Arguo
The open research intelligence toolkit.
Arguo plans searches, records what each source actually says, checks whether citations support their claims, and writes reports that state their own limits. It does not promise accuracy: it makes research auditable.
Built by Zynu and Shraddha. Maintained at github.com/izynu/arguo.
Every claim carries its evidence
Each run ends in a claim table where every row points to numbered evidence records and sources. This one is from the bundled sample report.
| Claim | Evidence | Sources | Verdict |
|---|---|---|---|
| Climate adaptation budgets across example cities | E1 E2 | S1 S2 | supported (medium) |
| Tree canopy share paired with summer surface temperature | E1 E2 | S1 S2 | supported (medium) |
- 10
- pipeline stages
Counts are generated from the repository contents by build scripts. They describe what ships here, not adoption or quality.
What ships in this repository
Agent Skill
A portable SKILL.md with 20 reference files that any compatible agent can load, following progressive disclosure.
Python research agent
A CLI and library that runs a 10-stage pipeline and writes auditable artifacts. Works offline against fixtures for testing.
Resource catalog
Real, schema-validated records of research resources: government statistics, academic indexes, archives, and APIs.
Evaluation suite
Structural benchmark cases that measure citation validity, unsupported-claim rate, and source diversity without fabricating scores.
How a research run works
Ten observable stages. Each one can be inspected, tested, and audited on its own.
Intake
Validate the question, depth, and date constraints
Planning
Produce a structured plan of subquestions and queries
Search
Query the configured provider, deduplicate results
Retrieval
Fetch pages with size, timeout, and SSRF limits
Evidence
Extract passages into the evidence ledger with provenance
Gap analysis
Find subquestions without evidence, run follow-ups
Verification
Source rubric, corroboration, contradiction scan
Synthesis
Assemble findings that only cite recorded evidence
Claim audit
Check every citation resolves and every claim is supported
Reporting
Write the report and machine-readable artifacts
What a run produces
Every run writes six artifacts. The report cites evidence identifiers that resolve into the machine-readable files.
report.md
The research report with 15 fixed sections, inline [S#] citations, and an audit summary.
sources.json
Every retrieved source with URL, retrieval timestamp, and explicit unknown fields.
evidence.json
Passages linked to claims, with independence groups and content hashes.
research_plan.json
The inspectable plan: subquestions, queries, limitations.
audit.json
Citation issues, support-review verdicts, and contradiction records.
manifest.json
Stage timings, budgets used, providers, and warnings for the run.
Honest limitations
- Automated citation checks verify internal consistency and passage support. They cannot guarantee truth.
- Offline mode runs against labeled fixtures. It demonstrates the pipeline; it does not perform research.
- Catalog records ship as “documented”, not “verified”, until the link checker has live-checked them.
- Heuristic quality signals are rough. The project deliberately avoids numerical trust scores.
- For medical, legal, financial, or safety decisions, treat output as a starting point for human review.