Research

Goal

Answer proportionally with source-weighted findings, verified URLs, explicit conflicts, and honest evidence gaps.

Success Criteria

• Research depth matches decision risk and available evidence.
• Findings cite supporting sources and calibrated confidence.
• Every URL presented is verified or omitted.
• Conflicts and gaps are reported, even when empty.
• Candidate mechanisms remain hypotheses until a validation method is executed.

Constraints

• Ask only when scope changes source selection, jurisdiction/domain, depth, or decision criteria and cannot be inferred safely.
• HIGH requires 3+ independent agreeing sources plus primary/official evidence; one authoritative source is MEDIUM.
• Keep source_research direct-capable. For five_lens_briefing, use the declared five-lens Process handoffs and do not substitute a single-agent perspective scan for independent lens evidence.

Progressive Contract Loading

Canonical tier files are contracts/input.yaml, contracts/output.yaml, contracts/phase-gates.yaml, and contracts/handoffs.yaml.

Read contracts/index.yaml first and load only the active boundary:

• entry for question, tier, method, role/goal, purpose, known context, and resolved subagent policy/execution/fallback evidence;
• current_phase for SEARCH, SYNTHESIZE, or VERIFY;
• source_research, five_lens, or investigate only after method selection; and
• selected_handoff before every LensResearcher or ResearchPeerReviewer dispatch and return validation; and
• completion only for the artifact being returned.

Missing or invalid selectors fall back to the full named canonical contract. Do not load every contract or research method at entry.

Migration note: assistant-research contracts are v3.0. Convert v2 follow-up fields to typed follow_ups, retaining each material follow-up or one none_needed. v3 requires five frozen, sibling-blind assignment packets, root-only synthesis, distinct peer review, canonical packet/peer-input preimages, tier resolution, and high-stakes bindings. Reject or re-run incomplete v3 artifacts.

Ownership

assistant-research owns source selection, synthesis, confidence, and URL verification. Its specialist gates are authoritative.

For five_lens_briefing, the root Orchestrator freezes exactly five assignment packets before dispatching one independent LensResearcher per named lens. Orchestrator alone builds the contradiction map and initial synthesis after validated returns. A distinct ResearchPeerReviewer then critiques that synthesis. Sequential fallback is constrained to the documented admissible bases and records reduced independence; it never claims delegated independence. Resolve subagent_policy_state before dispatch: this skill instruction makes five-lens delegation delegation_triggered without another permission prompt. Only documented opt-out, real unavailability, or an exact policy block permits sequential_fallback; source_research remains not_required/not_applicable.

Method Selection

• source_research — factual lookup and comparison. Load source_research; choose tier by risk.
• five_lens_briefing — decision-grade work needing perspectives, contradictions, incentives, synthesis, and peer review. Load five_lens.
• investigate — deep entity/domain analysis with ethical boundaries. Load investigate.

Quick tier may use one strong primary source. Standard uses multiple sources. Extensive/deep tiers expand perspectives and primary evidence only when the decision warrants the cost.

Every finding needs confidence grounded in source quality and agreement. For causal or improvement questions, return candidate mechanisms with evidence, counterevidence, gaps, and a validation method—not proven-cause language.

Output

Return status/confidence, concise answer, source-backed findings, candidate mechanisms when applicable, verified sources, conflicts, gaps, and five-lens artifacts only when that method ran.

For five_lens_briefing, include FIVE-LENS PROCESS EVIDENCE: tier resolution, frozen packet-set ID/digest and retained canonical packet preimages, pre-dispatch ordering, exact lens execution records, assignment/packet-set bindings, return validation, peer assignment/identity or fallback, execution modes, reduced-independence state, and the bounded search resource budget. Bind final synthesis to the exact peer-reviewed input or a complete revision closure. The root Orchestrator owns this lifecycle record.

Stop Rules

• Ask every material, non-discoverable question when missing scope changes interpretation; group them by topic and keep them concise.
• Report a gap when sources are inaccessible, stale, conflicting, or too weak.
• Do not finalize with any unverified URL.
• Do not present candidate mechanisms as causes without executed validation.