Research search

Research Investigate

Rigorous evidence-based research applying institutional methodology: evidence hierarchies, source quality grading, systematic search protocols, domain-aware database routing, PRISMA screening, and citation quality scoring.

Important

• Never fabricate citations. Every claim must trace to a verifiable source or be explicitly marked speculative. If a search returns nothing, report nothing.
• Never collapse uncertainty. If evidence conflicts, present the conflict. Do not resolve it artificially.
• Evidence hierarchy is law. Grade every source. Weight conclusions by evidence quality, not volume.
• Prioritize by citation impact. High-citation papers from Tier 1 venues outweigh a dozen low-quality sources.
• Take your time with each phase. Rigor is more important than speed.

Instructions

Step 1: Classify and Frame

Determine the research domain and formulate the question.

Domain detection — check in order:

1. Biomedical/Health: Mentions disease, treatment, drug, gene, symptom, clinical, diagnosis, supplement, nutrition, cancer, condition, pathology, mechanism of action
2. Technical/Engineering: Mentions algorithm, architecture, protocol, system design, performance, implementation
3. Markets/Economics: Mentions market, pricing, valuation, behavioral economics, prediction, forecasting
4. General: Everything else

State: "Domain: [X]. Routing to [relevant databases]."

Question formulation:

• For biomedical questions: Apply PICO framework (Population, Intervention, Comparison, Outcome). State each element explicitly.
• For technical questions: Define the system, the property under investigation, and the evaluation criteria.
• For markets questions: Define the market, the hypothesis, and the observable evidence.
• For general questions: Define the central question, scope boundaries, and 2-3 sub-questions.

Review type detection:

• If the user asks for a systematic review, literature review, meta-analysis, or meta analysis, activate the full PRISMA screening protocol in Step 3.
• Otherwise, proceed with a focused investigation (Steps 2 and 4 only).

Propose the framing to the user. PAUSE. Wait for confirmation before proceeding.

Step 2: Systematic Search

Dispatch 2-4 research subagents in parallel (type: general-purpose), each targeting a different source tier.

Search strategy per domain:

Biomedical/Health:

• Agent 1: PubMed search using MeSH terms and Boolean operators. Target systematic reviews and meta-analyses first.
• Agent 2: ClinicalTrials.gov for active and completed trials.
• Agent 3: Web search for recent guidelines (NIH, WHO, Cochrane) and expert consensus statements.
• Agent 4 (if genetic/molecular): ClinVar, KEGG pathways, Open Targets.

Technical:

• Agent 1: arXiv and OpenAlex for academic papers and benchmarks.
• Agent 2: Web search for implementation reports, documentation, standards bodies.
• Agent 3: Industry whitepapers and RFCs.

Markets:

• Agent 1: OpenAlex for behavioral economics and market microstructure literature.
• Agent 2: Web search for recent data, market reports, regulatory filings.
• Agent 3: Historical evidence and case studies.

General:

• Agent 1: Academic database search (OpenAlex, arXiv as appropriate).
• Agent 2: Web search for authoritative sources (institutional publications, expert analysis).
• Agent 3: Counter-evidence agent — explicitly search for evidence contradicting the emerging consensus.

Citation chaining: For any highly-cited foundational paper found, run a forward citation search (papers citing it) via Semantic Scholar or OpenAlex. Identify newer work building on seminal papers. Record papers cited across multiple sources — these are strong inclusion candidates.

Each agent receives these instructions:

"Search for evidence on: [specific sub-question]. For EVERY finding:

1. State the finding in one sentence
2. Cite the source: Author(s), Year, Journal/Publisher, DOI/URL if available
3. Classify the source type (see evidence hierarchy below)
4. Note sample size, methodology, citation count (if retrievable), and key limitations

Evidence hierarchy (grade every source):

• Level 1: Systematic reviews, meta-analyses
• Level 2: Randomized controlled trials (RCTs)
• Level 3: Cohort studies, case-control studies
• Level 4: Case series, case reports
• Level 5: Expert opinion, mechanism-based reasoning, animal/in-vitro studies
• Level P: Preprints (not yet peer-reviewed — flag prominently)
• Level G: Gray literature, news, industry reports

Source quality indicators:

• Peer-reviewed journal with impact factor: HIGH quality
• Preprint server (bioRxiv, medRxiv, arXiv): MEDIUM quality (flag review status)
• Government/institutional report (NIH, WHO, FDA): HIGH quality
• News article or blog citing primary research: LOW quality (trace to primary source)
• Anecdotal report, forum post, social media: VERY LOW quality (note but do not weight)

If you find NO relevant evidence, say so explicitly. Do not fabricate sources."

Step 3: Screening Protocol (Systematic Reviews Only)

Apply only when the user requested a systematic review or literature review. Skip for focused investigations.

Inclusion/Exclusion criteria (confirm with user or use defaults):

• Date range, language, publication types, study designs

Screening sequence:

1. Title screening: Exclude obviously irrelevant studies. Record count excluded.
2. Abstract screening: Apply inclusion/exclusion criteria. Record reasons for exclusion.
3. Full-text screening: Detailed review against all criteria. Record specific exclusion reasons.

PRISMA flow diagram (render as ASCII for systematic reviews):

Initial search results: n = X
├─ After deduplication: n = Y  (removed: X-Y)
├─ After title screening: n = Z  (excluded: Y-Z)
├─ After abstract screening: n = A  (excluded: Z-A)
└─ Included in review: n = B  (excluded at full-text: A-B)

Step 4: Evidence Synthesis

4a. Citation quality scoring:

Apply these thresholds to weight sources:

Paper Age	Citation Threshold	Classification
0–3 years	20+ citations	Noteworthy
0–3 years	100+ citations	Highly Influential
3–7 years	100+ citations	Significant
3–7 years	500+ citations	Landmark Paper
7+ years	500+ citations	Seminal Work
7+ years	1000+ citations	Foundational

Journal tiers:

• Tier 1: Nature, Science, Cell, NEJM, Lancet, JAMA, PNAS, Nature Medicine, Nature Biotechnology
• Tier 2: High-impact specialized journals (IF > 10), top conferences (NeurIPS, ICML for ML/AI)
• Tier 3: Respected specialized journals (IF 5–10)
• Tier 4: Lower-impact peer-reviewed venues

4b. Evidence inventory:

| # | Finding | Source | Level | Tier | Citations | Quality | Sample/Scope | Limitations |
|---|---------|--------|-------|------|-----------|---------|--------------|-------------|
| 1 | ...     | ...    | L1    | T1   | 847       | HIGH    | N=5000       | ...         |

4c. GRADE assessment (biomedical domain):

• High certainty: Consistent results from multiple Level 1-2 sources. Unlikely to change.
• Moderate certainty: Level 2-3 sources with some inconsistency. May change with new research.
• Low certainty: Level 3-4 sources or significant limitations. Likely to change.
• Very low certainty: Level 4-5 sources, serious limitations, or indirect evidence.

For non-biomedical domains:

• Strong evidence: Multiple independent, high-quality sources converge
• Moderate evidence: Several sources agree with notable gaps
• Weak evidence: Limited sources, significant methodological concerns
• Speculative: Extrapolation from adjacent evidence or expert reasoning only

4d. Conflict analysis:

For each conflict:

1. State the contradiction clearly
2. Assess which side has stronger evidence (by hierarchy level, journal tier, citation count)
3. Identify possible explanations (different populations, methodologies, timeframes)
4. Do NOT resolve — present the tension

Step 5: Deliver

## Investigation: [Topic]

### Question
[PICO framing or structured question from Step 1]

### Domain
[Detected domain] | [Databases consulted]

### Evidence Summary

#### [Sub-question 1]
[Synthesized findings organized by evidence strength, strongest first]

#### [Sub-question 2]
[Synthesized findings]

### Evidence Table
[Full table from Step 4b]

### Certainty Assessment
[GRADE or equivalent from Step 4c]
- Overall certainty: [HIGH/MODERATE/LOW/VERY LOW]
- Key factors affecting certainty: [list]

### Conflicts and Limitations
[From Step 4d — present contradictions and unresolved tensions]

### Conclusions
[What the evidence supports, weighted by hierarchy]
- Claims supported by strong evidence: [list]
- Claims with moderate support: [list]
- Claims that are speculative or contested: [list]

### What Would Change This Assessment
- [Most important missing evidence]
- [Ongoing trials or research to watch]
- [Key assumptions that could be wrong]

For systematic reviews, prepend a PRISMA flow diagram (Step 3) before the Evidence Summary.

For investigations spanning 3+ domains, produce a unified synthesis document with per-domain Evidence Summary sections and Evidence Tables, followed by a Cross-Domain Summary. The prescribed template applies within each domain section.

Ask: "Investigation complete. Want me to dig deeper on any finding, challenge a conclusion, or save this?"

PAUSE. Wait for user response.

Citation Style Quick Reference

APA (7th Edition)

• In-text: (Smith et al., 2023)
• Reference: Smith, J. D., Johnson, M. L., & Williams, K. R. (2023). Title. Journal, 22(4), 301–318. https://doi.org/10.xxx/yyy

Nature

• In-text: Superscript numbers^1,2^
• Reference: Smith, J. D., Johnson, M. L. & Williams, K. R. Title. Nat. Rev. Drug Discov. 22, 301–318 (2023).

Vancouver

• In-text: Superscript numbers^1,2^
• Reference: Smith JD, Johnson ML, Williams KR. Title. Nat Rev Drug Discov. 2023;22(4):301–18.

Default: Use APA 7th Edition unless the user specifies otherwise.

Error Handling

1. No evidence found: Report explicitly. State what was searched and what came back empty. Suggest alternative search terms or adjacent topics.
2. All evidence is low-quality: Flag prominently. Present what exists but lead with the certainty assessment. Recommend what type of evidence would be needed.
3. Domain ambiguous: Ask the user to clarify before proceeding. Do not guess.
4. Conflicting high-quality evidence: Present both sides with full citations. Do not pick a winner — let the user decide.
5. Citation count unavailable: Note the limitation; do not fabricate counts. Weight by journal tier instead.

Examples

Example 1: Focused Investigation (Biomedical)

Input: /investigate effect of creatine on cognitive function

• Domain: Biomedical/Health. No systematic review requested — skip PRISMA.
• PICO: P=healthy adults, I=creatine supplementation, C=placebo, O=cognitive performance measures
• 3 agents: PubMed meta-analyses, ClinicalTrials.gov active trials, web for Cochrane reviews
• Evidence table with Level/Tier/Citation columns
• GRADE assessment, conflicts noted (sleep-deprived vs rested populations)

Example 2: Systematic Review (Technical)

Input: /literature review WebSocket vs SSE for real-time data delivery

• Domain: Technical/Engineering. Systematic review requested — activate PRISMA.
• Screening: 3-stage title/abstract/full-text with documented exclusions
• PRISMA flow diagram rendered
• Evidence table comparing findings across independent sources
• Strong/Moderate/Weak/Speculative framework applied