Baldwinian vs Lamarckian A/B (Issue #849)

This protocol runs matched intrinsic-evolution sweeps under two inheritance modes to quantify when Lamarckian warm-starting is worth potential stability costs.

Issue: #849

Arms

• baldwinian: offspring inherit only the hyperparameter chromosome and start with a fresh policy.
• lamarckian: offspring still inherit chromosome dynamics, and additionally attempt compatible policy warm-start from the parent.

Matrix

• Profiles: conservative, balanced, buffered
• Seeds: 42 7 19 101 137 256
• Total: 2 × 3 × 6 = 36 runs

Held fixed

• num_steps=1000, warmup_steps=200, snapshot_interval=50
• Mutation: gaussian, mutation_rate=0.15, mutation_scale=0.10, boundary mode reflect
• Crossover disabled (to isolate inheritance mode)
• selection_pressure=low
• Initial diversity: independent mutation (rate=1.0, scale=0.25)
• Speciation: GMM, max_k=4

Ecology regime and divergence from the gate

The inheritance A/B runs with reproduction enabled and a starting population of 8 independent agents. The two ecology variants differ in how much the colony can grow:

Saturated ecology (default, --population 8 --max-population 32)

max_population=32 gives 4× growth headroom. In pilot runs, every seed fills to the cap early (32/32) and runs as a crowded, high-churn ecology for most of the horizon (see Issue #963).

This diverges from the precondition gate (transferable-signal-budget devlog), which ran with reproduction disabled in a fixed-8 uncrowded setting. The gate measured a modest within-life decision-quality signal (~+15–30 net reward at age 10) in an uncrowded, no-repro environment; the A/B measures offspring early-life fitness in a saturated colony where resources and spatial competition differ.

Practical implication: interpret early-life scores alongside the ecology context table in the analysis output (ecology_by_arm in the JSON, “Ecology context” section in the Markdown). A saturation_ratio of 1.0 means every arm ran fully saturated; any comparison is then ecology-matched by construction, but the regime differs from the gate.

Low-churn ecology (--population 8 --max-population 8 --low-churn)

Setting max_population = population (via --low-churn) prevents the colony from growing beyond its starting size. Reproduction can only replace dead agents, keeping ecology density comparable to the gate regime.

Use this variant when the research question is: “does the inherited payload help offspring survive in an uncrowded environment similar to the one where the parent learned?”

Use the saturated variant when the research question is: “does inheritance help under realistic evolutionary pressure in a full colony?”

Run commands

Note: Results under experiments/inheritance_ab_pre_fix/ were collected before a decision-path fix (2026-05-22) that prevented policy weights from influencing actions. That aggregate is invalid; use a fresh output directory after the fix lands.

1a) Run both arms — saturated ecology (default)

PYTHONHASHSEED=0 python scripts/run_inheritance_mode_ab.py \
  --output-dir experiments/inheritance_ab \
  --disk-database \
  --resume

1b) Run both arms — low-churn ecology variant

PYTHONHASHSEED=0 python scripts/run_inheritance_mode_ab.py \
  --output-dir experiments/inheritance_ab_low_churn \
  --population 8 \
  --low-churn \
  --disk-database \
  --resume

2) Compare paired deltas (treatment vs Baldwinian baseline)

python scripts/compare_inheritance_arms.py \
  --baseline-dir experiments/inheritance_ab/baldwinian \
  --baseline-label baldwinian \
  --treatment-dir experiments/inheritance_ab/lamarckian \
  --arm-labels lamarckian \
  --output-dir experiments/inheritance_ab/aggregate

Outputs

• experiments/inheritance_ab/inheritance_ab_manifest.json
• Per-arm sweep_manifest.json and stable-profile analysis artifacts
• experiments/inheritance_ab/aggregate/inheritance_ab_summary.json
• experiments/inheritance_ab/aggregate/inheritance_ab_summary.md
• experiments/inheritance_ab/aggregate/paired_delta_heatmap.png
• experiments/inheritance_ab/aggregate/speciation_trajectories_with_arms.png
• experiments/inheritance_ab/aggregate/startup_transient_comparison.png

Acceptance criteria

Use paired-seed deltas (treatment - baseline) with:

• 95% CI excluding zero, and
• sign agreement >= 75%

Primary readouts:

• Performance: population_mean, population_final
• Stability: startup_transient.peak_death_rate, startup_transient.oscillation_amplitude, lineage.churn_rate
• Diversity impact: speciation_final, speciation_slope
• Mechanism coverage: warmstart_rate, gate_hit_rate (P4), blend_alpha (P4)

Per-profile recommendation categories (label-aware):

• net recommend <treatment-label> (e.g. net recommend lamarckian)
• net recommend <baseline-label> (e.g. net recommend baldwinian)
• performance win + stability loss
• speciation collapse risk
• no robust effect

The classifier consults the metric groups documented above:

• Performance: population_mean, population_final
• Stability loss: startup_transient.peak_death_rate, startup_transient.oscillation_amplitude, lineage.churn_rate
• Speciation collapse: robust negative delta on speciation_slope