Evidence backed research

Evidence-backed research

Use this skill when an answer, brief, record, or article must be defensible at the claim level. The skill stops at structured evidence; it does not write promotional prose or decide that weak evidence is sufficient.

Inputs

• A bounded research question or target.
• A list of allowed source types or domains, if constrained.
• The date at which the research should be considered current.
• An optional output schema.

Outputs

• One evidence packet using the structure in examples/evidence-packet.json.
• A list of conflicts, stale sources, unsupported assumptions, and unanswered questions.
• A clear recommendation: ready_for_synthesis, needs_more_evidence, or blocked.

Workflow

1. Restate the exact question and define what would count as sufficient evidence.
2. Gather primary sources first, then credible secondary sources where necessary.
3. For every claim, record:
   • the normalized claim;
   • source title and URL;
   • publication or access date;
   • a short supporting excerpt or faithful paraphrase;
   • confidence: confirmed, inferred, or reported;
   • temporal status: current, historical, or unknown.
4. Record conflicting evidence side by side. Do not silently choose the preferred answer.
5. Separate source-backed facts from analyst inference.
6. Check whether precise numbers, dates, and superlatives are directly supported.
7. List evidence gaps and issue the readiness recommendation.

Confidence rules

• confirmed: the subject or authoritative primary source states it directly.
• inferred: strong technical evidence supports it, but the subject did not state it.
• reported: a credible third party states it.

Confidence describes the relationship between claim and source, not how persuasive the claim sounds.

Failure handling

• If a source cannot be opened, record it as unavailable; do not rely on a search snippet as confirmed evidence.
• If evidence is older than the user-defined freshness window, label it historical.
• If sources conflict, return needs_more_evidence unless the output can honestly retain both possibilities.
• If the central claim has no direct support, return blocked.

Boundaries

• Do not invent citations, quotes, dates, or source contents.
• Do not transform inference into fact.
• Do not write publication-ready narrative as part of this skill.
• Do not expose private documents or credentials in the packet.
• Preserve uncertainty when the evidence does not resolve it.