Cross-Validated Search

Evidence-aware web search, browsing, and claim verification for AI agents.

CLI + MCP + skill surfaces for Gemini, Claude Code, OpenClaw, Manus, and other agent runtimes.

Cross-Validated Search is the flagship repo for source-backed agent answers. It combines live web search, page reading, and claim checking so an agent can surface supporting evidence, conflicting evidence, and source-backed confidence before presenting factual answers.

Canonical names in v16: package cross-validated-search, module cross_validated_search, and MCP command cross-validated-search-mcp. Legacy free_web_search imports and free-web-search-mcp remain available as compatibility aliases.

Recommended free path: ddgs + self-hosted SearXNG. Configure CROSS_VALIDATED_SEARCH_SEARXNG_URL to unlock a free second provider and stronger evidence reports.

At a glance

• one install gives you search-web, browse-page, verify-claim, and evidence-report
• one repo covers CLI, MCP, Gemini, Claude Code, OpenClaw, Manus, and Copilot-adjacent workflows
• one flagship workflow turns raw search results into an evidence artifact with rationale, conflicts, and next steps

Workflow at a glance

Indexer Quick Review

If you are reviewing this repo for collection or ecosystem inclusion, the fastest path is:

1. verify the flagship workflow: evidence-report "Python 3.13 stable release" --claim "Python 3.13 is the latest stable release" --deep --json
2. review the ecosystem contract: docs/ecosystem-readiness.md
3. review the free dual-provider bootstrap: docs/searxng-self-hosted.md
4. review Gemini gallery readiness: docs/gemini-submission-checklist.md
5. review Claude Code and Manus setup notes: docs/claude-code.md, docs/manus.md

60-second verification

Install and verify the public surface:

pip install cross-validated-search
search-web "Python 3.13 release" --json
verify-claim "Python 3.13 is the latest stable release" --deep --max-pages 2 --json
evidence-report "Python 3.13 stable release" --claim "Python 3.13 is the latest stable release" --deep --json

Typical evidence-report JSON shape:

{
  "verdict": "contested",
  "confidence": "MEDIUM",
  "coverage_warnings": [
    "Single-provider evidence path. Add another provider when possible."
  ],
  "analysis": {
    "report_model": "evidence-report-v2",
    "provider_diversity": 1,
    "page_aware": true,
    "support_score": 1.42,
    "conflict_score": 0.61,
    "coverage_warning_count": 1
  }
}

If you are evaluating the repo for ecosystem collection, start with docs/ecosystem-readiness.md.

Ecosystem surfaces

Surface	Entry	Why it matters
CLI	search-web, browse-page, verify-claim, evidence-report	fastest way to verify the repo in under a minute
MCP	cross-validated-search-mcp	works with MCP-native agent runtimes
Gemini	GEMINI.md, .gemini/SKILL.md	gallery / extension readiness
Claude Code	.claude/skills/cross-validated-search/SKILL.md	local skill install path
OpenClaw	cross_validated_search/skills/SKILL.md	bundled skill surface
Manus	SKILL.md, docs/manus.md	GitHub-import-friendly skill

Demo transcript

$ search-web "Python 3.13 release" --json
... ranked sources with citations ...

$ verify-claim "Python 3.13 is the latest stable release" --deep --max-pages 2 --json
{"verdict":"likely_supported","confidence":"MEDIUM", ...}

$ evidence-report "Python 3.13 stable release" --claim "Python 3.13 is the latest stable release" --deep --json
{"verdict":"contested","confidence":"MEDIUM","analysis":{"provider_diversity":1,"page_aware":true}}

Why this exists

Most search wrappers stop at “here are some results.” This repository goes one step further:

• returns structured search results with citations
• reads full pages when snippets are not enough
• classifies evidence as supporting, conflicting, or neutral
• generates a higher-level evidence report with citation-ready sources and next steps
• exposes explainable confidence signals instead of a black-box claim
• works across CLI, MCP, Gemini, OpenClaw, and other agent workflows

Current capabilities

search-web

Use live search for factual or time-sensitive queries.

search-web "Python 3.13 release"
search-web "OpenAI release news" --type news --timelimit w
search-web "人工智能最新进展" --region zh-cn --json

browse-page

Read the full text of a URL when snippets are not enough.

browse-page "https://example.com/article"
browse-page "https://example.com/article" --json

verify-claim

Check whether a claim looks supported, contested, likely false, or still under-evidenced.

verify-claim "Python 3.13 is the latest stable release"
verify-claim "OpenAI released GPT-5 this week" --timelimit w --json
verify-claim "Python 3.13 is the latest stable release" --with-pages --max-pages 2

evidence-report

Generate a compact report that combines search, verification, citations, and follow-up guidance.

evidence-report "Python 3.13 stable release"
evidence-report "Python 3.13 stable release" --claim "Python 3.13 is the latest stable release" --deep --json

The report now includes:

• verdict rationale that explains why the score landed where it did
• stance summary buckets for supporting, conflicting, and neutral evidence
• coverage warnings when provider diversity, domain diversity, or page-aware depth look weak
• citation-ready source digests and recommended next steps

Free dual-provider setup

If you want the strongest free setup, self-host SearXNG and pair it with ddgs:

./scripts/start-searxng.sh
export CROSS_VALIDATED_SEARCH_SEARXNG_URL="http://127.0.0.1:8080"
./scripts/validate-free-path.sh

Or use the compose file directly:

cp .env.searxng.example .env
docker compose -f docker-compose.searxng.yml up -d

Full setup and validation guide: docs/searxng-self-hosted.md.

The current verification model is evidence-aware-heuristic-v3, and the flagship report surface is evidence-report-v2. Together they use:

• keyword overlap between the claim and returned evidence
• contradiction markers in titles and snippets
• source-quality heuristics
• source freshness when a parseable date exists
• domain diversity across the evidence set
• optional page text from top fetched sources
• optional provider diversity when a second provider is configured

Details and limitations are documented in docs/trust-model.md. For a quick product-level comparison with plain search wrappers, see docs/why-not-just-search.md. The next calibration step is outlined in docs/benchmark-plan.md. Quick repository-level context lives in docs/use-cases.md, docs/benchmarks.md, and docs/external-threads.md.

Installation

pip install cross-validated-search

Or install from source:

git clone https://github.com/wd041216-bit/cross-validated-search.git
cd cross-validated-search
pip install -e .

Python 3.10+ is required.

Platform support

Surface	Status	Notes
CLI	Yes	search-web, browse-page, verify-claim, evidence-report
MCP	Yes	cross-validated-search-mcp
Gemini CLI	Yes	gemini-extension.json, root skills/, and .gemini/SKILL.md
OpenClaw	Yes	cross_validated_search/skills/SKILL.md
Claude Code	Yes	.claude/skills/cross-validated-search/SKILL.md
Cursor / Continue / Copilot	Yes	Bundled instruction and skill files
Manus	Yes	Root SKILL.md plus docs/manus.md

Verification model

verify-claim returns one of five verdicts:

Verdict	Meaning
supported	Strong support, low conflict, and enough domain diversity
likely_supported	Evidence leans positive but is not decisive
contested	Support and conflict both carry meaningful weight
likely_false	Conflict is strong and support is weak
insufficient_evidence	Evidence exists but is too weak for a firmer claim

This is an evidence-aware heuristic system, not a fact-level proof engine. Today it still has three important limits:

• default install still starts with ddgs; the recommended collection-grade free path is ddgs + self-hosted searxng
• page-aware verification is optional and still heuristic rather than full-document entailment
• no benchmark-driven confidence calibration yet

MCP example

Add the server to your MCP client:

{
  "mcpServers": {
    "cross-validated-search": {
      "command": "cross-validated-search-mcp",
      "args": []
    }
  }
}

Development

Run the test suite:

python -m unittest discover -s tests -v

Build distributables:

python -m build

Run deterministic benchmark regressions:

python benchmarks/run_benchmark.py

Roadmap

The next upgrades needed for ecosystem-grade collection are:

1. calibrate provider weighting and add stronger provider-specific tests
2. improve page-aware verification beyond snippet and keyword heuristics
3. add benchmark fixtures and regression scoring
4. improve the flagship evidence-report workflow with richer source summarization and calibration

License

MIT License.