H and S1H was +1.85 (short of the +2 bar), and S2H won only 2 of the 3 surviving pairs. The pre-registered hypothesis was NOT CONFIRMED.

Analysis of this failure reveals crucial dynamics. First, heatshare adds nothing to complexity (S2N actually outperformed S2H). Second, while we over-read an n=2 in D-123, moderate seasons do provide a more consistent baseline for complexity: every S2 run maintained at least 6.6 routines, and all 8 S2 runs safely passed the population floor. However, the real complexity effect size is approximately +2 routines, not +4. Furthermore, this structural complexity comes at a trade-off: S2 lowers behavioral breadth (about 6.5 action kinds against 8.6) and depresses overall population numbers.

Limits: The sample size remains bound to 4 seeds per arm, and the behavioral breadth cost of moderate seasons presents a conflicting metric against routine growth.

Next steps: Heatshare is officially discarded. The next hosted candidate will be moderate seasons without heatshare. The code merge awaits the founder's approval (PR #35), and deployment will occur only when the current hosted run stops yielding insights.

5. The Mission, Sharpened

Complexity stands first and foremost. Following the founder's explicit directives, we are recalibrating our metrics. Dependence, time budgets, and storage are not the goals; they are strictly means and diagnostics for evaluating the overarching drive for complex program structures and behaviors.

This cycle has demonstrated the committee's commitment to honest, data-driven science. We have made two major, honest corrections: shrinking the previously overstated effect size of D-123, and dismantling the compelling but ultimately unsupported "complexity pump" framing within a day, when the live run weakened it.

To safeguard the ecological viability of these complex runs, a new F-c population floor rule has been enacted: the 5th percentile of the yearly mean must be at least 150, the hard minimum must never drop below 30, and there must be no extinction or reseed.

A suspected telemetry bug turned out to be by design (corrected after first publication). In the local chemistry world, crystalShare reads 0 because that world puts crystal behind a chemical chain: mined crystal dissolves into chemical C and pays out later as chemistry income, not as crystal income. The hosted miners' crystal shares come from a world without that chain.

Limits: Crystal-mining effort in chain worlds must be measured from mined mass, not income share.

Next steps: None needed; the analysis tools use mined mass in chain worlds.

What comes next

Scientifically, we remain in an observational holding pattern on the hosted v0.35.1 run; it will not be redeployed while the miners' ratchet continues to push boundaries and yield novel radiations. Once the founder decides this run has exhausted its insights, moderate seasons without heatshare is the candidate for the next hosted deployment; its code (PR #35, switchable and default off, gate-tested for byte-identity) awaits the founder's merge.