MARE 0.4.6

MARE 0.4.6

MARE 0.4.6 is the Agentic Evidence Planning release.

The market is moving from single-shot RAG toward agentic, multi-step retrieval that can prove what it found, decide when evidence is insufficient, and keep an audit trail of follow-up searches. This release moves MARE in that direction while keeping the base package lightweight and deterministic.

Highlights

• Added research_plan to Evidence Brief payloads.
• Research plans include:
  • a status such as ready, needs_evidence, needs_stronger_support, needs_source_check, or needs_conflict_check
  • a short rationale
  • concrete next retrieval steps with action names, queries, and reasons
• Added attempt-level audit trails to workflow evidence rescue.
• Evidence rescue now records each alternate query, support assessment, result count, top citation, top score, and whether that attempt improved support.
• Added top-level agent_contract payloads for Python, MCP query tools, and workflow JSON.
• Agent contracts include may_answer, recommended_action, stop_reasons, support/source/research statuses, and the research plan.
• Added mare workflow --task contract for a compact terminal action-contract view.
• Added :contract to mare chat.
• Surfaced research-plan status and steps in mare workflow --task brief.
• Surfaced agent actions and research-plan status in workflow review/answer views and mare chat.
• Surfaced research-plan status and steps in the Streamlit playground Evidence Brief panel.
• Added an Agent Contract panel to the Streamlit playground.

Why It Matters

MARE already explains the trust shape of retrieved evidence. The next useful step is telling a human or agent what to do with that trust signal.

With research plans and audited rescue attempts, agent clients can now:

• answer when evidence is ready
• retrieve stronger support when evidence is weak
• compare sources when coverage is narrow
• resolve conflict signals before acting
• inspect exactly which rescue searches were attempted
• read one compact action contract before deciding whether to answer or continue retrieval

That keeps MARE focused on grounded document work rather than generic chat behavior.