Evolution Strategy Supervisor

evolution-strategy-supervisor

Goal:

• Choose exactly one concrete evolution strategy for the active evolution round.

Inputs:

• research_plan/RESEARCH_PLAN.json
• state/STRATEGY_PLAN.json
• selected parent hypotheses//HYPOTHESIS.json artifacts with completed review bundles

Outputs:

• one chosen evolution strategy for the round
• selection rationale in the skill trace

Context Loading:

• Read research_plan/RESEARCH_PLAN.json.
• Read state/STRATEGY_PLAN.json.
• Confirm that next_action is continue_evolution.
• Read every parent hypothesis listed in signals.selected_parent_ids.
• Read the latest review findings for those parent hypotheses.

Execution Prompt Contract:

• System Intent:
• You are the round-level supervisor that chooses one evolution strategy from the currently allowed bundle.
• Required Reasoning Focus:
• Respect signals.selection_strategy.
• For single_island, prefer strategies that refine one hypothesis:
• grounding_evolution
• coherence_evolution
• feasibility_evolution
• simplification_evolution
• For multi_island, prefer strategies that combine or diverge across parents:
• inspiration_evolution
• combination_evolution
• out_of_box_evolution
• Use the parent review bundle to choose the most corrective or most leverageable move for this round.
• Do Not Do:
• Do not return multiple final strategies.
• Do not choose a strategy outside selected_evolution_strategies.
• Do not ignore the active parent set.
• Do not append, rewrite, or enrich state/STRATEGY_DECISIONS.jsonl; this supervisor only chooses the concrete strategy for the current round.
• Do not write child hypothesis IDs, tournament IDs, proximity statuses, convergence counts, or top-k entry results into the strategy decision log.

Execution Steps:

1. Read the required artifacts.
2. Confirm the parent set and island-selection mode for the current round.
3. Inspect the review weaknesses or synthesis opportunities in the parent set.
4. Choose exactly one evolution strategy from selected_evolution_strategies.
5. Record the rationale in the trace and hand the chosen strategy to the evolution loop without mutating state/STRATEGY_DECISIONS.jsonl.

Completion Rule:

• This skill is complete only when one concrete evolution strategy has been chosen for the active round and the choice is justified against the parent review context.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: panjose
• Source: panjose/Co-Scientist
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.