| name | by-hypothesis-debate |
| description | Hypothesis-debate pattern for design strategy selection. Spawns competing strategy agents and uses adversarial ranking to select the best approach before committing GPU compute. Use when starting a new campaign, facing a novel target, or when multiple viable approaches exist. |
Hypothesis-Debate Skill
When to Trigger
- Novel targets (0-1 PDB structures, no known binders)
- Contradictory research findings (CONTRADICTED confidence level)
- Multiple viable modalities (user hasn't specified VHH vs scFv vs de novo)
- User requests "explore options" or "compare approaches"
- Target difficulty rated "novel" or "exploratory"
When to Skip
- Well-studied targets with clear best approach
- User specifies exact parameters
- Preview/quick test campaigns
- Iteration rounds (strategy already established)
Protocol
- Orchestrator spawns 3 hypothesis agents in parallel (via spawnTeam)
- Each agent receives the same research but a different strategic directive
- Each outputs a structured JSON strategy proposal
- Orchestrator spawns 1 reflection agent sequentially
- Reflection agent ranks, critiques, and selects winner
- Winner's strategy becomes the campaign config
- Decision is logged to decision_log.jsonl
Output
The debate produces a concrete campaign config YAML ready for execution.