Stable Profile Comparison — Intrinsic Evolution

This document compares three intrinsic-evolution runs that share every evolutionary policy and selection-pressure setting and differ only in their stable initial-conditions profile (resource level, count, and regen rate). The goal is to isolate the effect of the resource buffer on what gets selected, not just on how many agents survive.

The three runs live under:

• experiments/intrinsic_evolution_stable_conservative
• experiments/intrinsic_evolution_stable_balanced
• experiments/intrinsic_evolution_stable_buffered

Setup

All three use the runner described in intrinsic_evolution.md, invoked through scripts/run_intrinsic_evolution_experiment.py with identical arguments except the stable profile manual overrides. From each intrinsic_evolution_metadata.json the resolved configuration is:

What changed

What was held fixed

Population dynamics

initial_population here is the post-warmup population at step 1 of logged steps. Birth/death rates are the per-step means across all 1000 logged steps.

The post-warmup populations track the resource buffer, but all three relax toward a similar steady state (88–99 alive at step 1000). Conservative pulls its population down to ~88 within the first 100 logged steps; buffered drops roughly twice as far in absolute terms (171 → ~99) and takes most of the run to bleed the surplus off. The balanced run is the only one whose death rate does not meaningfully exceed its birth rate on average, which matches its smaller post-warmup overshoot.

Speciation

speciation_index is the silhouette-based score described in intrinsic_evolution.md. Higher means clusters in chromosome space are better separated.

The trajectory shape is the cleanest signal in the comparison:

• Conservative: monotonically declining — clusters merge over the run.
• Balanced: V-shape — splits, partially reconverges, splits again.
• Buffered: monotonically rising — clusters separate further over time.

Conservative and buffered have a comparable number of unique clusters tracked over the whole run (75 vs. 76), but their endpoints are different because conservative’s clusters keep dying off and being replaced, while buffered’s persist and pull apart.

Lineages

All three runs hit the same maximum lineage depth of 3 (great-grandchildren), which is consistent with similar reproduction rates and run length. Buffered keeps the most founder lines alive (60% vs. 50%); conservative has the flattest tree (mean depth ~1.0 indicates many surviving founders have no descendants beyond themselves at the final snapshot).

Gene shifts

Means at step 0 → step 1000, computed from per_gene_initial_mean and per_gene_final_mean in each analysis_summary.json. Percent shift is relative to the initial mean.

Convergent shifts (all three runs agree on direction)

The cross-variant agreement on these is the most robust pattern in the comparison. Cheaper, less-aggressive, less-sharing, less-wandering, smaller networks; more prioritised replay and slower target syncing.

Direction-flipping shifts (resource regime decides the sign)

learning_rate is the clearest case: with more food, the population selects for faster learners; with less food, selection drifts toward slower, more conservative updates. ensemble_size collapses by ~26% only in the conservative run — a “cheap brain under stress” signature. The reproduce multipliers also tell a consistent story: buffered raises the bonus for breeding when wealthy and lowers it when poor (selective reproduction when surplus exists), while conservative does the opposite (breed whenever you can, especially when poor).

Per-variant artifacts

Each run has the standard analysis bundle. Quick links to the most useful plots:

Interpretation

1. Resource buffer scales the carrying-capacity transient, not the steady state. Post-warmup populations track the buffer (135 / 100 / 165) but all three relax to 88–99 alive by step 1000. Bigger buffers buy a longer, slower decay rather than a higher floor.
2. Resource buffer tunes whether speciation strengthens or collapses. Buffered keeps splitting clusters apart (final 0.753, rising). Conservative merges them (final 0.684, falling). Balanced is bistable (V-shape, ends at 0.711). With low selection pressure and abundant food, sub-populations have room to drift apart; under tighter food, the survivors converge on whatever happens to be working.
3. Selection moves the same direction on cheap-and-cautious traits regardless of the buffer. Attack, share, wandering-when-empty, memory size, and network width all drop in every variant. Prioritised replay per_alpha and target_update_freq go up. This is a robust convergent signal across the resource regime.
4. Learning rate is the cleanest direction-flipping locus. Buffered selects for faster learners (+23%); conservative and balanced select for slower (-8% / -6%). ensemble_size is a similar story: only the resource- tight run compresses it dramatically. There appears to be a real gradient between resource availability and the aggressiveness of the learning priors that survive.
5. Reproduction strategy splits along the buffer. Buffered selects for selective reproduction (breed when wealthy, hold off when poor). Conservative selects for opportunistic reproduction (the wealthy multiplier drops slightly while the poor multiplier rises by +21%). Balanced sits between them with neither shift exceeding 1%.

Caveats

• Single seed each. Every run uses seed=42. The qualitative pattern is consistent across the three buffer levels, but a multi-seed sweep is needed before treating any specific number as representative.
• No crossover. All three runs are mutation-only inheritance. The speciation pattern in particular may look very different with crossover enabled, since it provides another mechanism for clusters to mix or stay apart.
• Same warmup length. The 200-step warmup runs the simulation under each resource profile before logging starts, which is part of why the post-warmup populations differ so much. The buffered profile produces a larger founder cohort (because more agents survive warmup), and that initial size affects the first ~100 logged steps. None of the steady-state conclusions depend on the post-warmup population, but the early-window metrics do.

Suggested next steps

• Replicate each profile across a small seed sweep (e.g. seeds 42, 7, 19, 101) to confirm the speciation-trajectory direction is a property of the buffer rather than a single-seed artefact. See scripts/run_stable_profile_seed_sweep.py for the runner and scripts/analyze_stable_profile_seed_sweep.py for aggregation.
• Extend conservative to ~5000 steps to see whether the cluster-merging trend completes (single dominant cluster) or stabilises around k=2.
• Re-run the buffered profile with crossover enabled — done (2026-05-18); see crossover_rerun.md (Issue #845). Rising speciation trajectories survive both crossover arms (6/6 seeds diverging; no robust collapse of final index or slope). Note: conservative profiles do show robust speciation collapse under crossover in the same sweep.
• Add a stress and legacy profile to the comparison for a wider spread on the carrying-capacity axis.