AgentFarm

Sparse ecology doesn't save the ladder: warm-start still loses

The 07-08 saturated A/B left one confound open: every arm ran at full population saturation (final count 32/32 in essentially every cell), while the transferable-signal budget the ladder is graded against was measured in a sparse, fixed-8, no-reproduction regime. This post closes that confound — the decisive follow-up that re-runs the identical inheritance-ladder A/B with --low-churn (max_population = 8) so density matches the precondition gate.

Result: warm-start still loses. Zero robust positives, zero positive mean deltas across all 36 (profile, arm, age) cells. Absolute reward is higher in the sparse regime — cold-start captures that gain — but every warm-start payload still earns roughly half of what Baldwinian cold-start earns. Saturation was a confound for levels, not ranking. The barrier is real for this design.

The regime and what ran

Identical to the 07-08 saturated sweep except for the population cap:

  • Arms (5): baldwinian (P0), lamarckian (P1, weights), p2 (plasticity damping), p3 (optimizer state + replay slice), p4 (fitness-gated θ blend). Default warm-start knobs (damping 0.5, replay limit 256, blend 0.5, gate 1.0).
  • Population: 8 independent agents, max_population = 8 (--low-churn). Reproduction replaces rather than expands the colony — density matches the precondition gate. Warm-start fires at reproduction as before.
  • Horizon: 3000 steps, 200-step warmup, snapshot every 100.
  • Matrix: 3 profiles (conservative/balanced/buffered) × 6 seeds (42 7 19 101 137 256) × 5 arms = 90 runs.

Offspring cohorts are smaller than in the saturated sweep — the colony stays near its cap of 8 so reproduction events are less frequent — but cohort sizes are still adequate for the robustness gate (paired 95% CI excludes zero and within-profile sign agreement ≥ 0.75).

Headline result: still entirely negative

Across all 36 (profile, arm, age) cells, zero cells are robustly positive. Zero positive mean deltas. The profile and arm ordering is preserved exactly from the saturated sweep: buffered is the most damaging ecology for warm-start, P4 (the richest payload) is the worst arm in every profile. No payload wins, at any age, in any profile.

This is the same dose-response seen in the saturated sweep: the richer the payload imposed on the offspring, the larger the loss. If the problem were “insufficient transferred signal,” enriching the payload would help. It does the opposite.

Absolute reward jumps — cold-start takes it

The sparse regime raises absolute reward for everyone: cold-start (baldwinian) offspring earn substantially more than their saturated-regime counterparts, especially in conservative and balanced. The warm-start arms do share in the absolute gain — but they capture roughly half of what cold-start earns, the same fraction as in the saturated run. Reducing density did not shrink the relative gap; it just moved the floor up. The inherited policy is still calibrated to parent conditions that differ from what the offspring encounters, and the offspring still pays for it.

The confound is closed

Saturation was the most natural explanation for the 07-08 result: offspring born into a crowded colony face a different resource landscape than the parent, so a parent-calibrated policy would be systematically wrong. Reduce density so that conditions at birth better match conditions during the parent’s life, and the mismatch should shrink.

It does not shrink. The ranking is identical, the robustness pattern is identical, and the dose-response is identical. Crowding inflated the scale of the loss; it was not the cause.

Together with the 07-08 saturated A/B, the honest claim is: Lamarckian-style policy inheritance does not beat Baldwinian cold-start on early-life reward in either the saturated or the sparse learning-positive regime, and richer payloads make it worse.

What it means for #904

The decision rule — keep a richer payload only if it robustly beats P0 on early-life net reward without degrading stability — is not met in either regime. None of P1–P4 advance. Baldwinian cold-start stays the default. Richer payloads don’t help and usually hurt, worst of all in resource-rich ecologies.

The saturation open question (first bullet in the 07-08 post) is now closed. The remaining open questions concern different warm-start designs, not re-testing this one under yet another density condition.

Reproduce

Low-churn sweep (resume-safe; skips completed cells):

PYTHONHASHSEED=0 python scripts/run_inheritance_mode_ab.py \
  --arms baldwinian lamarckian p2 p3 p4 \
  --population 8 --max-population 8 \
  --num-steps 3000 --snapshot-interval 100 \
  --warmstart-replay-buffer-limit 256 \
  --output-dir experiments/inheritance_ab_lowchurn \
  --disk-database --resume

Grade it (same analysis scripts as the saturated sweep):

python scripts/analyze_early_life_fitness.py \
  --ab-dir experiments/inheritance_ab_lowchurn \
  --baseline-arm baldwinian --treatment-arms lamarckian p2 p3 p4

python scripts/compare_inheritance_arms.py \
  --baseline-dir experiments/inheritance_ab_lowchurn/baldwinian \
  --baseline-label baldwinian \
  --treatment-dir experiments/inheritance_ab_lowchurn/lamarckian \
  --treatment-dir experiments/inheritance_ab_lowchurn/p2 \
  --treatment-dir experiments/inheritance_ab_lowchurn/p3 \
  --treatment-dir experiments/inheritance_ab_lowchurn/p4 \
  --arm-labels lamarckian p2 p3 p4 \
  --output-dir experiments/inheritance_ab_lowchurn/aggregate

Outputs land in experiments/inheritance_ab_lowchurn/early_life/ (early_life_ladder_summary.json / .md) and .../aggregate/.

Open questions

  • Does the clamp relax with a weaker blend? P4 at blend 0.5 imposes half the parent policy; sweeping the blend toward 0 should interpolate back to P0. If even a small blend hurts in buffered, the imposition itself is the problem regardless of density.
  • Per-offspring warm-start contrast. Coverage is ~100%, so the arm-level comparison can’t separate “warm-started” from “born via the treatment path.” Per-offspring applied/skipped telemetry would let us compare warm-started and cold offspring within the same run (see the incompatible_state skip issue).
  • Different transfer designs. Distilled priors, delayed fitness-gated transfer, or partial-layer warm-start are untested; the barrier established here is for this warm-start-on-reproduction ladder.