1. Home
  2. Projects
  3. Arguo

Research toolkit

Arguo

A research pipeline where every claim carries its evidence and every citation is audited before the report is written. When the evidence is thin, the report says so.

Documentation site Source PyPI
Licence
Apache-2.0
Language
Python, with a Next.js docs site
Pipeline
10 stages, 6 artifacts
Tests
167 offline tests

Why it exists

Research produced by language models is often fluent, weakly sourced and impossible to audit. Arguo builds the audit layer in: instead of producing a free-standing answer, every run assembles a ledger of sources, evidence passages and claims, then checks its own citations before the report is written.

What ships

Claims carry their evidence

Every claim links to numbered evidence records and sources, with quotes, hashes and provenance recorded rather than reconstructed.

An observable pipeline

Ten stages, each with a status, timings and warnings, recorded in a run manifest you can inspect.

Audited citations

A citation checker validates that every citation resolves and that each passage supports the claim it is attached to.

Offline first-class mode

The full pipeline runs against labelled fixtures with no API keys and no network, so you can test and contribute without a provider.

Resource catalog

1,570 schema-validated records of real research resources across 50 categories and 82 regions, 369 of them with public APIs.

Agent skills

Three portable skill packages: the evidence-first research method, the catalog curation method and the citation audit method.

The ten stages of an Arguo run Intake, planning, search, retrieval, evidence, gap analysis, verification, synthesis, claim audit and reporting, in that order, laid out in rows that read left to right. Evidence and claim audit are drawn in the accent colour because they are the stages that carry the audit trail. 01Intake 02Planning 03Search 04Retrieval 05Evidence 06Gap analysis 07Verification 08Synthesis 09Claim audit 10Reporting
  1. Intakethe question, scope and a written plan
  2. Planningthe plan is turned into concrete search tasks
  3. Searchsources are found against the task list
  4. Retrievalpages are fetched and passages kept
  5. Evidenceeach passage becomes a numbered record, with quote and hash
  6. Gap analysiswhat the evidence does not answer is named
  7. Verificationclaims are checked against the evidence that backs them
  8. Synthesisthe report is written from recorded evidence only
  9. Claim auditevery citation is resolved and tested against its claim
  10. Reportingreport, audit trail and manifest are written to disk
Every run moves through ten stages in order and writes six artifacts: plan, sources, evidence, audit, report and manifest.

Six artifacts per run

Each run writes research_plan.json, sources.json, evidence.json, audit.json, report.md and manifest.json into its output directory. Nothing is published automatically.

Quickstart

git clone https://github.com/izynu/arguo.git
cd arguo
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

arguo doctor
arguo research "Does urban tree cover reduce summer heat in cities?" --offline --output-dir runs/demo

Honest limitations

What the checks cannot do. Automated citation checks verify internal consistency and whether a cited passage supports its claim. They cannot guarantee truth.

Offline mode is a demonstration. It runs against labelled fixtures. It shows the pipeline working; it does not perform research.

Not every record is verified. 1,420 of 1,570 catalog records are documented, not verified. A record moves to verified only after an actual live check.

Not advice. For medical, legal, financial or safety decisions, treat output as a starting point for qualified human review, never a conclusion.

Read the method, then the code

81 documentation pages cover the pipeline, the evidence model and the safety rules.

Read the docs View source