Scientific Method Selector

Scientific Method Selector

Goal

Choose the minimum useful scientific-evidence method or method sequence for the current research job.

The job of this skill is not to sound academic. The job is to classify the evidence task, pick the right route, and explain why that route fits better than broader or weaker alternatives.

This skill selects methods. It does not execute the full downstream literature workflow unless explicitly asked or paired with the selected child skill. It also does not decide the overall proceed / revise / stop verdict for the whole proposal; use quality-gate when a broader judgment is needed.

Default Posture

• search before synthesis
• appraise before trust
• separate discovery from proof
• prefer the minimum useful research route
• do not confuse "found papers" with "claim established"

When To Use

Run this skill when:

• the user asks for scientific backing or research support
• the task needs papers, studies, or evidence review
• a claim needs to be checked against scientific literature
• the right evidence workflow is unclear
• several evidence tasks are being mixed together

This is a mandatory routing gate before:

• running a literature search across scientific sources
• claiming that research supports or disproves something
• synthesizing multiple studies into one conclusion
• escalating from a simple paper search to a broader evidence review

Scope Boundaries

In scope:

• classify the research job
• choose the minimum useful route
• define the next evidence method or staged sequence
• explain why competing routes were not chosen

Out of scope by default:

• writing the final scientific conclusion itself
• treating preprints as equivalent to peer-reviewed evidence
• doing a full systematic review by default
• combining many research methods "just in case"

Problem Types

Classify the task into one primary type first:

• find_relevant_research
• check_claim_strength
• judge_study_quality
• synthesize_evidence
• identify_research_gap

Add a secondary type only if it materially changes routing.

Research Phase Types

Identify the current phase:

• discovery
• screening
• appraisal
• synthesis
• gap-analysis

If phase is unclear, default to discovery.

Core Routing Map

• find_relevant_research -> literature-search
• check_claim_strength -> literature-search -> study-quality-appraiser
• judge_study_quality -> study-quality-appraiser
• synthesize_evidence -> literature-search -> study-quality-appraiser -> evidence-synthesizer
• identify_research_gap -> literature-search -> evidence-synthesizer -> research-gap-detector

Routing Modes

Use one of:

• single
  • one primary method only
• staged
  • an ordered sequence across evidence phases
• hold
  • pause routing and return the task to clarification or narrowing

Decision rule:

• use single when one research job clearly dominates
• use staged when trustworthy output requires multiple evidence phases
• use hold when the research ask is too broad, vague, or premature to route responsibly

Staged Routing Rules

Use staged routing when:

• the user wants a conclusion stronger than a simple search can support
• claim checking depends on both finding studies and judging quality
• evidence synthesis requires screening and appraisal before summary
• research-gap analysis depends on a meaningful view of the current literature

Example routes:

• literature-search -> study-quality-appraiser
• literature-search -> study-quality-appraiser -> evidence-synthesizer
• literature-search -> evidence-synthesizer -> research-gap-detector

Hold Examples

Use hold for cases like:

• See what science says.
• Find all the research on this topic.
• Is this scientifically true?

In these cases the claim, domain, scope, or evidence standard is too weak to route responsibly.

Source Scope Heuristics

Prefer source scope by domain:

• PubMed / PMC for biomedical and health topics
• arXiv for computer science, math, physics, and early-stage technical research
• Semantic Scholar or OpenAlex for broader cross-domain discovery

If the user asks for "science" in a high-stakes domain, prefer higher-trust, domain-relevant sources before broad discovery layers.

Evidence Risk Notes

Always keep these distinctions visible:

• peer-reviewed vs preprint
• review article vs original study
• observational vs experimental evidence
• promising vs established
• direct evidence vs analogy

If the task depends heavily on strong claims in medicine, safety, or policy, say so explicitly and keep the route conservative.

Preprint-Aware Mode

When arXiv or similar preprint-heavy sources are likely to dominate:

• treat discovery value as high and certainty value as limited
• require explicit preprint flagging in downstream outputs
• prefer staged routing over single-step conclusions
• downgrade confidence if strong support depends mainly on non-peer-reviewed sources

Use this mode by default for fast-moving AI, ML, and technical topics unless stronger reviewed evidence is also available.

Fallback Rules

• If the ideal child skill does not exist yet, return the route anyway and mark the missing skill explicitly.
• If the claim is vague, route to hold or intent-clarifier before scientific work.
• If the task is really about general web research, say that a scientific route is not the best first step.
• If the strongest available evidence is mostly preprints, flag that as a limitation rather than hiding it.

Hold Conditions

Return hold instead of forcing a route when:

• the claim or research question is still vague
• the requested scope is unrealistically broad
• the user is asking for a final conclusion without enough evidence definition
• source scope is unclear in a way that would distort the search
• the task is mainly implementation, planning, or opinion rather than scientific evidence work

Output Contract

Always return:

1. research_job
2. problem_type
3. secondary_type (none if absent)
4. current_phase
5. routing_mode (single, staged, hold)
6. selected_methods
7. source_scope
8. missing_skills (none if absent)
9. why_this_route
10. why_not_others
11. evidence_risk_notes
12. next_action

Guardrails

• Do not treat a literature search as proof by itself.
• Do not mix preprints and peer-reviewed studies without flagging the distinction.
• Do not synthesize evidence before a meaningful appraisal step when quality matters.
• Do not force a broad scientific route when the task really needs clarification first.
• Prefer the narrowest trustworthy route that answers the user's real question.
• Keep evidence confidence separate from rhetorical confidence.
• If the evidence base is preprint-heavy, say so early instead of letting the limitation appear only at the end.

Anti-Patterns

Watch for these failure modes:

• paper dumping
• claiming consensus from a few search hits
• treating preprints as settled science
• doing synthesis before relevance and quality screening
• using scientific language to hide uncertainty
• forcing a scientific workflow onto a non-scientific task

Example

User request:

Can you check whether there is good research behind cold exposure improving mental health?

Expected shape of response:

1. research_job: determine how strong the scientific support is for cold exposure improving mental health outcomes
2. problem_type: check_claim_strength
3. secondary_type: none
4. current_phase: discovery
5. routing_mode: staged
6. selected_methods: literature-search -> study-quality-appraiser
7. source_scope: PubMed, PMC, broader scholarly discovery only if needed
8. missing_skills: none
9. why_this_route: the task requires both finding relevant studies and judging their quality before any conclusion is trustworthy
10. why_not_others: direct synthesis is premature before study quality is screened; research-gap analysis is not the user's main need
11. evidence_risk_notes: likely mixture of small studies, heterogeneous interventions, and possible preprint or low-power evidence
12. next_action: run literature-search with a narrow claim and explicit outcome terms