Define richer inherited-payload variants and evaluation framework in hyperparameter-genome devlog

[Copilot]

This updates the intrinsic-evolution devlog follow-up with a concrete plan for evaluating richer inheritance beyond chromosome-only transfer. It specifies what to test and how to judge whether richer payloads improve adaptation or just overfit local ecology.

• What this adds

  • Introduces a structured Issue #848 follow-up section in the devlog.
  • Replaces the open-ended “push farther” note with an actionable evaluation framework.

• Candidate payload variants + hypotheses

  • Baseline: chromosome-only inheritance (current behavior).
  • V1: chromosome + lightweight policy prior summary.
  • V2: chromosome + bounded module-state micro-transfer.
  • V3: gated transfer (conditional/clipped inheritance of richer payloads).
  • Each variant includes an explicit adaptation-vs-overfitting hypothesis.

• Comparative experiment design

  • Matched baseline-vs-variant runs under identical ecology/policy settings.
  • Cross-ecology transfer evaluation to separate local fit from generalization.
  • Multi-seed cohort framing so conclusions are distributional, not trajectory-specific.

• Decision criteria / trade-offs

  • Defines evaluation axes: adaptation speed, stability, and generalization.
  • Adds a clear selection rule: retain richer inheritance only when speed gains do not materially degrade stability or cross-ecology performance.

| Variant | Inherited payload | Hypothesis |
| --- | --- | --- |
| **Baseline (current)** | Chromosome only (learning + action priors), fresh decision module | Most robust cross-ecology generalization, but slower cold-start adaptation |
| **V2: module-state micro-transfer** | Chromosome + bounded module state slice (e.g., small distilled value head / compressed policy state) | Highest immediate adaptation speed, but higher instability and local-ecology lock-in |