Does the Resource Buffer Pick the Genes? Three Stable Profiles Compared

Update (2026-05-12): A 6-seed-per-profile follow-up revisited these claims. The robust signal is that speciation diverges across all three profiles; the single-seed learning_rate/ensemble_size direction-flip findings were not robust under replication. See When one seed disagrees with six.

Update (2026-05-18): Crossover follow-up is complete. Conservative profiles show robust speciation compression under gene flow, while buffered profiles keep diverging trajectories. See Gene flow and the buffer.

A natural follow-up to the hyperparameter genome experiments is whether the amount of food in the world changes what gets selected, not just how many agents survive. So I ran the intrinsic-evolution runner three times with everything held fixed except the stable initial-conditions profile — small adjustments to the starting resource pool, count, and regen rate — and compared the outputs.

Short answer: yes, and in more interesting ways than I expected. The direction of selection on most ecological/behavioural genes is shared across all three regimes, but a few core knobs (notably learning_rate and ensemble_size) flip sign with the buffer. The cluster structure of the population does the same: under abundant food clusters keep separating, under tight food they merge.

The raw numbers and side-by-side tables live in stable profile comparison. This post is the narrative.

What was held fixed, what changed

Every run is 1,000 logged steps after a 200-step warmup, single seed (42),
30 founders seeded by independent mutation, mutation rate 0.15 / scale 0.10,
crossover off, GMM
speciation tracking, and selection_pressure="low" (a small density-dependent reproduction cost, no global carrying-cap term).

The only thing that varies is the resource buffer:

The deltas are small on purpose. I wanted to see how sensitive the selection signal is to a modest change in food, not crash one of the simulations.

The three profiles, in numbers

After the 200-step warmup, the populations look very different. After 1000 more steps, they look much more alike:

So the resource buffer mostly stretches the transient. By the end, all three are in the 88–99 range. What it does not equalise is the genome each run ends up with.

Where the genes agreed

Across all three runs the population drifts in the same direction on the behavioural and ecological knobs. Cheaper, less aggressive, less social, less wandery, smaller networks, and a small but consistent shift toward more prioritised replay and slower target-network syncing:

• attack_weight and share_weight both fall — by 9–28% depending on variant. Aggression and altruism are both expensive when there is no explicit signal that either one pays off.
• move_mult_no_resources falls 15–22% in all three. Wandering when hungry doesn’t survive selection.
• memory_size falls 8–28% and dqn_hidden_size falls 12–22%. Smaller brains beat bigger brains when the rest of the policy is fixed and reproduction is what matters.
• per_alpha rises 4–12% and target_update_freq rises 5–34%. Both are consistent with selection for more stable, less-thrashy learning.

This is the most robust pattern in the comparison: it shows up regardless of how much food the world has. With low selection pressure and no crossover, agents drift toward the cheapest viable behavioural profile.

Where they disagreed

The fun part. A handful of genes flip direction by buffer:

learning_rate is the clearest one. With more food the population selects for faster learners (+23% in buffered); with less food it selects for slower, more conservative updates (-8% in conservative). Balanced sits in between. There seems to be a real gradient: how aggressive the learning prior should be depends on how much slack the environment offers a mistake.

ensemble_size is the most striking single-locus shift in the comparison. Conservative chops it by 26%, balanced trims it slightly, buffered does nothing. A “compress the brain when starvation gets close” signature.

The reproduce multipliers tell a coherent story too. Buffered evolves selective reproduction — breed when wealthy, hold off when poor. Conservative evolves opportunistic reproduction — when food is short, the survivors are the ones who reproduce whenever they can, especially when poor.

Reading the speciation traces

The speciation index is a silhouette score on chromosome-space clusters (see intrinsic evolution docs). Higher means clusters are better separated. Plotting it over time tells you whether the population is splitting into niches or collapsing back toward one type. The three runs all start in the 0.65–0.73 range and diverge:

• Conservative: 0.732 → 0.713 → 0.684 — monotonically declining. Clusters merge.
• Balanced: 0.708 → 0.654 → 0.711 — V-shape. The population briefly consolidates and then re-splits.
• Buffered: 0.653 → 0.729 → 0.753 — monotonically rising. Clusters separate further over time.

Resource buffer not only sets what evolves on average, it sets how much the population can afford to spread out in chromosome space. With abundant food, sub-populations have room to drift apart from each other. With tight food, the survivors look more and more alike.

Lineage trees

All three runs hit the same maximum lineage depth of 3 (great-grandchildren make it to the final snapshot), but the breadth differs. Buffered keeps the most founders alive at the end (18 of 30), conservative and balanced both keep 15. Conservative has the flattest tree (mean depth 1.011), which together with the falling speciation index suggests a population mostly made up of “founders who never became grandparents”.

What this is and isn’t

This is a single-seed comparison across three closely-spaced resource regimes. The qualitative pattern — convergent selection on cheapness, direction flips on learning_rate and ensemble_size, and a clean buffer-vs-speciation trend — is consistent across all three points, which is more informative than three runs at the same setting would be. But every individual number here should be replicated across seeds before getting too attached to it.

It’s also a low-selection-pressure regime. With a stiffer selection_pressure="medium" or "high", the convergent shifts on the ecological genes would probably be sharper, and the buffer-dependent splits might either amplify or wash out. I haven’t tested that yet.

What’s next

A few obvious follow-ups:

• A small seed sweep per profile (4-8 seeds) to bracket the speciation-trajectory direction and the learning_rate shift.
• Longer conservative runs to see if the falling speciation index resolves into a single dominant cluster or stabilises around k=2.
• Re-run buffered with crossover enabled — done (2026-05-18): rising speciation trajectories survive crossover; see crossover rerun.
• Add stress and legacy profiles to widen the buffer axis.

For the full numerical results, side-by-side tables, and per-variant artifact links, see stable profile comparison.

Related docs

• Glossary
• Intrinsic evolution experiment docs
• Hyperparameter chromosome design
• Hyperparameter genome devlog