AgentFarm

Are we measuring at the wrong level? Early-life offspring fitness under Lamarckian warm-start

The 05-21 inheritance A/B left a sharp open question. Lamarckian policy warm-start fired for ~85% of reproduction events, yet whole-population summaries (population, speciation) did not move robustly in any profile. One honest reading was that we were measuring at the wrong level: population and speciation are emergent, high-variance summaries, and any effect of inheriting a parent’s policy should show up most strongly right after birth, before ecology washes it out.

This post tests that directly. It reuses the same 36-run dataset, but instead of whole-population readouts it scores each offspring over its first N steps of life. The answer is a clean negative: at the newborn level there are two small, robust behavioral shifts under warm-start, but neither is a fitness gain — and they don’t even point the same way. “Measuring at the wrong level” does not rescue Lamarckian.

Two reward channels (don’t conflate them)

This sim logs two unrelated reward signals, and keeping them separate is the whole point of this post:

  • RL reward (agent_states.total_reward): the cumulative reward at that step (sum of per-step resource_delta + 0.5*health_delta + survival_bonus; see reward.py). This is the fitness-relevant “did the agent get richer/healthier/stay alive” channel.
  • Per-action module reward (agent_actions.reward): a separate per-action value that is nearly constant (~0.135 for ~89% of actions), with a minority of larger positives and some negatives. The only thing it supports here is a coarse “positive-action fraction” — how often an early action returned a positive value rather than a penalty.

An earlier draft of this post told a “more hits, smaller hits” story linking these two. That was wrong: they are different channels and shouldn’t be put in a single causal chain. The corrected reading treats them as two independent readouts, below.

What we measured

For every offspring (an agent born after the 200-step warmup) across both arms, scored at three horizons N ∈ {10, 25, 50} steps of life:

Readout Source Definition
survival to age N agents fraction of (uncensored) offspring alive at age N
net RL reward at age N agent_states.total_reward accumulated RL reward at age N
positive-action fraction agent_actions.reward fraction of actions in the first N steps with reward > 0
resource level at age N agent_states.resource_level instantaneous resources at age N
parent RL-reward gap lineage + total_reward \|offspring − parent\| RL reward over the same wall-clock window

Cohorts are the two arms: Lamarckian offspring (~85% warm-started: 0.852 / 0.849 / 0.849 for conservative / balanced / buffered) vs Baldwinian offspring (cold start by design). A true per-offspring “applied vs skipped” split isn’t recoverable from the stored artifacts — only aggregate warm-start counts are logged — so this is an arm-level comparison.

The parent-anchored readout is enabled by a small data-spelunking win: agents.genome_id supports 0–2 parents plus a counter (::n, parent:n, parent1:parent2:n). For this metric, we only use offspring with exactly one encoded parent, which still yields 100% linkage in these runs without a dedicated parent column.

Paired deltas (lamarckian − baldwinian) are computed per (profile, seed) and gated with the same rule used elsewhere in this project: 95% CI excludes zero AND within-profile sign agreement ≥ 75%.

Runner: scripts/analyze_early_life_fitness.py. Figures: scripts/plot_early_life_fitness.py. Outputs: experiments/inheritance_ab/early_life/.

Headline result

Two small, robust shifts, pointing in opposite directions, neither a fitness gain:

  • Net early RL reward is lower under Lamarckian. The paired delta is negative in all 9 profile×horizon cells, and robustly so for balanced (N=25: Δ −0.277, CI [−0.511, −0.042], 6/6 seeds) and buffered (N=25: Δ −0.449, CI [−0.645, −0.252], 6/6). This is the fitness-relevant channel, and it moves the wrong way for the warm-start arm.

  • Positive-action fraction is slightly higher under Lamarckian — about +1.0 to +1.5 percentage points. Robust in conservative (N=25: Δ +0.015, N=50: Δ +0.014) and buffered (N=25: Δ +0.012, N=50: Δ +0.014), each 6/6. Because actions are either positive or negative here (no zeros), this is equivalent to slightly fewer negative-reward actions. It is not a reward magnitude gain — most positive actions return the same ~0.135 — so it does not translate into more accumulated RL reward.

  • Survival, resource level, and the parent RL-reward gap don’t move. Survival is identical until offspring start dying around age ~42 (a population-wide attrition cliff, not an arm effect); resource level and the parent gap are flat and CI-wide everywhere.

Early-life paired deltas by profile and horizon

The two panels are measuring different things: the top (RL reward) is uniformly below zero, the bottom (positive-action fraction) is uniformly above it.

The RL-reward gap shows up as a small, persistent offset rather than a crossover — the Baldwinian trace sits just above the Lamarckian one and the two stay roughly parallel as offspring age:

Mean cumulative RL reward vs age

And survival is unaffected: both arms track each other until the same attrition cliff:

Offspring survival vs age

How to read this

The 05-21 post raised the possibility that whole-population summaries were simply too coarse to see a real Lamarckian advantage. This analysis does not support that escape hatch:

  1. Warm-start does change newborn behavior. Offspring take slightly fewer negative-reward actions early on. Arms are genuinely different at the action-selection layer for newborns, not just on paper.

  2. But it is not a head start. The fitness-relevant RL reward is, if anything, slightly lower, and resources and survival don’t move. Inheriting a parent’s policy reshapes how a newborn acts without making it better off in its first 50 steps.

  3. So the verdict from 05-21 holds, for the right reason. It isn’t that the effect hides below population-level noise; the newborn-level effects exist but are tiny and don’t favor warm-start on the channel that matters.

Practical takeaway, unchanged: keep Baldwinian as the default.

Caveats

  • Conditional on survival. Net RL reward, resource level, and the parent gap at age N are computed over offspring that reached age N. At N=50 that’s ~60-70% of (uncensored) offspring. Survival deltas are small and non-robust, so the two arms condition on similarly sized subpopulations, but these are survivor estimates, not whole-cohort ones.
  • The two headline metrics are measured on different cohorts. Net RL reward is survivor-conditioned (above), but positive-action fraction spans every offspring that took an action in its first N steps — including those that died before N. The runner records both cohort sizes (n_reached vs n_acted) for transparency. Because survival barely moves before the age-~42 cliff the two cohorts are close in size, but the “reward down, positive-actions up” framing is two readouts on overlapping-but-not-identical populations, not one population measured two ways.
  • Coarse action signal. Because ~89% of action rewards are the same value, “positive-action fraction” is a blunt instrument; treat the +1.5pp as a directional behavioral nudge, not a precise effect size.
  • No multiple-comparison correction. This runs 5 metrics × 3 profiles × 3 horizons = 45 paired tests at n=6. The trustworthy signal is the direction (RL reward negative in 9/9 cells; positive-action fraction up in 9/9); individual “robust” cells are weak on their own.

What shipped (analysis)

  • scripts/analyze_early_life_fitness.py — per-offspring early-life scoring from the run databases, paired deltas, and the same robustness gate as the other inheritance scripts (genome-id parent linkage, right-censoring at the simulation horizon, the two reward channels kept separate).
  • scripts/plot_early_life_fitness.py — the three figures above.

To reproduce:

python scripts/analyze_early_life_fitness.py \
  --ab-dir experiments/inheritance_ab

python scripts/plot_early_life_fitness.py \
  --summary experiments/inheritance_ab/early_life/early_life_summary.json

Open questions

  • What drives the lower RL reward? Net RL reward is resource/health-delta driven; a per-step decomposition (resource vs health vs survival bonus) would show whether warm-started offspring spend down resources faster early or simply gather less.
  • True applied vs skipped split. This stays an arm-level comparison because warm-start status isn’t logged per offspring. Adding that telemetry and re-running would isolate the ~15% cold-start fallbacks from the warm-started majority.
  • Richer inherited payloads (#848). If a bare policy copy reshapes behavior without a payoff, does inheriting optimizer state or replay buffer contents change the early-life picture?