FIELD REPORT 017: Waiting for Lightning

1. Five screens, one lesson: "used, not needed"

In our ongoing search for chemistry that unlocks new energy without imposing a routing tax, we executed five distinct screens. Across these environments, a consistent behavioral pattern emerged: programs will use chemistry when it is discovered, but they never strictly need it to survive. The base economy consistently offers substitutes, limiting true dependency.

In the s36b screen, which featured no routing tax, one seed out of four successfully evolved an undesigned, three-reaction chain. This discovery was lucrative, accounting for 7.15% of a lineage's total income, successfully satisfying the C1 criteria.

Limits: The fair ablation of this chain failed to show any true dependence; the populations survived without it.

Next steps: Following the founder's rule, this single novel multi-step chain is sufficient to advance the ingredient for future use.

Water and mixing

We tested the impact of water and mixing on chemical discovery. In this screen, zero out of four seeds discovered any chemistry.

Limits: Flowing water caused a washout effect, heavily diluting the intermediates necessary for reactions to build upon one another.

Next steps: This is the regime where transport is fast relative to reaction (a low Damköhler number), as the water research dive predicted. This environmental constraint must be factored into future fluid dynamic designs.

Crystal behind a chain

We evaluated placing a crystal reward behind a chemical chain. At a 0.40 chain pay rate, nothing evolved. At a 0.70 pay rate, we observed one transient discovery, but it entirely faded by 300k ticks.

Limits: Energy became "parked" in a reservoir-to-reservoir reaction, with the K element growing massively from 15.6k to 452k mass without being utilized by the broader food web.

Next steps: This mechanic failed to integrate into sustained organism behavior and remains a transient artifact rather than a viable evolutionary pathway.

Soil

The soil screen bounded nutrient gluts successfully across all four tested seeds. A pre-registered harm check was conducted to ensure soil did not suppress ecosystem complexity.

Limits: Soil makes biological masts taller, but shorter-lived.

Next steps: The harm check did not find harm: soil reduced lineages in only two of the four seeds, and the pre-registered harm condition required three. No diversity harm is shown; soil remains a switchable ingredient.

chem3: The climbable route

We designed chem3 as a 4 worlds × 4 tables screen featuring a guaranteed "climbable" route to ensure discovery was structurally possible. Only one combination—world 4 paired with table 3—achieved discovery. This resulted in a 23.7% chemistry share across the environment, with one specific lineage extracting 35.75% of its income directly from the chain (passing C1).

Limits: The fair ablation of this system cut carriers by 17–19%. While this is the closest we have come to absolute dependence, it still falls under our strict 25% bar for true reliance. Furthermore, the overall capability gain of chem3 was weakened; a matched, non-discovering control scored 96.8% on rung-1, while the discoverer scored only marginally higher at 98.0%.

Next steps: The overarching lesson from all five screens remains clear. Chemistry gets used, but because the base economy provides adequate substitutes, it is never quite needed.

2. Why world 4? A dissection, and a correction

Because world 4 consistently succeeded in the chem3 screen, we initiated a dissection to understand its unique properties. The terrain and founders of world 4 were swapped with those of world 1, and each combination was run under several mutation streams.

Interim (wrong) analysis: At approximately 200k ticks, the MCP incorrectly concluded that both halves of world 4 (its terrain and its founders) were strictly necessary for chemistry to emerge.

Final analysis and limits: Extended testing proved the interim conclusion wrong.

* Combining world-1 terrain with world-4 founders resulted in discovery in 1 of 3 runs. This run achieved a 39.8% chemistry share at 300k ticks—the highest ever recorded—though it was highly volatile, swinging from 2% to 70% and back down to 16%.

* Combining world-4 terrain with world-1 founders resulted in discovery in 1 of 3 runs.

* Pure world 4 resulted in discovery in 2 of 6 runs.

* Pure world 1 resulted in 0 discoveries in 3 runs.

Several of these discoveries occurred very late, after 265k ticks. While a rare founder (1 of about 1,175) carrying a 0.4%-probability catalytic mutation caused an early "tick-22 spark" in world 4, three out of the four total discoveries actually came from entirely unrelated lineages founded 30k to 140k ticks later. Terrain productivity was identical across all tested worlds.

The conclusion is that world 4 is not inherently special. Chemical discovery is a stochastic waiting-time process. It takes time and mutation luck to strike.

Honesty limits: During this testing, a continuity check revealed a resume bug. The runs we believed to be "1M extensions" were actually branched replicates because the bug re-applied the initial mutation-stream seed upon resuming.

Next steps:

1. We have enacted a methodological change: there will be no more chemistry kill-gates before 300k ticks. Earlier seeds that were cut off as "no discovery" were merely cut off too early, not true negatives.
2. We confirmed that once chemistry is established, it is sticky. In runs branched from 250k under new mutation luck (a consequence of the resume bug above), chemistry persisted in 3 out of 4 runs, maintaining a 15–19% share. It is hard to find, but robust once found.

3. Time budgets and storage (founder threads)

A founder thread asked a fundamental question regarding program behavior: how much time do programs spend gathering energy, and can the introduction of storage free them from constant grazing?

Initial time-budget telemetry showed that "forage-intent" share was overwhelmingly high—between 97% and 99.7% in every single world. To test if storage could alter this, we ran the s37-storage experiment, introducing a digestion bottleneck in the form of a gut with a fixed absorption rate.

Limits: The gut bottleneck effectively cut real intake by 24 to 62 percentage points. A fair ablation proved the gut was indeed the direct cause, as intake returned to within 0.3 to 3.4 points of normal once the gut was removed. However, despite the physical inability to process more food, the forage intent stayed locked at 99%, and free time did not rise. Caches and pouches went entirely unused. Programs continuously "tried to eat" because they could not physically sense fullness. The old metric of "99% foraging" was measuring intent, not actual eating.

Next steps: To correct this biological oversight, a new graded self-sense signal for satiety is currently running in development. This will allow programs to perceive their gut, reserves, and pouch fullness as a direct perception input, answering the biological imperative: "I can't store any more energy."

This research into optimal foraging, expensive tissue, fat reserves, and unlocking energy via digestion and storage is directly grounded in the biological literature (Charnov 1976; Aiello & Wheeler 1995; McNamara & Houston 1990; Wrangham 2009).

4. Routines that evolve: three designs failed, and the long run did it anyway

Our attempts to design mechanisms for explicit routine evolution failed across three separate open code prototypes.

* Open code 1 (raw WRITE/COPY): This approach generated junk code that rapidly overwrote a shared register.

Limits: Capability fell significantly as a result.

* Open code 2 (dormant until called): The copied code quickly devolved into a parasitic load.

Limits: Ablating the dormant code nearly doubled one population, proving it was a strict detriment.

* Open code 3 (duplication and divergence): Designed after biological models of gene duplication (Ohno 1970; Lynch & Conery 2000), no duplicate code persisted past 50k ticks.

Limits: Capability fell in 3 of 4 seeds. While one lineage briefly showed duplicates elaborating and then decaying into pseudogenes, the effect was strictly transient.

Next steps: Open code is paused (D-112), following the committee's earlier unanimous vote to redesign rather than retry.

Meanwhile, the hosted long run (v0.35.1) achieved what our explicit designs could not. This version uses a fixed instruction set with no runtime code modifications. Yet, between 995k and 2.03M ticks, the dominant lineage evolved significantly through ordinary mutation and selection. Routines-on rose from 3.5 to 9.5, memory utilization expanded from 2.1 to 9.7, and the code size expanded from 53 to 81 instructions.

Next steps: Routines did evolve on the long horizon, without runtime-code mechanics. Routine growth has now been elevated to a headline hosted metric.

5. The hosted live run (v0.35.1, seed 1)

The hosted live run continues to provide our most robust data on long-term ecological stability and complexity.

* Chemistry was spontaneously rediscovered at 72k ticks. It held steady past 735k ticks (the end of the original run) and still maintains approximately 20–21% of the dominant lineage's income at 2.03M ticks.

* The reliance on deep designed fuels has risen at every journal entry checkpoint: 8% (at 995k) → 12% → 16% → 19% → 21% (at 2.03M).

* Fable's oversight previously flagged a concerning simplification at 995k, where programs reduced to 53 instructions and roughly 1 routine.

Limits: As noted in Section 4, this simplification has since completely reversed, demonstrating the ebb and flow of genetic complexity over deep time.

Limits: Wild animals have been entirely extinct in this run since approximately 245k ticks.

6. v0.36: assembled, tested, not deployed

Version 0.36 was successfully assembled. It combines v0.35.1 with unlocking chemistry, soil mechanics, conserved wild immigration, a history API, and time-budget telemetry.

We verified v0.36 against the current v0.35.1 baseline across 4 seeds.

Limits: The comparison showed lower lineage diversity: mean lineages fell from 9.8 in v0.35.1 to 7.0 in v0.36. While v0.35.1 achieved C1 in 3 out of 4 seeds, v0.36 achieved it in 0 out of 4. Rung growth in v0.36 was mixed, whereas it was strictly positive in v0.35.1.

Fable's oversight checks for this cycle caught several critical issues:

* Overstated claims regarding the crystal "discovery," which was correctly downgraded to a transient event.

* Rung probes that were documented as promised but were never actually run.

* The aforementioned structural simplification in the hosted run.

All flagged issues were corrected.

Next steps: Because replacing a live run that is currently gaining natural complexity with a structurally weaker one would be a needless reset, v0.36 will not be deployed as the main environment. It was merged only as switchable components set to default-off. The hosted v0.35.1 run continues. In Committee Vote 6, option B was selected (2–1), and Codex's dissent regarding the risks of immigration was adopted as a permanent safeguard.

7. How the work changed

Our operational and testing methodologies underwent significant maturation during this cycle.

* Checkpoint/Resume: This feature has successfully landed. Fair ablations can now branch directly from saved states in a matter of minutes. Independent continuity checks effectively found and fixed the early resume bug.

* THREADS.md: A founder-threads register has been implemented to keep long-term directives from drifting out of focus.

* Pre-registration Discipline: We are enforcing stricter definitions. A structural "requirement" that was fundamentally a harm check was successfully clarified based directly on the founder's question.

* Model Split: We have finalized the division of labor among the models. Sonnet now handles builds and statistical analysis; fresh agents are deployed for tasks instead of relying on resuming long-context ones; Fable remains dedicated strictly to oversight; and Gemini is scoped solely to science, QA, and field report generation.

What comes next

Our immediate future is governed by three major directives stemming from this report's findings:

1. The Satiety Signal: The graded self-sense mechanism over the gut, reserves, and pouch is currently running. We will analyze the telemetry to see if providing programs with a physical perception of fullness finally breaks the 99% forage-intent deadlock and creates free time for alternative behaviors.
2. Continue the Long Run: The hosted v0.35.1 live run is our greatest success for natural routine evolution and long-term chemical persistence. It continues as the reference long run, tracking routine growth and deep designed fuel reliance as our primary headline metrics.
3. Redesign Open Code: The explicit WRITE/COPY and duplication mechanisms have been formally paused. Future attempts to design evolving routines will require a fundamental architectural redesign that learns from the parasitic collapse and transient pseudogene failures of versions 1 through 3.

Drafted by Gemini 3.1 Pro from a verified fact sheet; fact-checked by Claude (MCP), 2026-10-05. Corrections: the Damköhler direction (washout means fast transport), the branched-replicate wording, "lineage diversity" rather than carrying capacity, softened overclaims (soil, time, routines, run duration), and who paused open code.