Skip to content

hybrid_agent.md

Latest commit

 

History

History
41 lines (29 loc) · 1.29 KB

hybrid_agent.md

File metadata and controls

41 lines (29 loc) · 1.29 KB
Raw

Model Card: Hybrid LLM-BO Agent

Intended Use

The hybrid agent uses BO for numerical candidate generation and an LLM-style orchestrator for history inspection, candidate review, explanations, and bounded interventions. The fake LLM is the default for reproducible tests and notebooks.

Inputs

  • Optimizer-visible Stage 2 history.
  • BO candidate summaries.
  • Optional local or mock literature notes.
  • Hard bounds and validated soft bounds.

Outputs

  • Tool calls over strict schemas.
  • BO candidate requests.
  • Candidate execution by immutable candidate ID.
  • Tested feasible final recommendations or no-window declarations.
  • Concise rationales only.

Assumptions

  • Numerical conditions are generated by BO, not arbitrary LLM text.
  • Hard bounds are immutable.
  • Final recommendations must reference tested feasible experiments.
  • Oracle-only information is excluded from LLM context.

Known Limitations

  • Default tests do not make live API calls.
  • Fake-LLM behavior is deterministic and not a proxy for all model behavior.
  • Literature retrieval is local or mock unless a future task adds a live provider.
  • The system does not automate reactor control.

Reproducibility

Hybrid runs record modes, seeds, state-machine events, candidate IDs, observations, warnings, and concise rationales.