[RESEARCH-INFRA] Enhancing Evolution Diversity via String Seed of Thought (SSoT) for Robust DAG

[RUFFY-369]

Description

The self-evolution pipeline for Hermes Agent—specifically the SyntheticDatasetBuilder and EvolveSkill DAG—currently suffers from deterministic bias. This limits Diversity-Aware Generation (DAG), causing synthetic task distributions to collapse into a narrow set of redundant scenarios. This "Exploration Deficit" prevents the GEPA optimizer from discovering truly novel prompt adaptations, especially on 3B-parameter models which lack the inherent entropy of larger frontier models.
Following the SSoT (String Seed of Thought) protocol (arXiv:2510.21150), I propose integrating a structured entropy-generation mechanism to improve Probabilistic Instruction Following (PIF) within the evolution loop.

Objective

Implement SSoT-based prompting to:

1. Generate Internal Entropy: Force the LLM to output a 16-character alphanumeric seed as a primary randomness source.
2. Extract Randomness: Instruct the model to manipulate this seed (via Polynomial Rolling Hash) to derive a diverse sub-strategy for skill evolution.
3. Scale via CoT: Leverage the model's reasoning traces to improve the alignment of empirical choice frequencies with the target exploration distribution.