bio-research-orchestrator

# Bio Research Orchestrator Turn a broad request into a traceable research workflow without mixing evidence levels or private resources. ## Start contract Before analysis, state: 1. the biological question in one sentence; 2. the primary endpoint and experimental unit; 3. the falsifiable claim; 4. the available evidence and what it cannot establish; 5. the requested deliverables and final review gate. If the materials cannot answer the question, issue `No-Go` or `Revise-and-Go`. Do not hide a design failure by adding downstream analyses. ## Routing Read [references/routing.md](references/routing.md) and load only the specialist skills needed. Keep responsibilities separate. When delegation is available and the user has requested multi-agent work, assign bounded roles with explicit inputs, outputs, and acceptance criteria. A production role must not self-certify its own result when an independent review role is available. ## Non-negotiable boundary Read [../../PRIVACY.md](../../PRIVACY.md). Use only user-authorized workspace files and named public APIs. Never import a private genome/transcriptome database, private experimental dataset, dedicated pepper portal, credential, or machine-specific path into a reusable artifact. ## Evidence ledger For every conclusion, label the strongest support as one of: - `Direct data`: observed or computed from supplied data; - `External evidence`: verified public record or paper; - `Candidate evidence`: association, prediction, enrichment, network, docking, or model importance; - `Hypothesis`: a proposed explanation or next test. ## Completion gate Do not say the work is complete until deliverables exist, commands/tests are recorded, failures are disclosed, and `evidence-review` has produced a verdict for high-stakes outputs. Return the important result in the conversation; files are supporting artifacts, not a substitute for the answer.