Cross-Check Claims Across Independent Sources

Multi-Source Search cross-validates web research using the search and page-reading capabilities already available to the host agent, instead of treating a single result as an established fact. It runs a five-step workflow: define the claim, a search budget (up to six search calls and six page opens by default), and a stop condition; search across at least two distinct retrieval capabilities, tracing articles back to common origins so circular reporting is only counted once; build claim-level evidence, linking each material claim to source IDs, classifying each as supporting or contradicting, marking the claim sourced or inference, counting genuinely independent sources, and assigning low/medium/high confidence (one independent source for low, two for medium, three for high - a conflicting claim can never be high confidence); validate the resulting JSON report against references/report-schema.md with a bundled, zero-dependency, read-only validator; and present a sourced synthesis organized by confidence, with citations kept adjacent to claims, agreements/disagreements/gaps disclosed, and the search date noted for time-sensitive questions.

Use it when a claim needs fact-checking against independent sources, when the research needs to expose disagreements and evidence gaps rather than just summarize a single result, when the output needs a machine-checkable link between claims and sources, or when the host provides at least two distinct search/retrieval capabilities. Do NOT use it for a simple lookup where one authoritative primary source fully answers the question, or when the user has prohibited external search. High-stakes medical, legal, or financial conclusions still require qualified expert review - the validator only checks internal structure, it does not prove a claim is true, fetch URLs, judge publisher credibility, or detect hidden common sources.

Input is the claim or decision to investigate, plus the host agent's available search/retrieval capabilities (the skill needs at least two distinct ones - separate queries to the same capability do not count as provider diversity). Output is a UTF-8 JSON research ledger following references/report-schema.md: a question, searched_at date, a providers list (and unavailable_providers), a sources array (id/url/publisher/source_type), and a claims array where each claim carries its kind (sourced/inference), confidence, source_ids (the disjoint union of supporting_source_ids and contradicting_source_ids), independent_source_count, and a conflict flag - plus an explicit gaps list for anything the search could not resolve. That report is validated with:

python3 scripts/validate_report.py research-report.json

The command is read-only except for reading the named local report; the skill notes to inspect the path before running it when the report location is supplied by another party.

The skill is provider-agnostic: it works with whatever search or page-reading tools the host agent already has (listed as supporting claude, cursor, gemini, and codex in its metadata), and treats SandBase as optional - it remains useful with native agent tools alone. It names three related skills for adjacent jobs: @efficient-web-research when token-efficient retrieval is the primary concern, @deep-research when a Gemini-backed autonomous research job is specifically required, and @audit-agent-run-evidence when auditing claims and evidence from an existing agent run rather than conducting new web research. Source: community skill from the sandbaseai/sandbase-skills repository, licensed Apache-2.0.

Anyone using an AI coding or research agent who needs to fact-check a material claim against independent sources and hand off a synthesis with an offline-checkable evidence trail, rather than trusting a single search result.