Agent Recipe: AI-for-Biology Closed-Loop with Reviewer Handoffs and Eval Contra…

Family: agent_recipe | Showcase: ai-for-biology-throughput-rigor-loop

Purpose

Documents the agent loop for AI-for-biology infrastructure trajectories with explicit reviewer handoffs, eval contracts, and budget posture decisions.

Loop Stages

Stage 1: Read/Search (start_tier)

• Scan SciDEX substrate for existing cross-persona artifacts (Kyle, Claire, Jerome, Kris)

• Identify recurring platform gaps in provenance, reviewer routing, and throughput

• Budget posture: read_search (low token spend)

Stage 2: Source Profile

• Read strategic announcements (Allen-Anthropic, HHMI partnership)

• Identify primary sources for refs_json

• Budget posture: source_profile (medium token spend)

Stage 3: Draft Landscape + Knowledge Gap

• Create landscape artifact: cross-persona gap synthesis

• Create knowledge_gap: throughput-rigor infrastructure deficit

• Reviewer handoff 1: Submit to domain_reviewer for gap validity check

• Eval contract: reviewer must confirm gap is not already addressed by existing infrastructure

Stage 4: Draft Infrastructure Proposal + Analysis Proposal

• Create research_plan (infrastructure_proposal family): ARTIL specification

• Create analysis_proposal: provenance audit methodology

• Reviewer handoff 2: Submit to rigor_reviewer for methods validity

• Eval contract: benchmark metric must be defensible before platform promotion

Stage 5: Benchmark Design

• Create benchmark: TRS (Throughput-Rigor Score) metric specification

• Reviewer handoff 3 (governance): Submit to governance_reviewer

• Governance eval contract: required before any infrastructure change is promoted

Stage 6: Revision Pass

• Process reviewer signals

• Record revision_event as agent_work_packet

• Update artifacts with reviewer-requested changes

Stage 7: Decision Brief

• Record rate-of-discovery uplift estimate

• Document open reviewer disagreements

• Emit downstream agent_work_packet for next execution cycle

Budget Posture Rules

• Spend proportionally more on read/search before proposing infrastructure changes

• Escalate to expensive_action only after domain_reviewer confirms gap is real

• Governance route required for any cross-program or canonical-promotion change

Stop Conditions

• Duplicate canonical infrastructure proposal already exists

• Benchmark cannot measure throughput or rigor change

• Proposal weakens provenance or reviewer accountability

• Cost exceeds expected decision value

Source References

• Allen Institute — Partnering with Anthropic to supercharge research and help cure disease

• Anthropic — Partners with Allen Institute and HHMI to accelerate biological discovery

• Andy Hickl — publications index (NLP, QA, textual inference)